The magnitude of this change and the accelerating speed of it is shocking everyone in data science.

As I’ve mentioned in the past, for most of human history intelligence was expensive because it came packaged inside a person.

If you wanted more analysis, you hired an analyst. If you wanted more code, you hired an engineer. If you wanted more research, planning, design, writing, or judgment, you found people with the right skills and paid for their time. Software made those people more productive, but the person remained the scarce unit at the center of the work.

That arrangement is starting to break.

Epoch AI has assembled one of the clearest measurements of the change. Its inference price analysis asks a simple question: what was the lowest API price required to reach a fixed level of model performance at different points in time?

For a GPT-4-level threshold on MMLU, the answer fell from $37.50 per million weighted tokens in March 2023 to $0.175 in February 2025.

That is a 214× decline in less than two years.

The number matters because it holds capability roughly constant. It is not saying that the same model became cheaper. It is saying that the market learned how to deliver a comparable benchmark level of intelligence for dramatically less money.

That is the curve we need to understand.

The AI conversation still focuses mostly on which model is smartest. But the larger economic question is how quickly adequate intelligence is becoming cheap enough to use everywhere. Once capable inference becomes inexpensive, we stop rationing it. We put it inside every product, every workflow, every decision, and every machine.

The cost curve bends down at the same time the demand curve bends up.

How the 214x Analysis Works

The logic is simple. Epoch combines public model benchmark scores with API pricing data. It sets a capability threshold, then identifies the lowest-priced non-reasoning model available at each point in time that meets or exceeds that threshold.

For the series we are looking at, the threshold is an MMLU score of 86, roughly the level reached by the original GPT-4 release.

MMLU is a broad general-knowledge benchmark. It is not intelligence in full. No benchmark is. But it gives us a consistent measuring stick across model generations.

The progression is striking:

• GPT-4 launched at a weighted price of $37.50 per million tokens.

• GPT-4 Turbo brought that down to $15.

• GPT-4o brought it to $7.50.

• Gemini 1.5 Pro reached the threshold at $2.19.

• Gemini 2.0 Flash reached it at $0.175.

Epoch calculates the token price as a 3:1 weighted average of input and output prices. First-party pricing is used when available; otherwise, the analysis uses the median across API providers. Reasoning modes are excluded because they can generate far more internal and output tokens, making a simple price-per-token comparison misleading.

Epoch then repeats the process across multiple benchmarks and capability thresholds. The measured price declines vary widely, from 9× to 900× per year, with a median of 50× per year. The underlying notebooks and data are public, which makes the result more useful than another floating chart with no visible method.

There are important limits.

1. A token is not a completed task.

2. MMLU does not measure every form of reasoning, reliability, tool use, latency, context management, or domain expertise.

3. API list price is not the same as the provider’s underlying economic cost.

4. A model that meets the threshold on a benchmark may still behave very differently inside a real workflow.

The analysis does not prove that all intelligence became 214 times cheaper. It shows that one useful, fixed level of machine capability became available at 1/214^th of its prior listed token price.

That is enough to change behavior.

The Price of Adequate Intelligence

The frontier gets the headlines, but in reality, the economy is transformed by what happens several steps behind it.

Most work does not require the smartest model in the world on every call. It requires a model that is good enough for the task, connected to the right context, given the right tools, bounded by clear instructions, and checked by an evaluation loop.

This is the idea of adequate intelligence.

An accounts-payable workflow does not need to solve open problems in physics. A sales-research agent does not need to write a great novel. A code-review agent does not need to manage a board meeting. Each system needs enough cognition to perform a bounded job reliably.

When that level of cognition is expensive, we use it sparingly. We reserve it for high-value queries and keep humans inside every intermediate step. When it becomes cheap, we redesign the workflow around its availability.

We let the agent read every document instead of a sample.

We let it test ten approaches instead of drafting one.

We let it monitor continuously instead of checking weekly.

We let it retry, compare, criticize, reconcile, and escalate.

This is not just a cheaper answer. This is a larger amount of cognition applied to the problem.

That distinction is where the economic implications begin.

Why Chat Becomes Command

In The Shift From Chat to Command, I called out that the interface to AI is changing from conversation to delegation.

A chatbot produces a response. An agent produces an outcome.

The prompt becomes the work order. The thread becomes the workspace. The human moves from performing each task to defining the objective, supplying context, granting tools, supervising execution, and deciding what ships.

But command is inference-intensive.

A chat answer may require one model call. An agentic workstream may require hundreds or thousands. The agent reads files, searches sources, calls tools, writes artifacts, runs tests, checks results, repairs failures, and asks for approval. A multi-agent system may explore several paths in parallel and discard most of them.

None of that works economically if capable inference remains precious.

The move from $37.50 to $0.175 changes the feasible “design space”. A workflow that would have been irresponsible to run continuously at the old price can become infrastructure at the new one. The system can spend more tokens thinking, checking, and trying again while still producing an outcome that is cheap relative to human labor.

This is why the shift from chat to command is not just a product-design choice. It is being pulled forward by the intelligence cost curve.

Cheap answers create chatbots while cheap sustained cognition creates agents.

Why Agents Become an Energy Problem

In Agents Per Gigawatt, I argued that the future of productivity will not be measured only as output per person. It will also be measured by how much useful synthetic labor a person, company, or country can command.

If agents are labor, inference is the fuel. Compute consumes energy. Energy flows through data centers. Data centers become factories for cognitive work.

The intelligence cost curve strengthens that thesis because it accelerates demand for inference. As the dollar price of adequate cognition falls, more use cases clear the economic threshold. More agents run. They run longer. They work in parallel. They monitor systems that were previously checked by people only when something went wrong.

This is where the apparent contradiction resolves.

Intelligence can become cheaper while the physical infrastructure behind it becomes more valuable.

The token price falls. Total token use explodes. Model efficiency improves. Aggregate compute demand rises. Data centers consume more power even as each unit of useful cognition costs less.

This is classic abundance behavior. When the cost of a useful input collapses, we rarely hold consumption constant and pocket the savings. We discover new ways to spend it.

Cheaper lighting did not make societies want less light. Cheaper bandwidth did not make us transfer less data. Cheaper compute did not make us run fewer calculations.

Cheaper intelligence will not make us think less.

It will create ambient cognition.

That is why power generation, grid access, chips, cooling, land, fiber, transformers, and data-center efficiency become part of the labor stack. The more affordable intelligence becomes in dollar terms, the faster demand runs into physical constraints.

Walking the Curve Forward

We should be careful here. Exponential curves make fools of anyone who turns them into a precise point forecast.

The right way to walk the intelligence cost curve forward is to separate three different curves that are often collapsed into one.

The first is price per token at a fixed capability level.

The second is cost per completed task.

The third is cost per reliable outcome inside a real operating system.

The first curve is falling fastest. Model compression, better architectures, improved hardware, higher utilization, batching, competition, and lower provider margins can all reduce the listed price of a token. Epoch explicitly notes that it did not attempt to isolate how much each factor contributed.

The second curve falls more slowly because tasks expand to consume available intelligence. A better agent may read more context, perform more tool calls, run more checks, and explore more alternatives. Tokens get cheaper, but the workflow deliberately uses more of them.

The third curve is the one businesses ultimately care about. It includes reliability, supervision, integration, security, failure recovery, and the cost of the surrounding harness. A cheap model inside a weak system can still be expensive because humans must repair the output. A more expensive model inside a disciplined workflow can be cheaper because it finishes the work correctly.

This is why price per token will eventually become a background metric, like the price of a floating-point operation. Important to infrastructure planners. Nearly invisible to the end user.

The useful unit will become cost per accepted outcome.

Still, the token curve shows the pressure underneath the system. Using the study’s February 2025 GPT-4-level observation of $0.175 per million weighted tokens as a starting point, even a much slower 3x annual decline would bring the price to roughly 1.9 cents by 2027 and less than one-tenth of a cent by 2030.

A 10x annual decline would make the raw token price economically trivial far sooner.

Those are scenarios, not forecasts. They are also the point.

If the direction persists at even a fraction of the historical rate, the economy stops asking whether it can afford a model call. It starts asking how much cognition should be applied to each objective, how many parallel attempts should run, and what system will judge the results.

The constraint moves from generation to selection.

The Demand Explosion

The first-order effect of cheaper intelligence is cost savings. The second-order effect is demand creation. The second-order effect will be much larger.

Today, companies ration analysis because people are expensive and attention is finite. Customer research is periodic. Compliance review is sampled. Forecasts are updated on a schedule. Software tests cover the paths the team had time to write. Managers accept incomplete information because gathering more is not worth the labor.

When cognition becomes cheap, the default changes.

Most of the technologists I know along with the broader futurist and “future of work” community believes (with good evidence) we will see these soon:

1. Every customer can receive individualized support.

2. Every contract can be reviewed.

3. Every product decision can be challenged by competing models.

4. Every code change can be tested against a growing library of failure cases.

5. Every operational system can be watched continuously.

6. Every executive can receive a live synthesis of the company instead of waiting for the monthly deck.

The geeky way to put this is the “long tail becomes economically reachable” but I think it’s going to feel like a never-ending renaissance of new experiences and improved living.

Markets that were too small to support human specialization can support agentic service. Languages that were too costly to localize can receive native-quality interfaces. Diseases with too few patients to attract large research teams can receive persistent machine attention. Small businesses can access analysis, design, software, finance, and operations capacity that previously belonged to large enterprises.

The cost decline makes previously uneconomic work exist.

This is why aggregate inference spending can rise even as the unit price collapses. The world will consume vastly more cognition because the number of useful places to apply it is close to unlimited already, and we’ll use our rising intelligence to find + create more surface area.

Human Capital Gets Repriced

When execution is expensive, the person who can produce the artifact holds leverage. When execution becomes cheap, leverage moves to the person who can define, direct, verify, and integrate the work.

This does not mean human knowledge becomes worthless. It means the value moves up the stack. Domain knowledge is what lets an operator specify the real objective, notice the hidden constraint, reject a plausible error, and understand which tradeoff the model cannot make on its own.

The analyst of the next decade does not compete with one agent. The analyst competes with another analyst who can direct fifty agents, compare their work, and find the signal inside the volume.

The manager does not win by protecting a larger headcount. The manager wins by building a system that turns a small amount of human judgment into a large amount of reliable execution.

The company does not win because it bought access to the best model. Model access diffuses. The company wins because it has captured context, mapped workflows, clear permissions, proprietary data, strong evaluations, trusted decision rights, and people who know what good looks like.

The model is the engine.

The harness is the productive system.

As intelligence gets cheaper, weak harnesses become more dangerous. A system that can generate one bad answer is annoying. A system that can generate, distribute, and act on a million bad answers is an operational failure.

Abundance punishes weak judgment.

The organizations that understand this will invest in evaluation and control at the same time they invest in generation. They will not measure adoption by the number of employees with chatbot accounts. They will measure it by reliable workstreams, accepted outcomes, cycle-time compression, and the ratio of machine execution to human review.

The Bottleneck Migration

The intelligence cost curve is one of the clearest empirical signs that scarcity is moving.

First, execution becomes abundant. Models can draft, analyze, code, classify, compare, and plan at a cost that keeps falling.

Then demand concentrates on the physical layer: compute, chips, electricity, grid access, cooling, and capital.

Then it moves again into the sovereign layer: trusted data, secure systems, institutional competence, decision rights, legitimacy, judgment, and the ability to turn machine output into real-world action.

For an individual, the bottleneck shifts from hours available to clarity of intent and capacity to review.

For a company, it shifts from hiring enough people to encoding enough context and building enough trust in the workflow.

For a country, it shifts from population alone to energy systems, compute infrastructure, capital formation, data access, security, and institutions capable of deploying synthetic labor without losing control of it.

This is the Sovereign Singularity thesis in economic form. Digital intelligence becomes abundant. Physical capacity and human agency become more valuable because they determine how abundance is aimed.

The winners will not control intelligence. They will command it.

What to Build Now

The practical response is not to wait for the curve to finish. Curves like this never finish. They change the design assumptions underneath the work while most organizations are still budgeting around last year’s costs.

Start with six moves.

1. Measure outcomes, not prompts. Track the cost of completed, accepted work. Token price is an input. The result is the product.

2. Build evaluation before volume. Cheap generation without reliable selection creates cheap slop at industrial scale.

3. Capture context. Write the operating procedures, connect the source systems, define permissions, and preserve the institutional knowledge agents need to act.

4. Route intelligence. Use the cheapest model that can perform each step, then escalate difficult cases to stronger models and humans.

5. Spend the savings on search. Run competing approaches, test assumptions, explore edge cases, and use abundance to widen the solution space.

6. Protect the final mile. Keep human accountability where trust, taste, ethics, relationships, and irreversible decisions matter.

The goal is not maximum automation. The goal is maximum useful agency.

The best operators will use cheap intelligence to expand what they can attempt while becoming more deliberate about what they approve. They will not try to personally touch every intermediate step.

They will design the system, set the standard, review the exceptions, and decide where the machine workforce should go next.

Intelligence Becomes an Input

The 214x decline does not tell us when artificial general intelligence arrives. It does not prove that benchmarks equal human judgment. It doesn’t give us a clean forecast for the price of every task.

It tells us something more immediate.

A useful level of machine cognition is moving from scarce capability to cheap input.

That transition is enough to reorganize work. It is enough to turn chat into command, software into labor systems, data centers into workforce infrastructure, and energy policy into economic strategy. It is enough to let small teams command output that once required large institutions. It is enough to make evaluation, judgment, context, and trust the new scarce assets.

We spent the first years of the AI era asking how smart the models could become.

The next phase will be defined by how cheaply adequate intelligence can be deployed, how much of it we can run, and how well we can aim it.

The cost of intelligence is collapsing. The value of judgment is compounding.

That is the curve to watch.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

Why our family office invested in the company building power for the intelligence economy

The AI boom looks like software from a distance.

Up close, it looks like a construction site.

It is concrete, copper, transformers, cooling systems, gas lines, permits, switchgear, and power. Enormous amounts of power. The most advanced chips in the world do nothing while they sit in a warehouse waiting for an energized data center.

That physical constraint is why our family office invested in Bloom Energy.

We did not invest because fuel cells are new. They are not. We did not invest because Bloom has discovered free or carbon-free electricity. It has not. We invested because the unit of competition in AI is changing.

The first phase of the AI race was about access to models.

The second was about access to chips.

The next is about access to energized megawatts.

Bloom sells those megawatts faster than much of the traditional power industry can deliver them. Its solid oxide fuel cells can be built in a factory, shipped in modular blocks, installed onsite, and turned on without waiting for a major transmission project or a new central power plant.

The product is electricity.

The value is time.

The grid runs on industrial time

Electric utilities were not designed for the current moment.

They were built around long-range forecasts, measured demand growth, centralized generation, and infrastructure that could take years to approve and build. That model made sense when electricity demand grew slowly and predictably.

AI broke the forecast.

A single new AI campus can require hundreds of megawatts. The largest proposed campuses are measured in gigawatts. They appear faster than utilities can build substations, transmission lines, and generation. At the same time, gas turbines face multi-year order backlogs. Nuclear power takes longer. Renewables can be built quickly, but a data center needs firm power at night, during bad weather, and when the grid is already strained.

The old energy system asks customers to wait.

The AI economy cannot wait.

This is not impatience for its own sake. An unpowered AI factory is stranded capital. The land has been bought. The building is going up. The GPUs may already be committed. The customer contracts often have delivery schedules and performance requirements. Every month without power is a month without compute revenue.

That changes the purchasing decision.

A data center developer is no longer asking only, “What is the cheapest electricity over twenty years?” The developer is also asking, “What can I turn on next year?” A solution can cost more per megawatt-hour and still create far more value if it begins earning revenue several years earlier.

Time to power has become part of the cost of power.

What Bloom builds

Bloom’s Energy Server uses solid oxide fuel cells, or SOFCs, to convert fuel into electricity through an electrochemical reaction.

Most power plants burn fuel to create heat, use that heat to spin machinery, and turn mechanical motion into electricity. Bloom skips much of that chain. Its cells operate at high temperatures and convert the chemical energy in natural gas, biogas, or hydrogen into electrical energy without combustion at the point of generation.

That distinction creates several advantages.

The system has few moving parts. It uses little water during normal operation. It produces very low levels of local pollutants such as nitrogen oxides, sulfur oxides, and particulate matter compared with combustion engines. It is quiet, compact, modular, and capable of running around the clock.

It is also a factory product.

Bloom does not need to design a new power plant from scratch for every customer. The same cell, stack, module, server, and power-block architecture can be repeated from a small commercial installation to a campus measured in hundreds of megawatts.

That modularity matters because it changes how power capacity gets built. Traditional generation arrives as a project. Bloom capacity can arrive as a product.

Print the cells. Build the stacks. Assemble the servers. Ship another block.

That is the mechanism behind the speed.

Oracle moved from trial to platform

Oracle is the clearest proof of what changed.

In July 2025, Oracle and Bloom announced that Bloom would deploy onsite fuel-cell power at selected Oracle Cloud Infrastructure data centers. The original promise was aggressive: power an entire data center within 90 days.

Bloom says the first system was operational in 55.

That execution turned a vendor relationship into a platform decision. In April 2026, the companies expanded the partnership. Oracle contracted for an initial 1.2 gigawatts of Bloom capacity deploying across U.S. projects through 2027, under a master agreement supporting up to 2.8 gigawatts.

Those numbers are enormous for the fuel-cell industry. Bloom had deployed roughly 1.5 gigawatts across its entire history when the original Oracle collaboration was announced. One customer’s framework is now almost twice that cumulative figure.

Oracle has a reason to move this aggressively. Its cloud infrastructure business is growing faster than its power supply can comfortably follow. Oracle finished fiscal 2026 with $638 billion of remaining performance obligations, up 363% from the prior year, while quarterly infrastructure-as-a-service revenue grew 93%.

That backlog is not an abstract accounting number. It is a claim on future data centers.

Oracle needs buildings, chips, cooling, networking, and electricity to turn that contracted demand into revenue. Bloom helps remove one of those dependencies from the utility timeline.

The relationship has also moved beyond using fuel cells as a bridge. Oracle and its partners selected Bloom to provide up to 2.45 gigawatts for Project Jupiter in New Mexico. The revised plan replaces proposed gas turbines and diesel generators with a large fuel-cell microgrid.

That is not backup power. That is the power plant.

Nebius is buying certainty

Nebius shows the same shift from the perspective of an AI-native cloud provider.

The company has signed large, long-term compute agreements and plans to deploy more than five gigawatts of NVIDIA systems by the end of 2030. Its growth is limited not only by demand for compute but by its ability to deliver that compute on schedule.

In May 2026, Nebius selected Bloom for its first U.S. fuel-cell deployment. The project is designed to provide approximately 250 megawatts of guaranteed capacity using about 328 megawatts of installed equipment. The capacity is expected to come online in three phases, and Nebius says the full installed system is planned to operate this year.

The planned fuel cells replace reciprocating engines.

The commercial structure is revealing. According to Nebius’s SEC disclosure, Bloom will install, operate, and maintain the systems. Each phase has a ten-year supply term. Nebius will pay aggregate monthly service fees of up to $2.6 billion over the agreement.

Nebius is not simply buying boxes. It is buying capacity, electricity, installation, operations, maintenance, and performance.

In other words, it is transferring a hard infrastructure problem to a specialist.

That matters for a company trying to scale quickly. The more of the power stack Nebius can turn into a contracted service, the more attention and capital it can put into compute, software, customers, and deployment.

Oracle is using Bloom to control the power architecture.

Nebius is using Bloom to control the delivery schedule.

Both are buying certainty.

Power is becoming part of the computer

There is another reason this matters.

The boundary between a data center and its power system is starting to dissolve.

AI workloads create dense and rapidly changing demand. A conventional data center pulls alternating-current power from the grid, moves it through transformers and switchgear, converts it to direct current, maintains batteries and uninterruptible power systems, and converts voltage again before electricity reaches the chips.

Every step adds equipment, space, losses, and failure points.

Fuel cells naturally produce DC power. Bloom has been developing its platform around emerging 800-volt DC data center architectures, load following, and grid-independent microgrids. The company’s systems can ramp with changing loads and use inverter controls to manage voltage and frequency without relying on rotating machinery.

The long-term opportunity is not just putting a different generator beside the same data center.

It is redesigning the power path from fuel to compute.

If onsite DC generation can remove conversion steps, reduce switchgear, shrink the backup architecture, and preserve more campus space for revenue-producing equipment, then the comparison with a turbine becomes too narrow. The power system begins to improve the data center itself.

The power architecture becomes part of the compute architecture.

That could become one of Bloom’s strongest advantages. Speed gets the company into the project. Integration can keep it there.

Can Bloom build fast enough?

Large orders are useful only if Bloom can manufacture and install the product.

The formal plan has been to double annual production capacity from roughly one gigawatt to two gigawatts by the end of 2026. Bloom says its existing manufacturing footprint can eventually accommodate approximately five gigawatts of annual output.

Those statements are easy to mix up.

Bloom does not have five gigawatts of fully installed annual capacity today. It has facilities with enough space and infrastructure to scale toward that level. According to the company’s 2025 annual report, each additional gigawatt within that footprint is expected to take six to nine months and require $100 million to $150 million of capital.

That is a compelling expansion model. Adding a gigawatt of manufacturing capacity for a fraction of the capital required to build a gigawatt-scale generating asset gives Bloom a way to follow demand without making one irreversible factory bet.

But the ramp will not be frictionless.

An exit run rate of two gigawatts does not mean Bloom can ship two gigawatts during 2026. Capacity is being added throughout the year. The average available capacity will be lower than the December run rate, with the fuller benefit appearing in 2027.

Manufacturing is also only one link. Bloom depends on specialized materials, suppliers, installation contractors, electrical equipment, gas infrastructure, permits, financing, and customer-site readiness. Product can leave a factory faster than a data center campus can become ready to receive it.

This is why we view two gigawatts by year-end as a credible operating target and five gigawatts as valuable optionality, not near-term production.

The distinction matters.

Capacity is real when it can make, ship, install, and service product at quality. Floor space is only the beginning.

The cost curve has three layers

SOFC costs should fall as production scales, but “cost” needs to be unpacked.

There is the factory cost of the hardware.

There is the installed cost of the full system.

There is the lifetime cost of the electricity.

The first curve should fall fastest. Higher production volume improves factory utilization. Better cell-printing yields reduce waste. New stack designs can produce more power from the same material. Standardized power blocks simplify assembly and installation. Suppliers can invest against larger and more predictable orders.

Bloom says it has delivered double-digit product-cost reductions for many years. Our base case is more conservative going forward: roughly 8% to 12% annual reductions in factory cost per kilowatt over the next three to five years, with installed costs falling more slowly.

The lifetime cost will move slower still.

Natural gas has a market price. Financing has a market price. Installation includes local labor and civil work. High-temperature fuel-cell stacks degrade and require periodic replacement. Lower hardware cost helps, but durability, efficiency, and service labor will determine how much of that improvement reaches the cost of electricity.

This is why stack life may matter more than stack price.

A longer-lasting module reduces replacement material, truck rolls, field labor, downtime, and service reserves. It also maintains better efficiency for longer, creating more electricity from the same fuel. One materials-science improvement can lower several costs at once.

The U.S. Department of Energy’s long-term SOFC targets include $225 per kilowatt for the stack and $900 per kilowatt for the system, along with low degradation and greater than 60% efficiency. Those are research targets, not Bloom’s current commercial pricing. They show how much room the technology may still have to improve.

There is also a difference between cost and price.

When customers are desperate for power, Bloom does not need to pass every manufacturing gain through immediately. Some cost reduction can show up first as higher margins. That is normal. Prices fall fastest when supply expands, competition arrives, and Bloom begins pursuing markets where the customer is less willing to pay for speed.

The cost curve expands the market.

The scarcity premium funds the expansion.

What comes after data centers

AI is the first large market willing to pay heavily for time to power. It will not be the last.

As SOFC costs decline, adoption should move outward from the places where reliable electricity has the highest value.

Advanced manufacturing is an obvious next step. Semiconductor plants, pharmaceutical facilities, food processors, and chemical sites need steady power and cannot tolerate grid interruptions. Many also need heat, hot water, steam, or cooling. Bloom’s high-temperature exhaust can be captured in a combined heat and power system, pushing total energy efficiency above 90% in the right application.

Hospitals, universities, and critical infrastructure value resilience. EV fast-charging hubs and fleet depots can face years of distribution-grid upgrades before receiving the capacity they need. Utilities can use load-following fuel cells as modular local capacity where demand is growing faster than wires can be built.

Biogas creates another path. A dairy, landfill, or wastewater plant already produces methane. Converting that fuel onsite avoids flaring or expensive upgrading while producing firm electricity.

Carbon capture may become important because the SOFC process can create a concentrated carbon dioxide stream that is easier to separate than the exhaust from a conventional combustion plant. Marine auxiliary power, hydrogen, and fully fuel-cell-powered ships are longer-term possibilities.

Some markets will remain difficult.

We do not expect SOFCs to sweep into ordinary homes or displace cheap wholesale grid power soon. Small systems lose scale. Natural gas still costs money. Maintenance still matters. Solar, batteries, turbines, engines, nuclear, and the grid will keep improving.

Bloom does not need to win everywhere.

It needs to win where waiting is expensive.

This is good news, not perfect news

The Oracle and Nebius agreements validate Bloom’s central claim: onsite fuel cells can move from niche resilience equipment to primary power for the AI economy.

They also expose the risks.

Bloom issued Oracle a fully vested warrant for approximately 3.53 million shares at $113.28 as part of the relationship. Oracle later disclosed that it sold Bloom warrants during fiscal 2026. That commercial incentive improved Oracle’s economics and should be included in any honest reading of the deal.

Project Jupiter is also not finished. Its revised fuel-cell plan still faces an air-permit process, and a proposed natural-gas pipeline route encountered a right-of-way denial from the New Mexico State Land Office. Fuel cells shorten one critical path. They do not repeal permitting, fuel infrastructure, or local politics.

Natural-gas SOFCs are not carbon-free either. They produce far fewer local pollutants and use far less operational water than many combustion alternatives, but they still emit carbon dioxide. Their climate value depends on what they replace, how efficiently they run, what fuel they use, and whether carbon capture becomes economic.

There are business risks too: customer concentration, AI capital-spending volatility, manufacturing defects, stack-life assumptions, service costs, supplier quality, competing technologies, and the possibility that turbine availability or grid capacity improves faster than expected.

We are investors, not believers. The difference is vital.

Belief looks for reasons the thesis must be right. Investing looks for the mechanism, the evidence, the price, and the ways the thesis can break.

The evidence today says Bloom has moved from selling a cleaner onsite generator to selling a faster path to productive infrastructure.

That is a much larger market.

The intelligence economy needs physical rails

The popular story of AI is that intelligence is becoming abundant.

That is true at the model layer. It is not yet true at the infrastructure layer.

Intelligence still has to run somewhere. It needs chips. Chips need buildings. Buildings need cooling. Everything needs power.

The more abundant digital intelligence becomes, the more pressure it puts on scarce physical systems. This is the paradox at the center of the AI buildout. Software can scale instantly. Electricity cannot.

The companies that close that gap will capture enormous value.

Some will build chips. Some will build data centers. Some will build turbines, reactors, transmission lines, batteries, cooling equipment, and gas infrastructure. Bloom’s opportunity is to build the modular power layer that lets a customer stop waiting for the full energy system to catch up.

That is why the Oracle expansion matters.

That is why the Nebius adoption matters.

That is why our family office invested.

Bloom is not merely participating in the AI boom. It is working on one of the constraints that determines how fast the boom can continue.

The old model treated electricity as a bill.

The new model treats power as infrastructure, architecture, and schedule.

AI has made intelligence cheaper. It has made time more valuable.

Bloom Energy sells time.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

That sounds strange at a moment when model capability is rising so fast. Every few weeks, a new system appears that is cheaper, faster, more open, or better at a task that looked difficult six months ago.

Founders naturally want to attach themselves to the winner.

But the winner keeps changing.

Think about the stack that should matter to every entrepreneur building with AI: an open-source model trained on domain data, frontier models sitting behind a router, and a system that sends each piece of work to the intelligence best suited to handle it. The combination can produce the same or better outcomes at a lower cost than sending every request to the frontier.

This is not just a cheaper way to build an AI product.

It changes where enterprise value lives.

If the underlying models continue to improve and their prices continue to fall, raw intelligence becomes less defensible. Access spreads. Performance gaps narrow. Today’s advantage turns into tomorrow’s API option.

The durable company will not be the one that rents the smartest model first. It will be the one that builds the best system for turning changing models into a better customer outcome.

The model is becoming a component. The learning loop is becoming the company.

The Stack Has Split

The first generation of AI products was built around a simple idea: choose the strongest available model you could afford, place an interface in front of it, add some context, and sell the result.

That was enough when capability was scarce and the model did most of the visible work. It is no longer enough now that intelligence has become a portfolio.

The emerging stack has at least three layers.

At the base is an open-weight worker. It handles the high-volume work: reading, drafting, extracting, classifying, searching, reconciling, and calling tools. It can be hosted with more control, adapted to a domain, and run at a fraction of frontier cost.

Above it sits frontier intelligence. The frontier is not removed. It is used selectively for the hardest reasoning, the ambiguous exception, the final review, or the moment when an error would be expensive.

Between them sits the routing and evaluation layer. It decides which model should perform which task, measures the result, manages context, preserves state, calls tools, escalates uncertainty, and records what happened.

Open models do the volume.

Frontier models supply scarce judgment.

The system decides where each earns its keep.

Fireworks and Harvey recently published a useful demonstration of this architecture on a 100-task slice of Harvey’s Legal Agent Benchmark. Their open-model worker called a frontier model only when it needed advice. According to their reported results, the hybrid system passed 18 tasks completely at a total inference cost of $368, compared with 14 tasks and $954 when the frontier model ran end to end.

A separately post-trained open model also improved its all-pass result at roughly unchanged inference cost. The sample was limited and the work was published by the vendors involved, so it should not be mistaken for a universal law. But the mechanism is the important part: open weights at the core, frontier intelligence called only where it changes the answer.

This turns model selection from a product commitment into an operating decision.

That is a much stronger place to build.

Intelligence Becomes a Variable Input

Software companies once chose a database, a cloud, and a programming language.

Those choices mattered, but customers did not buy the stack. They bought the capability the whole system delivered.

AI is moving toward the same structure, only faster.

A model is an input into production. Different work deserves different intelligence. A routine contract extraction does not need the same reasoning budget as a novel regulatory question. A first-pass support response does not need the same model as a threatened enterprise renewal. A common code migration does not need the same model as a security-sensitive architecture change.

Sending everything to the frontier is like staffing every task with the most expensive expert in the firm. It may feel safe. It does not scale.

The whole reason we built companies in the first place was in recognition of this fact: work can be orchestrated, and there is profit in managing work intelligently.

The router creates a new form of managerial leverage. It can assign cheap cognition to common work, specialized cognition to repeated domain problems, and frontier cognition to the narrow band of tasks where it changes the outcome. Over time, it can learn the boundary.

That boundary is valuable.

Factory, an agent-native software development company, reported that routing work across open and frontier models lowered its average task cost by 30% to 40% in private preview. Its data suggested that roughly a third of tasks needed the frontier, a third could use the cheapest reliable option, and a third belonged somewhere between. The specific percentages will vary by domain, but the economic lesson travels: the cost of intelligence should match the difficulty and consequence of the work.

This is how AI moves from a premium feature to a labor layer.

Companies are beginning to treat tokens as a substitute for labor. The important question is no longer “How do we get employees to use AI?” The important question is “How do we turn tokens into labor at the lowest sustainable cost?”

Tokenomics is the future.

Wealth Systems

Token Economics Will Drive Everything

Brian Armstrong, CEO of Coinbase made an X post recently containing a blueprint for the next business operating system…

Read more

a month ago · 2 likes · Matt McDonagh

When every task requires the most expensive model, the product remains constrained by inference economics. When the system can blend models, improve the cheaper worker, and reserve the frontier for exceptions, the company can automate more work while protecting margin.

The addressable market expands because the cost of serving it falls.

The Moat Moves Into the Loop

Falling model costs are good for builders.

They are also dangerous for weak businesses.

If a product is only a thin interface around one model, every model release threatens it. The vendor may add the feature. A competitor may switch to a better API. The customer may build the workflow internally. Capability rises, but the company captures little of the increase.

The answer is not to predict which model wins.

The answer is to own the assets that become more useful no matter which model wins.

I see five assets that the companies I am building and investing in can compound.

Read more

This piece ends with 3 bold predictions for the next 12-months that you don’t want to miss if you are an investor, or if you are building something.

Sometimes a small technical change reveals the whole shape of the future.

OpenAI recently tested GPT-5.6 Sol on ARC-AGI-3, a benchmark built from unfamiliar 2D puzzle games. The model must explore each game, infer its rules, and improve through action without being told how the game works.

Using the official benchmark harness, GPT-5.6 Sol scored 13.3% on the public task set. OpenAI then changed two settings. It retained the model’s reasoning between actions and replaced rolling context truncation with compaction.

The score rose to 38.3%.

At the same time, output tokens per game fell from roughly 2.9 million to 485,000.

The same model produced nearly three times the score with about one-sixth of the output.

The weights did not change. The training run did not change. The games did not change.

The system around the model changed.

That is the story.

In Agents Per Gigawatt, I argued that the decisive economic metric of the AI era will be how much useful machine labor we can produce from energy. In Scientific Time Dilation, I argued that persistent AI systems will compress discovery by running more scientific work in parallel and feeding each result into the next attempt.

OpenAI’s ARC-AGI-3 experiment connects those two ideas.

Agents per gigawatt is about how much cognition a system can afford to run.

Scientific time dilation is about whether that cognition compounds.

The harness determines both.

The Model Is Not the System

We all still talk about AI capability as if it lives entirely inside the model.

A new model is released. A benchmark score goes up. A parameter count leaks. A training cluster grows. We compare one set of weights against another and treat the result as a clean measurement of intelligence.

But a model is not an agent any more than an engine is a car.

The model supplies the reasoning capacity. The surrounding system supplies memory, context, tools, permissions, feedback, and continuity. It decides what the model can see, what it can do, what it remembers, when it retries, and whether a useful lesson survives long enough to shape the next action.

That surrounding system is the harness.

We wrote in Agents Per Gigawatt that the harness is the agent. OpenAI has now given us one of the cleanest demonstrations of that claim.

The official ARC-AGI-3 harness was intentionally generic. That makes sense from the benchmark designer’s perspective. A common interface can make comparisons cleaner and expose what a model can do without a stack of custom tools.

But generic does not mean neutral.

In the official setup, the model’s private reasoning was discarded after each game action. GPT-5.6 Sol could see previous moves and brief notes, but it could not see the thinking that produced them. It had to reconstruct its understanding of the game again and again.

The harness also used a rolling truncation window. Once the history grew too large, the oldest actions disappeared. The agent lost not only its prior reasoning, but eventually parts of its experience.

Imagine evaluating a human by wiping their working memory after every move and tearing pages from their notebook whenever it became too thick.

You would not be measuring the person in isolation. You would be measuring the person inside an amnesia machine.

The same problem applies to agents.

Intelligence Needs Continuity

A single response can be impressive without memory. A long task cannot.

Long-horizon work depends on continuity. The system must remember what it tried, why it tried it, what happened, which assumptions failed, what patterns emerged, and what the current plan is. Without that chain, every new action begins too close to zero.

This is why retained reasoning mattered.

When GPT-5.6 Sol could carry its prior thinking forward, it no longer had to reinterpret the game from scratch after each action. It spent less time recreating old insight and more time applying what it had learned. Its behavior became more coherent because its strategy had a history.

Compaction solved the second problem.

A long-running agent cannot keep every raw observation in active context forever. The record becomes too large, too expensive, and too noisy. Rolling truncation solves the size problem by throwing away the oldest material. But time is not a reliable measure of importance. The first observation may contain the rule that makes the hundredth action intelligible.

Compaction takes a different approach. It compresses the history into a smaller representation that preserves the information needed to continue. Done well, it keeps the map while discarding the footsteps.

These two settings create a simple operating loop:

Act. Remember. Compress. Continue.

That loop is more important than another isolated burst of intelligence. It turns cognition from a sequence of disposable moments into a cumulative process.

The agent does not think.

It learns across the work.

More Work From Every Token

The score improvement will get most of the attention. The efficiency gain may matter more.

Under the official harness, the max-effort run used about 2.9 million output tokens per game to score 13.3%. With retained reasoning and compaction, it used about 485,000 output tokens per game to score 38.3%.

17x Improvement

A crude score-per-output-token calculation suggests an improvement of roughly seventeen times.

That is not a full economic measure, and the benchmark score should not be treated as linear units of value. But the direction is unmistakable: the better harness produced far more useful progress from far less inference.

This is agents per gigawatt at the scale of a single workflow.

Every time an agent forgets its plan, the data center pays for it to think again. Every time a system reloads a bloated history, it spends compute attending to information that may no longer matter. Every time a weak orchestration layer sends a capable model down the same failed path, energy becomes waste heat instead of useful cognition.

Model efficiency matters. Better chips matter. Cooling, networking, and power generation matter.

Workflow architecture matters too.

The effective supply of intelligence can rise even when the physical supply of compute stays fixed. Preserve useful state. Compress context. Route work to the right model. Cache stable results. Give agents the tools to verify their own output. Stop bad attempts early. Reuse the learning from one run in the next.

These are not cosmetic optimizations. They determine how many competent workstreams a gigawatt can support.

The AI economy will not be won only by whoever builds the largest cluster. It will also be won by whoever wastes the least cognition.

Time Dilation Requires Memory

In Scientific Time Dilation, I described a future where AI closes the loop between science and technology.

An agent reads the literature, finds a gap, generates hypotheses, builds simulations, proposes experiments, studies the results, and updates its theory. Better science creates better tools. Better tools improve the next cycle. Discovery begins to accelerate the machinery of discovery.

But parallelism alone does not create that future.

A million agents that cannot remember what they learned can produce a million attempts and still fail to compound. They may repeat the same mistakes, rediscover the same facts, and flood the system with plausible but disconnected work.

Scientific time dilation requires institutional memory at machine speed.

A useful research agent must preserve more than its final answer. It needs the shape of the search: rejected hypotheses, failed experiments, unexplained anomalies, uncertainty, provenance, and the reasons one path was chosen over another. It must compress that history without erasing the inconvenient result that later becomes the breakthrough.

This is the deeper meaning of the ARC-AGI-3 result.

The benchmark games are small. The operating principle is not.

An agent learning the rules of a puzzle and a research system learning the rules of a material share the same basic problem. Both must act under uncertainty. Both must use feedback. Both must carry insight across time. Both become more capable when the result of one attempt changes the next.

Memory turns attempts into experience. Compaction turns experience into usable context. Evaluation turns context into better decisions.

Combined, they turn raw inference into a discovery loop.

The Hidden Frontier Is Architecture

AI progress now has at least two frontiers.

Read more

More on what we’re launching in a moment! Let’s focus first on what’s happening in technology that makes launching an AI-Native Media Company a smart move.

The most important fact about artificial intelligence is not that machines are becoming more capable.

It is that human beings can become radically more capable with them.

An individual can now access intelligence, software, research, automation, and economic leverage that once required an entire institution. A small team can command resources that would have been unavailable to a Fortune 500 company a decade ago.

This changes who can build, and it also changes what can be built.

It changes how quickly an idea can become a system capable of producing real value.

But access to intelligence is not the same as agency.

AI can generate options. It cannot supply conviction.

It can analyze a market. It cannot decide what you are willing to risk.

It can build a system. It cannot choose the life that system should serve.

The machine multiplies the operator. That makes the quality of the operator more important, not less.

The Operator Behind the Machine

On the surface, my career looks like a sequence of increasingly successful bets.

I launched a hedge fund.

The fund failed, so I pivoted into technology investing.

I invested in a cybersecurity company that was eventually sold to Facebook.

I built RevSystems into a premier AI consultancy.

Today, I have investments in more than twenty-five companies and access to an extraordinary network of brilliant, passionate people.

That is the clean version.

It’s also incomplete.

I was twenty-five years old when I launched the fund. I was surrounded by people with more experience, more capital, better credentials, and deeper networks.

I dealt with an incredible amount of imposter syndrome. I constantly wondered whether I had earned the right to be in the room.

Then the fund failed.

The private fear that I might be a fake suddenly had public evidence.

At least, that was how it felt.

Failure did not just damage the business. It multiplied every doubt I already carried.

I moved into technology investing.

The first three companies I backed went to zero.

Three decisions. Three failures. Three more pieces of evidence for the prosecution.

Then one of the cybersecurity companies I invested in attracted Facebook. It should have been a clean breakthrough.

Instead, I came close to being ripped off the cap table.

Even success arrived carrying another test.

This is how life actually works. The breakthrough and the crisis travel together. Opportunity does not wait for your fear to disappear. It arrives while you are still deciding whether you belong.

The Window

The most important window is the universe is the small window between stimulus and response.

Something happens.

The fund fails. The investment goes to zero. The deal turns against you. The company breaks. The technology changes the rules.

Then there is a window.

It may last a month. It may last five minutes. Sometimes it lasts only long enough to choose whether you will act from fear or from agency.

You do not control the stimulus.

You control what you build in response.

Courage is not the absence of fear. Courage is the decision you make while fear is still present.

Confidence usually arrives later.

Confidence is the receipt generated by action.

You are brave in the window.

You are consistent after it.

You are resilient when the result is not what you wanted.

Those three forces compound.

AI Compresses the Window

Technological change creates the same window at a much larger scale.

The old tools are breaking. The old jobs are changing. The old assumptions about what a person, company, or institution can do are becoming unreliable.

AI compresses the distance between stimulus and response.

A new model launches.

A capability that seemed impossible becomes cheap.

A workflow that required thirty people can suddenly be run by three.

An industry reorganizes around a new cost structure.

A competitor begins operating at a speed your organization was not designed to match.

The window opens.

Some people wait for certainty.

Others act.

They test the tools. They rebuild their workflows. They connect models to memory, software, permissions, data, and evaluation systems. They turn AI from an occasional chatbot into an operating layer.

They learn while everyone else is debating whether the change is real.

The advantage does not come from having access to the model. Access will become increasingly universal.

The advantage comes from what you build around it.

Judgment.

Systems.

Context.

Taste.

Relationships.

Capital.

The ability to decide what deserves to exist.

AI amplifies all of these things. It also amplifies their absence.

A capable operator becomes more capable. A confused operator produces confusion faster.

Adversity Is Expensive Information

Failure is information delivered at maximum volume.

It shows you where the system is weak. It exposes bad assumptions. It reveals who can be trusted. It separates real conviction from borrowed confidence.

None of this means failure should be romanticized. Losing hurts. Betrayal hurts. Being wrong hurts.

Pain is not automatically productive.

You must convert it.

Adversity becomes valuable only when you turn it into better judgment, stronger systems, and more capable action. Otherwise, it is merely damage.

The fund taught me lessons about risk that success could not have taught me.

Those first failed investments forced me to rebuild the way I evaluated companies, founders, markets, and my own judgment.

The Facebook deal taught me that creating value is not enough. You must understand incentives, ownership, power, and the machinery surrounding the value you create.

Every hard experience became part of a new operating system.

That system eventually became leverage.

Technology works the same way.

The tool is not the operating system.

The model is not the operating system.

The operator learns, adapts, integrates, and builds the operating system.

Why Sovereign Singularity Media Exists

I am excited to announce that Life in the Singularity has joined Sovereign Singularity Media.

Sovereign Singularity Media is built around a simple belief:

The greatest opportunity created by technological change is the opportunity to become more capable.

Life in the Singularity studies the future we are entering and how to build with artificial intelligence.

Wealth Systems explores how knowledge, capital, technology, and disciplined action can produce autonomy.

Mastering Revenue Operations examines how value becomes repeatable revenue through good systems.

These are not separate subjects. They are parts of the same architecture.

You must understand the world.

You must develop the capability to build within it.

You must create systems that turn judgment and effort into durable value.

You must capture enough of that value to preserve your freedom to act.

That is sovereignty.

Not complete independence. Not isolation. Not the fantasy that nothing can hurt you.

Sovereignty is the ability to absorb change without surrendering your agency.

It is the capacity to take a hit, learn faster than the damage compounds, and build something stronger in response.

What Life in the Singularity Is Becoming

Life in the Singularity will remain focused on the future.

But the publication will move deeper into the systems required to operate inside that future.

AI agents.

Human-machine coordination.

Digital twins.

Workflow architecture.

New forms of individual and organizational leverage.

The changing economics of knowledge work.

The systems surrounding the model: memory, tools, context, data, permissions, evaluation, and human judgment.

The public essays will continue to explain the major shifts.

Paid membership will go further.

Paid members will receive deeper research, implementation frameworks, operating models, and practical tools for turning technological capability into action.

The promise is not more information. There’s too much of that already.

The promise is greater capability.

Subscribe now

Two Publications. One Operating System.

Understanding the future is not enough.

You must also build your position within it.

Life in the Singularity helps you understand what is changing, what new capabilities are becoming possible, and how to build with artificial intelligence.

Life in the Singularity

Build the future with AI.

By Matt McDonagh

Wealth Systems helps you turn those capabilities into capital, ownership, resilience, and personal sovereignty.

Wealth Systems

Build wealth systems to power your life.

By Matt McDonagh

One studies the emerging world.

The other helps you own more of the outcome.

Life is the capability layer. Wealth is the capital layer.

Paid membership in either publication provides the deeper research and practical systems associated with that publication.

Paid membership in both unlocks the singularity where they converge.

The Sovereign Operator Briefing

The Sovereign Operator Briefing is where Life in the Singularity and Wealth Systems become one operating system.

Beginning in August 2026, paid members of both publications receive a monthly briefing that translates the most important shifts across AI, markets, business, and capital into decisions:

• What changed?

• Why does it matter?

• Where is the leverage?

• What could break?

This is not another roundup of headlines.

It is a decision tool for people building companies, deploying capital, and designing more sovereign lives.

Dual members will also receive members-only conversations with founders, investors, builders, and experts from across my network. Interview series and other formats (a video podcast) are in development across Sovereign Singularity Media. All with the people operating inside the changes we are studying.

Life helps you see the future. Wealth helps you build your position in it. The Sovereign Operator Briefing connects the two.

If you want to understand the future and build with artificial intelligence, become a paid member of Life in the Singularity.

If you want to turn those capabilities into capital, ownership, and autonomy, become a paid member of Wealth Systems.

Wealth Systems

Build wealth systems to power your life.

By Matt McDonagh

If you want the complete architecture, join both.

See the future. Build with it. Own the upside.

What We Believe

We believe agency matters more than certainty.

We believe systems outperform intentions.

We believe technology should increase human capability, not weaken human judgment.

We believe intelligence becomes valuable when it is converted into action.

We believe capital is stored choice.

We believe adversity can become leverage, but only when it is converted.

We believe consistency is one of the most powerful forces available to an individual.

We believe the best networks are built from brilliant, passionate people who create, challenge, teach, and compound together.

We believe the future belongs to creators and operators.

People willing to see clearly.

People willing to decide.

People willing to build.

People willing to remain in the arena after the first plan fails.

The Future Is Built Through Action

AI is creating an extraordinary abundance of intelligence.

It will not create an abundance of courage.

It will not create an abundance of judgment.

It will not decide which risks are worth taking, which systems deserve to be built, or what kind of future we should create.

That work still belongs to us.

Technology expands the surface area for action. It gives individuals more leverage, small teams more power, and capable operators a greater ability to shape the world around them.

But the window remains.

Something changes.

The opportunity appears.

Fear arrives.

Then you decide.

Be brave in the window.

Build the system.

Compound the response.

The machine multiplies the operator.

Make yourself worth multiplying.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

They have stories about workflows.

The real process lives across documents, software, meetings, inboxes, habits, and the heads of experienced employees. The official procedure says one thing. The work itself moves through a web of judgment calls, hidden prerequisites, remembered exceptions, and manual handoffs.

That arrangement was good enough when humans were the only intelligence available to the firm. A person could read a policy, infer what it meant, ask a colleague what was missing, remember where the work stopped, and carry the process forward.

AI exposes how fragile that system always was.

Give an agent a long document and tell it to follow the rules, and it may perform brilliantly. It may also skip a check, lose its place, take an action too early, or confidently continue after the task should have stopped. The failure looks like a model problem. Often it is an operating-model problem.

The company gave the machine prose when the machine needed a program.

That is the next shift in the future of work. We are moving beyond AI as a tool employees use and toward AI as a labor layer the company must govern. To make that layer useful, work has to become more explicit, more modular, more testable, and more executable.

The future of work has a runtime.

The future of work is runtime.

From Workflow to Work Program

In AI Leverage Is an Operating Model Problem, I argued that the unit of change is not the prompt. It’s the workflow.

Every company now has access to capable models. Access is no longer the scarce thing. The scarce thing is knowing where intelligence belongs, how the work should change, what judgment must remain human, and what platform needs to exist underneath the process so the gains compound.

Then, in Productive Time Dilation, I pushed the argument further. Intelligence is becoming infrastructure. Labor is becoming programmable. The people who own systems that turn machine intelligence into repeated economic output will capture more of the surplus than people who merely use AI inside someone else’s system.

The system is the wealth.

A new paper makes the mechanism behind those claims much more concrete. In Compile, Then Page: Executable SOP Programs and a Capability-Gated Runtime for Procedural LLM Agents, Chenglin Yu and his coauthors study what happens when standard operating procedures stop being treated as text and start being treated as programs.

The researchers begin with a familiar enterprise problem. Banks, clinics, service desks, and other high-consequence operations depend on long procedures with conditional paths. Verify this before doing that. Ask for missing information. Refuse the request if a required condition fails. Use one tool if the user is authenticated and another if they are not.

The normal approach is to place the whole procedure in the model’s context and hope it follows the instructions.

The paper tests a different approach. A deterministic compiler turns the procedure into pseudocode. Each user goal becomes a process function. Each rule becomes a subroutine with a specific verification method. Each decision returns not only a verdict, but the evidence supporting it.

The agent does not simply receive a policy.

It enters a work program.

That distinction is larger than it sounds. A policy tells the worker what should be true. A program specifies what to check, in what order, which conditions open or close a path, what evidence must exist, where the process should resume, and when the system must stop.

This is not prompt engineering.

It is workflow architecture.

Prose Was Built for Human Labor

The modern company runs on text because text is flexible.

Job descriptions, playbooks, policies, handbooks, process maps, strategy documents, and meeting notes all assume a human reader. They leave room for inference because people are good at filling gaps. The experienced operator knows which rule matters, which database is reliable, which exception is real, and which step can safely be skipped.

That flexibility carries a hidden cost. The organization becomes dependent on interpretation.

Two employees can read the same procedure and execute it differently. A new employee needs months of repetition before the document becomes practical judgment. When an expert leaves, part of the operating system leaves with them. Managers compensate through training, supervision, meetings, and institutional memory.

AI does not remove the cost of ambiguity. It scales it.

A machine can read every policy in the company and still fail at the exact moment the policy must become action. The model has to translate language into state: where am I, what has already happened, what remains unresolved, which tool should I use, what did the result prove, and what am I allowed to do next?

That translation is the work.

The paper’s compiler makes the hidden structure explicit. It preserves sequence. It distinguishes an ordinary choice from a gate that should stop at the first valid path. It moves authentication checks in front of actions that require authentication. It binds every conclusion to evidence. It breaks large procedures into smaller functions that can be called and completed.

This is what companies will have to do with their own operations.

Not every sentence needs to become code. Not every human judgment should be reduced to a Boolean rule. But any process assigned to machine labor needs enough structure to run, enough context to act, enough state to continue, and enough evidence to be reviewed.

The SOP becomes software.

The Active Frame Is the New Unit of Management

Compiling the procedure is only the first half of the paper’s idea.

A complete program can still be too large and distracting. If the agent sees every instruction at every moment, the current step competes with the rest of the process, the conversation, the tool history, and the user’s latest request.

The researchers solve this with a stack-based runtime. The runtime keeps track of the call stack, the current position, the variables, and the results returned by prior checks. It then “pages” only the active frame into the model’s immediate context.

In plain language: show the model the part of the job it needs now, preserve everything else outside its attention, and restore the right state when the current subtask is complete.

You don’t think about every single part of your job while working, do you? Why should the AI?

The model remains the semantic executor. It speaks with the user, chooses tool arguments, reads evidence, and makes local judgments. But it no longer has to hold the entire procedure and its current location inside a rolling conversation.

The system remembers where the work is.

This matters because attention is part of the production function. Giving a model more information is not the same as giving it the right information at the right time. A large context window is storage. It is not management.

Humans already understand this intuitively. A great manager does not walk into every meeting and recite the entire corporate strategy, employee handbook, customer history, and project plan. The manager brings the relevant objective forward, clarifies the current decision, preserves the broader context, and keeps the work moving.

Agents need the same discipline, but encoded in the system.

The active frame becomes a new unit of management. It is the bounded piece of work the agent is authorized and equipped to perform now. It contains the immediate objective, the required inputs, the available tools, the relevant constraints, the expected evidence, and the definition of done.

This changes how we should think about jobs.

A job is a bundle of responsibilities wrapped around one human. A work program is a graph of objectives, checks, decisions, actions, and exceptions that can move across humans and machines.

The job was the old container.

The active frame is the new one.

Management Becomes Architecture

The first industrial managers allocated bodies, machines, and time.

The next generation of managers will allocate human and machine cognition across a live graph of work.

That means management decisions will increasingly be expressed as architecture. Which system gets the task? What context can it see? What tools can it use? What action requires approval? What evidence must come back? What happens when a check fails? Which model is capable enough for the procedure? Where does the work pause? Who owns the exception?

These questions used to live in training and supervision. Now they also live in permissions, schemas, workflow definitions, evaluation suites, and runtime controls.

The org chart will not disappear. It will be joined by the workflow graph.

The workflow graph will show how intent becomes action. It will include human employees, agents, databases, APIs, approval gates, memory, policies, and review queues. Some nodes will generate. Some will verify. Some will decide. Some will act. The highest-risk nodes will remain hard-gated by software or a human.

This is where the future-of-work debate usually becomes too shallow. We ask whether a model will replace an accountant, lawyer, analyst, salesperson, or software engineer. But jobs are not atomic. They are political and administrative packages assembled around the old cost of coordination.

AI will not simply replace those packages one at a time.

It will unbundle them.

Research may move to one system. Drafting to another. Reconciliation to a third. A human may handle the exception, the client conversation, and the final judgment.

The work will be recomposed around capability, risk, cost, and trust.

Some jobs will shrink. Some will become more powerful. New roles will appear around AI taste testing, workflow design, context engineering, evaluation, model operations, agent supervision, data quality, and exception handling. The important divide will not be between people who use AI and people who do not.

It will be between people who remain trapped inside tasks and people who can design the systems that perform them.

Capability-Gated Autonomy

The paper includes an important warning for anyone who believes more agent scaffolding is always better.

Recent evidence says it isn’t.

On one bank-task evaluation, converting the official procedure from prose into a compiled representation raised one model’s pass rate from 70.4% to 86.4%. Adding the runtime raised it again to 92.8%, with perfect refusal accuracy on the screened refusal set. Across seven domains, two stronger models benefited from the program-guided runtime.

Weaker models were harmed.

For two of them, the runtime reduced performance by 14 to 26 points on the bank tasks. The problem was not that they could never reconstruct the state. When directly tested, they could. The problem was spontaneous discipline during execution. One model under-attended to the active frame. Another kept verifying after the goal had already been achieved.

The system exposed a capability boundary.

This is a critical lesson for the future of work. Autonomy is not a product setting. It is an earned property of a specific system performing a specific class of work under specific controls.

The right question is not, “Are agents ready?”

Ready for what? With which model? Inside which workflow? Using which data? With what tools? Under which permissions? Evaluated against what standard? At what cost of failure?

The paper’s deployment rule is simple: compile first, then page only after a model-level discipline check. Even then, the runtime provides soft enforcement. It makes deviations visible, but it does not prevent every bad action. Irreversible actions still need hard controls at the tool layer.

That is how serious companies should approach autonomy.

Drafting can be broad. Recommendations can be bounded. Reversible execution can expand with evidence. High-consequence actions should remain permissioned, observable, and interruptible.

Autonomy should rise with demonstrated reliability.

Judgment Becomes Even More Valuable

When execution becomes cheap, judgment becomes expensive.

The model can generate the report, inspect the records, prepare the options, and run the checks. But someone still has to decide what goal matters, what tradeoff is acceptable, what evidence is sufficient, and what should enter reality.

This moves human value up the stack, it doesn’t erase it entirely as AI alarmists claim.

The most valuable employees will know how to define objectives, decompose work, supply context, set constraints, inspect evidence, resolve ambiguity, and recognize when a plausible output is wrong. They will understand the domain well enough to supervise machine labor and the system well enough to improve the workflow.

Taste matters because output becomes abundant.

Trust matters because execution becomes autonomous.

Agency matters because someone has to decide what the machine should do.

This does not mean every employee becomes a philosopher-manager floating above the work. Domain expertise still comes from contact with reality. A lawyer cannot supervise legal work they do not understand. An operator cannot verify a process they have never run. A manager who only reads agent summaries will slowly lose the ability to detect when the system is drifting.

The apprenticeship problem is real.

Junior employees historically built judgment through repetition. They drafted the memo, checked the numbers, updated the CRM, prepared the analysis, made mistakes, and watched experienced people correct them. If agents absorb the entry-level work, companies must create a new path for humans to develop the judgment required to supervise it.

The best organizations will not remove people from the work entirely. They will design deliberate learning loops: simulation, shadow execution, sampled review, failure analysis, rotation through exceptions, and direct exposure to customers and consequences.

You cannot govern what you no longer understand.

The Economics of Executable Work

Once work can be compiled, routed, run, checked, and improved, the economics of the firm change.

Headcount becomes a poor proxy for capacity. A small team with strong workflows can operate more surface area than a large company built on meetings and manual coordination. The minimum viable organization shrinks while the maximum viable ambition of an individual expands.

But the deeper change is that the workflow becomes an asset.

The first time a task runs, the system produces an output. The tenth time, the team has a repeatable process. The hundredth time, it has execution history, edge cases, evaluations, trusted data paths, permission boundaries, cost curves, and a growing map of what works.

That operational knowledge compounds.

The prompt does not.

This is why ownership matters. If you own the workflow, the customer relationship, the proprietary context, the feedback, and the asset produced by the process, machine labor can create durable leverage. If you merely perform tasks inside someone else’s runtime, you may become more productive without capturing the surplus.

The industrial economy rewarded owners of factories. The software economy rewarded owners of applications and networks. The agentic economy will reward owners of executable operating systems.

That can mean a company’s internal platform. It can mean a vertical AI business built around trusted execution. It can mean a single high-agency operator who has encoded enough context and workflow to function like a small institution.

Compile the Company

Most companies are not ready to turn every process into an agent workflow. That’s fine. The work begins with a single valuable loop.

Find a process that repeats, consumes meaningful human time, has a clear economic outcome, and can be observed. Map what actually happens, not what the official document claims happens. Identify the inputs, the decisions, the tools, the hidden prerequisites, the failure states, the exceptions, and the definition of done.

Then make the work executable.

Turn vague policies into explicit conditions. Tie each condition to a way of checking it. Separate reversible actions from irreversible ones. Move state out of the conversation and into a system that can preserve it. Bring the current objective forward. Require evidence. Record the trace. Test the workflow against cases that should succeed and cases that should refuse.

Do not start with maximum autonomy.

Start with visibility.

Let the agent prepare the work while a human approves it. Measure where it succeeds, where it stalls, where it invents, and where the procedure itself is ambiguous. Turn failures into evaluations. Turn corrections into context. Turn repeated decisions into rules.

Expand autonomy only when the evidence supports it.

Just like with human employees.

The paper also reveals a tradeoff that executives should not ignore. Its paging method improved accuracy and traceability, but it used more model calls and more total prompt tokens. Better systems do not always look cheaper at the level of one task. The return comes from fewer errors, safer refusal, clearer audits, faster scaling, and the ability to move more work through the organization without equivalent coordination cost.

Optimize the business system, not the token bill.

In fact, tokens are a poor way to measure inputs or outputs.

The company that does this well will become easier to run because it becomes easier to see. Work will have state. Decisions will have evidence. Policies will have executable structure. Exceptions will have owners. Every important workflow will produce both an outcome and a trace.

That is not just automation inside a black box.

It is a more legible firm.

The Future Belongs to the Architects

The future of work will not be defined by a single model, a universal digital employee, or one dramatic moment when every job changes.

It will arrive workflow by workflow.

A policy becomes a program. A program gains tools. The tools gain permissions. The work gains memory. The system gains evaluations. The agent earns more autonomy. The human moves from direct execution toward direction, review, exception handling, and ownership.

Over time, the company stops treating intelligence as something employees occasionally access and starts treating it as infrastructure that work runs on.

That’s the shift we are seeing happening.

The old company described how work should happen and relied on people to reconstruct the process.

The new company will encode how work happens, preserve its state, expose the right instruction at the right moment, and learn from every execution.

The job description was a document.

The future of work is a runtime.

Imagine the businesses people are going to build with this?

Wealth Systems

Build wealth systems to power your life.

By Matt McDonagh

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

AI is closing the loop between science and technology. The result will not just be faster invention. It will be an acceleration in the rate of acceleration itself.

We are beginning to run centuries of science in parallel.

That sounds like hyperbole until you look at what scientific work actually is. Read the literature. Find a gap. Form a hypothesis. Build a model. Design an experiment. Run it. Study the result. Change the hypothesis. Try again.

Every step in that chain has historically been constrained by human time.

A researcher can read only so many papers.

A lab can run only so many experiments.

An engineer can test only so many designs.

A scientific institution can pursue only so many ideas before money, attention, equipment, or patience runs out.

In my last essay, I called the ability to command parallel machine labor productive time dilation. One person can now operate workstreams that would have once required a large team. The human still experiences a normal day, but the system working on their behalf can produce hundreds or thousands of hours of output inside it.

That same force is now moving into science.

Productive time dilation compresses work.

Scientific time dilation compresses discovery.

And discovery has a property that ordinary work does not. A completed task ends. A scientific breakthrough becomes a tool for producing the next breakthrough. Better science creates better technology. Better technology creates better instruments. Better instruments produce more data. More data creates better science.

AI is entering every part of that cycle at once.

This is the real intelligence explosion. It is not simply a model getting smarter inside a data center. It’s intelligence entering the machinery of progress and increasing the speed at which civilization learns how reality works.

Science and Technology Have Always Fed Each Other

We talk about science and technology as if they are separate domains.

Science discovers. Technology builds.

That distinction is useful, but incomplete. In practice, the two have always formed a reinforcing loop.

The microscope opened biology. Biology improved medicine. Medicine created new instruments, new datasets, and new questions about the body. The transistor made modern computing possible. Computing gave scientists the ability to simulate systems that could never be calculated by hand. Those simulations helped engineers design better chips, which created more computing power.

Science turns the unknown into knowledge.

Technology turns knowledge into capability.

Capability becomes a new instrument for exploring the unknown.

That loop is the engine of modern civilization. Clean water, antibiotics, electricity, computation, modern agriculture, and spaceflight all sit downstream of it.

But the loop has always moved at biological speed. Humans had to carry knowledge from one stage to the next. Someone had to read the papers, make the connection, write the code, build the instrument, run the test, interpret the output, and persuade an institution to fund another cycle.

AI changes the clock speed of the loop.

AI Closes the Discovery Loop

Most discussion about AI and science focuses on isolated abilities. A model can summarize a paper. Predict a protein structure. Write analysis code. Suggest a molecule. Optimize an equation.

Each ability matters. But the larger shift appears when they are connected.

An AI system can search the literature, compare findings across fields, generate candidate explanations, turn those explanations into simulations, rank the most useful experiments, send instructions to laboratory equipment, analyze the results, and use those results to decide what to try next.

That is not a better search engine.

It’s a discovery system.

The architecture looks familiar because it is the scientific method turned into a persistent workflow:

Observe. Hypothesize. Model. Test. Measure. Update. Repeat.

Models provide reasoning. Databases provide memory. Simulators provide cheap experimental environments. Lab robotics reaches into the physical world. Instruments return measurements. Humans define the objective, inspect the evidence, and choose which paths deserve reality.

The individual pieces are uneven. Models still hallucinate references. Robots struggle with tasks a graduate student handles without thinking. Experimental data is messy. Many scientific questions do not have clean evaluation functions. A plausible theory can survive thousands of lines of elegant analysis and still be wrong.

But the direction is clear.

In 2026, a system described in Nature could move through an end-to-end machine-learning research cycle: generating ideas, writing code, running experiments, analyzing results, drafting a paper, and performing automated review. It was not replacing the best scientists. Its results were inconsistent, and a workshop paper is not a fundamental discovery. But the chain had been connected.

Once the chain is connected, every improvement to any component raises the value of the whole system.

Better reasoning improves the hypotheses. Better tools improve execution. Better instruments improve the data. Better evaluations improve selection. Better memory prevents the system from repeating old mistakes.

The loop learns how to loop.

I’ve been working on an autonomous research lab since the end of last year.

The Cost of Asking Nature Collapses

In the old model of science, experiments were precious.

Before spending months in the lab, a researcher had to choose a narrow path. Before using an expensive instrument, a team had to decide which question mattered most. Before synthesizing a material or testing a drug candidate, someone had to reduce an enormous possibility space to a handful of plausible options.

Scarcity forced scientists to make large bets with limited information.

AI changes the economics of the search.

It can screen millions of candidates in simulation before a physical experiment begins. It can run many competing models instead of asking one theory to carry the entire burden. It can search combinations that human intuition would never prioritize because the space is too large and the relationships are too strange.

This is already visible in biology. The AlphaFold Protein Structure Database provides open access to more than 200 million predicted protein structures. That does not solve biology. A predicted structure is not a cure, and living systems remain staggeringly complex. But it turns a once-scarce form of analysis into widely available research infrastructure.

The same pattern is appearing in materials science. Google DeepMind’s GNoME system identified 2.2 million candidate crystal structures and 381,000 new entries on an updated stability frontier. A connected autonomous lab then combined computation, knowledge extracted from the literature, machine learning, and robotics to synthesize 36 target materials during 17 days of continuous operation.

Prediction fed experiment. Experiment created evidence. Evidence improved the next decision.

That’s the loop in action.

The important number is not how many papers an AI can write. It is how many useful questions a scientific system can ask nature, how quickly nature can answer, and how much the system learns from every response.

When the cost of asking falls, science changes from a sequence of carefully rationed attempts into a portfolio of continuous exploration.

The Lab Becomes a Learning Machine

A traditional lab contains knowledge, but much of that knowledge is trapped in people.

It lives in the hands of the technician who knows that a sample looks wrong. It lives in the postdoc who remembers why an experiment failed two years ago. It lives in handwritten notes, undocumented adjustments, old spreadsheets, and the judgment of people who eventually leave.

The autonomous lab converts more of that tacit process into a system.

Every experiment becomes part of a structured memory. Every failed attempt can update the next one. Every instrument can feed a shared model. The lab stops being a place where isolated experiments happen and becomes a machine that gets better at experimentation.

This makes negative results more valuable.

Failed experiments often disappear because journals and institutions reward novelty more than useful dead ends. An intelligent research system can learn from every failed synthesis, unstable design, bad reaction, or broken hypothesis.

Failure becomes training data.

That changes the compounding structure of research. A lab with the best memory of what did not work may move faster than a lab with the largest archive of polished papers. The moat is not simply proprietary data. It is a closed loop between data, action, feedback, and improved judgment.

The winning research organization will not be the one that produces the most activity. It will be the one that learns the most from each cycle.

Intelligence Starts Improving Intelligence

In Living Through An Intelligence Explosion, I argued that the decisive threshold arrives when silicon begins helping to improve silicon. Once AI can automate meaningful parts of AI research, the system that produces intelligence becomes a beneficiary of the intelligence it produced.

That loop is no longer theoretical.

AI systems already help write training code, optimize data-center operations, design chips, discover algorithms, generate synthetic data, and evaluate other models. DeepMind’s AlphaEvolve has been used to improve algorithms involved in data centers, chip design, and AI training, including parts of the computing stack beneath systems like itself.

Better AI helps create better algorithms.

Better algorithms use compute more efficiently.

More efficient compute makes better AI cheaper to train and run.

Better AI then searches for the next improvement.

This does not mean capability rises without limits or that every recursive loop explodes overnight. Physical constraints still matter. Energy, fabrication, capital, data quality, and the stubborn difficulty of new ideas remain real. But the research process now contains a positive feedback channel that did not exist at this scale before.

And AI research is only the inner loop.

The larger loop runs through the physical economy. AI can help discover better cooling systems, power electronics, battery chemistries, superconducting materials, optical components, and manufacturing methods. Those technologies can improve the infrastructure used to build and run AI. The improved AI can then attack the next set of scientific constraints.

Materials improve chips. Chips improve intelligence. Intelligence improves materials.

Energy improves compute. Compute improves science. Science improves energy.

Robotics improves experiments. Experiments improve models. Models improve robotics.

Science and technology are not merely accelerating beside each other. They are beginning to pull each other forward.

Quantum Gives the Loop a New Gear

AI is not the only new engine entering the discovery system.

Quantum computing is approaching the loop from the other side.

A quantum computer is not a faster classical computer. It will not make every spreadsheet, simulation, or AI model run instantly.

Its importance is more specific: Quantum changes which problems can fit inside a computer.

What is Quantum?

Nature is quantum. Molecules, materials, chemical reactions, superconductors, and the machinery of life emerge from interactions that become brutally expensive to represent with classical machines. As the number of interacting particles rises, the space required to describe them can grow beyond anything a conventional computer can hold.

Quantum computers are built from the same underlying physics.

That creates the possibility of simulating parts of nature in nature’s native computational language.

The implications reach directly into drug design, materials science, energy, chemistry, and fundamental physics. An AI system could generate candidate molecules, design the quantum circuits needed to model them, allocate each part of the calculation between classical and quantum processors, interpret the results, and send the most promising candidates into an automated laboratory.

1. AI supplies the reasoning.

2. Quantum supplies a new experimental instrument.

3. Robotics supplies the hands.

The closed loop supplies the memory.

The relationship also runs in the opposite direction because quantum hardware is extraordinarily difficult to operate. Qubits drift. Noise accumulates. Errors destroy useful information. Control systems must continually tune thousands of parameters while error-correction systems infer what went wrong.

AI is beginning to absorb that complexity.

Google DeepMind’s AlphaQubit used a neural network to decode quantum errors more accurately than previous leading methods. In July 2026, Google researchers reported another step: a reinforcement-learning system that continuously adjusted control parameters on the Willow processor and improved the logical stability of its quantum memory 3.5-fold.

AI is helping quantum computers remain coherent long enough to become useful.

Quantum computers may eventually return the favor by opening scientific and computational spaces that classical AI systems cannot efficiently search.

We can already see the outline. Google’s Quantum Echoes experiment took roughly two hours on Willow and was estimated to require 13,000 times longer on a classical supercomputer. That was a specific quantum-physics task, not a general-purpose productivity benchmark. But that specificity is the point.

For the right problem, the difference is not twenty percent faster.

It is reachable versus unreachable.

Now place that capability inside the larger discovery architecture. AI generates the hypothesis. Classical computers coordinate the workflow. Quantum processors explore the hardest physical state spaces. Robotic labs test the result. Experimental evidence returns to the models. The system learns and begins again.

AI improves quantum.

Quantum expands science.

Science improves the materials, energy systems, sensors, and fabrication methods used to build better AI and quantum computers.

The loop gains another gear.

Every Field Becomes Upstream of Every Other Field

The deepest breakthroughs often happen when knowledge crosses a boundary. A method from physics unlocks biology. A materials advance changes energy. Better sensors create new medical data.

Human institutions are bad at searching these combinations. Knowledge is divided into disciplines, journals, departments, conferences, budgets, and professional identities. A researcher can spend a lifetime mastering one narrow field and still miss the useful idea sitting two literatures away.

AI is unusually well suited to search the spaces between domains.

It can hold concepts from several disciplines inside one working context, translate a problem into the language of another field, and propose connections at a scale no single person could read their way into.

Many of those connections will be nonsense. Some will be obvious to specialists. A few will open new paths.

That is enough.

When the cost of generating and testing connections falls, progress begins to look less like a single curve and more like a network of interacting curves. Each breakthrough becomes an input into the entire discovery network.

This is why the acceleration can feel discontinuous. A better battery does not just improve a phone. It changes electric transport, grid storage, robotics, drones, remote infrastructure, and the economics of renewable power. Better energy storage then makes new machines practical. Those machines collect more data and perform more work. That work supports more research.

The breakthrough propagates.

The Bottleneck Moves To the Physical

The digital parts of science will accelerate first because software can run at machine speed.

Literature review, code, simulation, data analysis, and model comparison can be copied and parallelized. Atoms are less cooperative. A reaction still takes time. A telescope cannot observe a sky that is not visible. A clinical trial must respect biology, statistics, safety, and human consent. A new factory still needs concrete, machines, permits, supply chains, and power.

Scientific time dilation does not abolish the physical world.

It exposes the physical bottlenecks more clearly.

When an AI can generate ten thousand promising material designs, synthesis capacity becomes scarce. When models can propose hundreds of drug targets, wet labs and clinical validation become scarce. When engineering agents can design new hardware every hour, fabrication and testing become scarce.

This will pull capital toward the bridge between intelligence and matter: self-driving labs, robotic factories, advanced simulation, high-throughput testing, new energy systems, semiconductor fabrication, scientific instruments, and the infrastructure required to verify machine-generated ideas.

The next great technology companies may not look like software companies. They may look like automated foundries for turning machine intelligence into physical progress.

AI does not make scientific judgment less important.

It makes cheap scientific output abundant.

That creates a dangerous temptation to confuse volume with progress. A system can generate a thousand hypotheses that share the same hidden assumption. It can produce convincing analysis around contaminated data. It can optimize a measurable proxy while moving farther away from the question that matters. It can flood journals with papers faster than the world can verify them.

The first draft of a hypothesis gets cheaper.

Deciding what counts as knowledge gets more valuable.

The scientist moves up the stack. The job becomes less about manually executing every step and more about choosing the question, designing the search, building the evaluation, inspecting anomalies, demanding replication, and knowing when a result is important enough to change the map.

Taste matters. Skepticism matters. Domain knowledge matters. Ethics matters. The ability to distinguish a clever result from a useful truth matters.

The best scientists will command portfolios of machine exploration without surrendering the standards that make science trustworthy. They will use abundance to test more ideas, not to lower the bar. They will build systems that preserve provenance, record failures, invite adversarial review, and make replication part of the workflow.

Machines can widen the search.

Humans still choose what is worth finding.

Progress Becomes an Operating System

The economic implications are difficult to overstate.

In the industrial economy, nations competed on factories, energy, logistics, capital, and labor. In the information economy, they added software, networks, and data. In the intelligence economy, they will compete on the speed and quality of their discovery loops.

Who can turn a question into an experiment fastest?

Who can turn an experiment into reliable knowledge?

Who can turn knowledge into deployed technology?

Who can feed the result back into the system and begin again?

The organization that owns this loop gains more than a one-time advantage. It gains an engine for generating advantages.

This applies to countries, companies, universities, laboratories, and individuals. A small research team with strong models, clean data, automated instruments, and rigorous evaluation may explore more of a problem than a much larger institution built around grants, committees, and manual handoffs.

The individual scientist becomes a lab.

The lab becomes a network.

The network becomes a continuously learning institution.

We will need new funding and review systems, as well as safeguards around dual-use biology, dangerous chemistry, and autonomous experimentation. Faster discovery increases both capability and responsibility.

But caution should guide the loop, not stop it.

The answer to powerful science is better governance, stronger verification, clearer permissions, and more capable institutions. It is not choosing ignorance as a safety strategy.

We Can Build Faster Than Our Problems Compound

Techno-optimism is often mistaken for the belief that technology will automatically save us.

Even us geeks and nerds know better than that!

Technology expands the set of actions available to us. It gives people more ways to solve a problem and more power to create one. Wisdom, incentives, institutions, and human choice still determine what happens next.

But the expansion of possibility matters.

For most of history, humanity lived inside constraints it could barely understand. Disease looked like fate. Crop failure looked like fate. Darkness after sunset looked like fate. Distance, cold, infection, hunger, and infant mortality were treated as permanent features of existence.

Turns out they were not permanent.

They were problems waiting for knowledge, tools, and coordinated effort.

We still face enormous constraints: cancer, dementia, energy scarcity, fragile supply chains, new pathogens, and most importantly: billions of people whose talent is limited by poverty. AI does not guarantee that we solve them. It gives us a new way to aim far more intelligence, iteration, and experimentation at them.

That should make us ambitious.

Imagine materials discovered for properties we can describe but have never been able to build. Medicines designed around the biology of an individual rather than the average of a population. Energy systems optimized across generation, storage, transmission, and use. Robots that construct infrastructure in places too dangerous, distant, or expensive for people. Scientific instruments designed by systems that understand what today’s instruments cannot see.

None of this arrives by magic. We must build the labs, power the compute, run the trials, manufacture the hardware, govern the risks, and decide that broad human flourishing is the objective.

But for the first time, the intelligence available to help with that work can scale.

That is the source of my optimism.

Productive time dilation gives the individual access to more work than one lifetime could hold. OpenAI’s internal usage data already shows people at the 99th percentile generating more than seventy hours of Codex agent turns per day across parallel tasks.

Agent runtime is not identical to accepted human output. But it proves that the twenty-four-hour day is no longer the unit of productive capacity.

We will see individuals command the equivalent of 1,000 days of machine work inside a single five-hour session.

Scientific time dilation gives civilization access to more attempts than one century could hold.

For the right class of problem, an AI-led discovery system using quantum computation may cross in seconds a possibility space that would trap classical systems for 100,000 years.

Not 100,000 years of complete science.

Something more precise: 100,000 years of serial search collapsed into a single computational step.

The hypothesis must still survive experiment. The medicine must still survive a trial. The material must still be manufactured. Reality retains the final vote.

But questions that were computationally impossible can finally enter the laboratory.

The first form of time dilation changes who can build.

The second changes what can be known.

Together, they change what humanity can become.

We are not watching science reach its end. We are watching it gain a new operating system. AI gives the system more intelligence. Quantum gives it access to new regions of reality. Robotics gives it hands. Every discovery improves the machinery available to produce the next discovery.

The tools are improving the science.

The science is improving the tools.

The loop is tightening.

And humanity is learning how to build faster than its problems can compound.

We are winning again thanks to the singularity.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

The United States government does not owe OpenAI, Anthropic, or any other lab a durable business model. It certainly does not owe their investors a protected return. National security cannot become a magic phrase that freezes competition, blocks cheaper models, and turns a temporary lead into a permanent entitlement.

If Chinese and open models can do useful work at a fraction of the cost, American companies should use them. If frontier labs cannot justify premium prices for routine workloads, the market should compress those prices. That is competition doing its job.

Chamath is right about more of this than the defenders of frontier labs may want to admit.

I made much of this same argument in Frontier AI vs Chinese AI vs Open Source Self-Hosted AI. The model is not the whole product. The harness matters. Routing matters. Trust, permissions, context, memory, tools, evaluations, and workflow design matter. Most importantly, cost per token is a distraction. The useful economic measure is cost per completed unit of work.

A cheap model that finishes the task is valuable. A premium model used where a cheaper model would work is waste.

Chinese competition is not a boogeyman. It is a price-performance signal. Open models and efficient Chinese labs are attacking the assumption that intelligence must remain scarce, closed, and expensive. That pressure is good for builders, good for enterprises, and ultimately good for America.

But this is where I part company with Chamath’s argument.

A weak token-selling business does not prove that frontier research is a bad investment. It may prove that we are measuring the investment inside the wrong boundary.

The Frontier Creates the Commodity

The mistake is treating a frontier lab as if it were only a SaaS company.

If the lab sells access to a model, we look at API revenue, inference costs, margins, and customer retention. Those are sensible measures for the commercial product. They are not enough to measure the value of the research engine producing the next class of capability.

Frontier labs do the expensive, uncertain work of finding out what is possible. They train at scales that expose new behaviors. They build new architectures, new data systems, new evaluation methods, new inference techniques, and new ways to coordinate models with tools. They discover the capability before the rest of the market learns how to compress it, distill it, route around it, or reproduce it cheaply.

This does not mean every Chinese or open model is merely a copy. It does not mean the American labs deserve ownership of every idea that follows. Innovation is distributed, and competitive pressure often produces real breakthroughs in efficiency.

But commoditization is still downstream of invention.

The cheap layer spreads capability. The frontier layer creates new capability to spread. We need both.

In my earlier piece, I argued that the premium model is becoming the escalation path rather than the default. That remains true. A config change can go to a cheap model. A routine bug fix can go to a local model. A hard architecture problem can escalate to the frontier.

But an escalation path only works if it keeps escalating.

If nobody funds the uncertain work at the top, the routing table eventually stops improving. We get better at allocating yesterday’s intelligence while someone else determines tomorrow’s frontier.

America cannot accept that trade.

The Infinite Loop Is the Product

At the deployment layer, the harness is the agent.

At the frontier, the loop is the product.

A more capable model does not simply produce a better chatbot. Place it inside a research harness and it can propose experiments, write code, search design spaces, test hypotheses, analyze failures, and generate the next round of work. Connect it to simulators, scientific instruments, chip-design tools, software repositories, robotics facilities, and automated evaluators, and the model becomes part of a research system.

Then the system begins to feed itself.

A better model helps design better algorithms. Better algorithms make training and inference more efficient. Better agents improve software, data pipelines, and evaluation systems. Better chip layouts and compilers produce more useful compute. Better compute supports the next model. The next model becomes a stronger research operator, and the cycle begins again at a higher level.

This is not science fiction. The early mechanisms are already visible. Google DeepMind’s AlphaChip has helped design layouts used in multiple generations of Google’s TPU hardware. Chips helped build AI, and AI now helps build better chips.

AlphaEvolve makes the loop even clearer. It has been used to improve data-center scheduling, modify a circuit for a future TPU, and optimize a core Gemini training kernel. The system used frontier models to improve the infrastructure and algorithms used to train frontier models.

Do you see what’s already happening?

We can build autonomous research labs that run large portfolios of software and scientific experiments. We can build applied engineering labs where agents test thousands of designs across hardware, networking, compilers, energy, cooling, robotics, and manufacturing. We can build AI-enabled scientific facilities where models help form hypotheses, schedule instruments, read results, diagnose failed experiments, and decide what to test next.

Human judgment remains essential. Researchers choose the important problems, build the evaluation systems, inspect strange results, and decide what is safe and worth pursuing. But the volume and speed of attempted work can rise by orders of magnitude.

This is accelerating acceleration.

Each gain in intelligence improves the machinery used to search for the next gain. Each improvement in hardware, software, and research operations shortens the cycle again. The strategic asset is not one model checkpoint. It is the compounding system that keeps producing better ones.

Some Strategic Labs Will Look Uneconomic

The market is very good at pricing value a company can capture.

It is less reliable at pricing value that spills across an economy.

A frontier lab may spend billions creating a capability, only to watch much of the benefit flow to cloud providers, chip companies, power producers, application developers, hospitals, manufacturers, defense systems, universities, and millions of individual operators. The lab bears a concentrated cost while the country receives a distributed return.

In economics, that is a positive externality. In plain language, the lab creates value that someone else gets to keep.

This is why a strategically vital lab can look like a bad stand-alone business. Its API margins may be thin. Its training costs may be brutal. Its direct revenue may never match the value created downstream.

That does not make the work uneconomic.

It means the accounting boundary is wrong.

The most important return may appear as faster drug discovery, better logistics, stronger cyber defense, more efficient energy systems, improved industrial design, scientific breakthroughs, and whole new categories of companies. OpenAI’s recent FrontierScience work describes early evidence of frontier models shortening parts of scientific workflows from days or weeks to hours while also making clear how much open-ended research capability remains to be built.

Now compound those gains across every major research field and critical industry.

Even a small number of labs that never become attractive token businesses could still produce enormous national returns. Their economic impact would show up everywhere except their own income statements. Their national security value would be even harder to capture through an API invoice.

Non-economic does not mean worthless.

Sometimes it means the asset is infrastructure.

Protect the Capacity, Not the Cap Table

Chamath’s warning still matters because any national strategy can be captured by incumbents.

“Protect the frontier” can become a polite way of saying “protect our margins.” That would be a mistake.

The answer is not to declare two companies too important to fail. It is not to ban cheap competitors, force every workload through premium American APIs, or use China to excuse weak execution.

The answer is to fund frontier capacity without guaranteeing incumbent economics.

Government can use competitive, milestone-based contracts to buy research outcomes. It can expand shared national compute for universities, startups, and independent researchers. It can create autonomous research facilities around energy, materials, biology, manufacturing, and defense. It can fund hard evaluations, secure test environments, semiconductor research, power infrastructure, and the applied engineering needed to turn model capability into real systems.

Public money should also buy public value. That can mean research access, broad licensing, shared infrastructure, security testing, or the release of selected results after a sensible lead time. The goal is not to socialize every model. It is to prevent public investment from becoming a blank check for private scarcity.

Competition should remain brutal.

Labs should have to prove progress. New entrants should be able to challenge incumbents. Open models should keep pressing prices down. Enterprises should route each workload to the system that completes it best. Bad management should still fail.

Protect the capability, not the capitalization.

That is the policy line.

America Needs the Whole Stack

The United States does not have to choose between abundant intelligence and frontier leadership.

At the deployment layer, optimize for cost per completed task. Let cheap models handle the work they can handle. Build routers, evals, private deployments, and agent harnesses that make intelligence legible and manageable.

At the research layer, optimize for the rate of capability growth. Build the models, tools, labs, chips, power systems, and autonomous facilities that shorten the cycle from idea to experiment to result.

At the national layer, optimize for the long-term spillovers: scientific progress, industrial capacity, economic leverage, security, and the ability to set the direction of the most important technology of this era.

The White House’s AI Action Plan connects AI leadership to both economic competitiveness and national security. That connection is real even if some companies invoke it too conveniently. China should not be used as a sales tactic. It should be understood as a capable strategic competitor operating across models, chips, energy, manufacturing, and state capacity.

American leadership does not mean every useful model must be American. It means the United States retains the ability to run the most important experiments, attract the best researchers, build the full stack, and keep pushing the edge outward.

Chamath is right that the government does not owe frontier-lab investors a return. He is right that cheaper intelligence will create value across CPUs, GPUs, clouds, energy, applications, and skilled trades. He is right that markets should punish a business model that depends on permanent scarcity.

But “let the market sort it out” stops one layer too early.

Let the market sort out API prices. Let it sort out routine workload routing. Let it sort out which company has the best product, the best harness, and the best operating discipline.

Do not let a quarterly income statement decide whether America continues to fund the machinery that creates the next class of intelligence.

Cheap intelligence is the dividend.

The frontier is the engine.

America needs both.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

It is not.

A model can propose, calculate, summarize, criticize, and sometimes surprise us. It can read more material than any individual. It can generate hundreds of hypotheses before lunch. It can write code, design experiments, compare results, and explain its reasoning in language that sounds remarkably authoritative.

But it cannot become a credible research institution merely by producing more impressive answers.

A lab requires mandates. Tools. Memory. Budgets. Permissions. Evidence. Adversarial review. Independent verification. Claim boundaries. Human authority.

The model is one worker inside that system.

The AI research lab is the thing we are building.

The Chatbot Frame Is Too Small

The chatbot was a useful way to introduce AI

A person types a question. A model returns an answer. The interface is simple, the feedback is immediate, and the value is easy to understand.

That interaction trained us to judge AI at the level of the response.

Was the answer accurate? Was it useful? Did it sound intelligent? Did it save time? Could a better prompt produce a better result?

Those were the right questions for the first phase.

They are too small for the next one.

Research is not a single response. It is a chain of work performed over time. A serious research effort may include source acquisition, data cleaning, hypothesis generation, experiment design, implementation, testing, falsification, replication, interpretation, review, and publication.

Each step can fail in a different way.

The wrong source can enter the system. A dataset can contain leakage. An experiment can test something other than the stated hypothesis. A result can be statistically real but economically meaningless. A valid observation can be summarized badly. A narrow finding can become an expansive claim. An impressive output can inherit authority it never earned.

No individual answer can manage all of that.

Once AI begins performing research work rather than merely discussing research questions, the unit of design changes.

The answer is no longer the product.

The workstream is the product.

Intelligence Is Not Reliability

The distinction at the center of this project is simple:

Intelligence and epistemic reliability are different engineering problems.

Intelligence helps a system generate, transform, and interpret information. It enables the model to identify patterns, propose explanations, write code, reason through alternatives, and navigate unfamiliar domains.

Epistemic reliability is the system’s ability to determine what deserves belief.

Those capabilities overlap, but they are not the same.

A highly intelligent model can still rely on the wrong source. It can make an invalid inference. It can accept a convenient explanation without testing alternatives. It can produce a persuasive summary that hides uncertainty. It can follow a flawed method with extraordinary competence.

Better reasoning does not automatically create better provenance.

More context does not automatically create independent verification.

Greater fluency does not automatically create calibrated claims.

The most capable model in the world still needs to know which files are authoritative, which actions are permitted, which results have been independently reproduced, which objections remain open, and where its mandate ends.

These are not model-weight problems.

They are institutional-design problems.

The Institutional Formula

The research system we are building can be described with a compact formula:

Research system = models + mandates + tools + evidence + memory + critics + gates + authority

Every term matters.

Remove the models and little work gets done.

Remove the mandates and the work loses direction.

Remove the tools and the system cannot act on the world.

Remove the evidence and it has only assertions.

Remove the memory and it cannot compound.

Remove the critics and errors survive unchallenged.

Remove the gates and weak results become stronger claims.

Remove authority boundaries and the system begins deciding things it has no right to decide.

The model is important.

It is simply not the whole machine.

Models Are Workers

We should think of models as cognitive workers.

Some are fast and broad. Some are slow and careful. Some are good at code. Some are better at reviewing arguments, finding inconsistencies, searching literature, translating formal notation, or generating alternatives.

The choice of model matters in the same way that the choice of worker matters. Capability, specialization, cost, speed, and judgment all affect the result.

But no serious institution would define itself by the intelligence of one employee.

A brilliant researcher operating without source controls, budgets, peer review, or research ethics does not become a lab. A room full of brilliant researchers without coordination does not become one either.

Models need roles. They need work orders. They need access to the right materials and restrictions against the wrong actions. They need to produce outputs that can be inspected by people and machines that were not involved in creating them.

A model should not be asked to be the worker, manager, critic, auditor, and executive at the same time.

That may create the appearance of completeness.

It does not create independence.

Mandates Turn Questions Into Work

A question asks for an answer.

A mandate defines a body of work.

That distinction becomes essential when AI systems operate with tools, files, compute, memory, and time. “Investigate this problem” is not enough.

What exactly is the research question? Which sources may be used? Which inputs are frozen? What counts as relevant evidence? What actions are permitted? How much compute and money may be spent? Which metrics will be evaluated? What conditions require the run to stop?

A good mandate also defines what the system is not authorized to do.

It may permit reproduction but forbid discovery. It may permit candidate generation but forbid promotion. It may permit local testing but forbid external publication. It may allow an evaluator to produce a report while preventing it from rewriting the underlying evidence.

The prompt is becoming the work order.

The quality of the work order determines whether machine intelligence becomes leverage or noise.

Tools Create Consequences

A chatbot produces text.

A research agent can execute code, query a database, transform a dataset, inspect a repository, call a scientific library, run a simulation, or coordinate additional workers.

That is far more useful.

It is also far more consequential.

Once a model can act, permissions become part of the research architecture. The system needs to know which tools are available, which data may be read, which files may be changed, which operations require approval, and which boundaries are absolute.

Tool access should follow the mandate.

A worker reproducing a frozen result may need read access to source files and permission to create a temporary output. It does not need the ability to rewrite the frozen inputs. An evaluator may need to inspect a candidate package. It does not need permission to promote that candidate or alter the evidence ledger.

Capability should not imply authority.

The fact that a system can take an action does not mean it should be allowed to take it.

Evidence Must Be More Durable Than Prose

Models are exceptionally good at producing explanations.

That creates a temptation to treat the explanation as the evidence.

It is not.

A research summary may describe what happened, but it cannot substitute for the underlying files, hashes, commands, measurements, controls, and receipts. A polished conclusion can conceal a missing input just as easily as it can explain a valid result.

The evidence layer must be more durable than the prose layer.

For every consequential result, we should be able to ask:

What were the exact inputs? Where did they come from? Were they altered? Which tool versions were used? What operations ran? What failed? What was excluded? Which controls passed? How much did the run cost? Can another system reproduce the result from the receipt?

These questions are not bureaucracy.

They are the difference between a claim and an auditable claim.

This becomes more important as machine-generated work increases. A human researcher may produce a handful of significant artifacts during a project. A machine research system may produce thousands.

Without structured evidence, the volume becomes unmanageable. The institution starts trusting summaries because reconstructing the work is too expensive.

That is how output abundance becomes epistemic debt.

Memory Must Become a Ledger

Most AI memory is designed to improve continuity.

It remembers preferences, prior conversations, project context, and earlier decisions so the user does not have to repeat them.

Research needs something stronger.

A lab must remember not only what it believes, but why it believes it.

It must remember which sources were authoritative, which hypotheses failed, which controls were missing, which objections remained unresolved, which candidate was held, and which decision changed the direction of the work.

That memory cannot depend on a model retelling the past.

It needs a ledger.

The ledger should preserve artifacts, provenance, decisions, costs, failures, and authority. It should allow future workers to inherit verified state without inheriting unsupported conclusions.

This changes the economics of failed research.

A failed experiment no longer disappears into a folder or a researcher’s memory. Its design, result, and failure mode become reusable institutional knowledge. A later worker can avoid repeating the same mistake or test whether changed conditions alter the outcome.

The institution compounds.

Not because the model remembers more tokens, but because the system retains better evidence.

Critics Must Be Designed to Disagree

Adding another model does not create independent review.

Two agents can share the same context, the same framing, the same training biases, and the same unexamined assumptions. They may agree because the evidence is strong. They may also agree because they were constructed to see the problem in the same way.

Agreement is not independence.

Criticism must be designed into the process.

One worker should test source integrity. Another should attack reproducibility. Another should search for leakage and hidden dependencies. Another should challenge the relationship between the evidence and the claim. Another should deliberately construct alternative explanations.

These critics need explicit adversarial mandates. Their job is not to make the original work sound better. Their job is to find a reason it should fail.

In some cases, they should be isolated from one another’s conclusions until their reviews are complete. Different model families or providers may be useful when procedural independence matters. Their outputs should be frozen before a final evaluator compares them.

The objective is not performative disagreement.

It is error detection.

A credible system does not ask: “Do several agents like this result?”

It asks: “Did sufficiently independent attempts to break this result fail?”

Gates Convert Evidence Into Authority

Evidence and authority are different things.

A result can be real without authorizing publication. A candidate can be interesting without authorizing more spend. A research package can pass technical checks without justifying a claim of human verification.

This is why the system needs gates.

A gate asks a specific decision question and accepts only the evidence relevant to that question.

The Research Evidence Gate asks whether the evidence supports the stated research conclusion.

The Credibility Gate asks whether a specific external credibility claim is justified.

The Publication Gate asks whether a release is operationally and ethically ready to publish.

The Planning Gate asks whether the next bounded discovery campaign should be authorized.

These gates must remain separate.

A missing hosting configuration should not invalidate a local reproduction. A screen-reader review should govern an accessibility claim, not whether private research can continue. A technical result should not silently inherit permission to publish itself.

Coupled gates create two opposite failures.

They allow unrelated requirements to block legitimate work.

And they allow evidence from one domain to grant authority in another.

Good gate design prevents both.

Authority Must Remain Explicit

Autonomous systems create pressure to make continuation automatic.

If a candidate passes, run the next experiment. If the experiment succeeds, expand the search. If the search produces a strong result, prepare the publication. If the publication package is complete, release it.

This feels efficient.

It is also how local success becomes uncontrolled authority.

Every transition changes the risk.

Reproduction becomes discovery. Discovery becomes validation. Validation becomes publication. Publication becomes reputation. In commercial or financial settings, a research result might eventually become a real-world action.

Those transitions should not occur because a model inferred that continuation was probably intended.

Authority must be explicit.

The system should know who can approve additional spend, broaden a mandate, make an external claim, publish a result, or stop the program entirely. It should record that decision and make the resulting permissions visible to every downstream worker.

A gate without an authority model is only a checklist.

An authority model without a gate is only hierarchy.

A credible institution needs both.

Humans Do Not Leave the System

The point of an AI-native lab is not to remove humans from research.

It is to move human effort to the places where it has the greatest leverage.

Machines can search, calculate, transform, compare, reproduce, and criticize at extraordinary scale. They can run many more bounded attempts than a human team could afford.

Humans still choose what matters.

We define the problem. We decide which risks are acceptable. We judge whether a technically valid result is meaningful. We decide when procedural independence is enough and when genuine human expertise is required. We choose what the institution will claim in public.

This is not a sentimental boundary.

It is an architectural one.

Judgment, taste, responsibility, and legitimate authority do not become unnecessary because cognitive work becomes cheaper. They become more important because the volume of possible action expands.

The human role shifts from performing every unit of work to designing and governing the institution that performs it.

That is a higher-leverage role.

It is also a more demanding one.

Why a Swarm Is Not a Lab

The easiest version of agentic research is a swarm.

Give many agents a problem. Let them explore in parallel. Ask other agents to rank the answers. Aggregate the results.

This may produce useful work. It may even produce breakthroughs.

But scale alone does not create institutional reliability.

A thousand agents operating without frozen mandates, provenance, budgets, controls, critics, or gates are simply a thousand opportunities to create convincing error.

The differentiator will not be the number of models deployed.

It will be the quality of the conversion process.

How efficiently can the system turn machine effort into validated observations? How much does each experiment reduce uncertainty? How often do results reproduce? How quickly are false positives killed? How much operator time is required? Can another evaluator reconstruct what happened without trusting the original workers?

The next generation of research systems will compete on research yield, not artifact volume.

More intelligence creates more possibilities. Better institutions determine which possibilities survive.

The Lab We Are Building

CORTEX is our attempt to build this machinery.

It is an AI-native research operating system made of workers, managers, critics, gates, evidence contracts, budgets, permissions, and durable memory. It is designed to run bounded research campaigns without confusing machine activity with established knowledge.

The Frontier Problems Lab is the institution we are building around it.

The destination is a lab designed to attack problems that resist ordinary research workflows. Its advantage will not come from pretending that a model is an autonomous scientist. It will come from coordinating many forms of machine intelligence inside a system built to preserve evidence, invite attack, control claims, and retain human authority.

Our first calibration run was intentionally narrow.

CORTEX reproduced 56 frozen records and the corresponding table exactly. The source hashes matched. The deterministic evaluator agreed. The system then refused to claim that the underlying geometry had been proven.

That refusal was not a disappointing ending.

It was evidence that the surrounding institution had begun to work.

The model completed the assignment. The lab controlled the meaning.

Build the Institution

AI will become more intelligent.

The models will reason better, use tools more effectively, retain larger contexts, and coordinate more complex work. Tasks that currently require elaborate orchestration will become routine model capabilities.

We should welcome that progress.

But more intelligence will not eliminate the need for institutional architecture.

It will increase it.

The faster the workers become, the more important the work orders become. The more candidates the system can generate, the more important falsification becomes. The more persuasive the outputs become, the more important evidence and claim boundaries become. The more actions the system can take, the more important explicit authority becomes.

The future of AI research will not be built by choosing between human scientists and machine scientists.

It will be built by designing institutions where machines can perform enormous amounts of cognitive work without being allowed to manufacture certainty, erase provenance, or grant themselves authority.

That institution will have models.

But it will also have mandates, tools, evidence, memory, critics, gates, and accountable human judgment.

The model is not the research system.

The model is a worker.

We need to build the lab.

That’s what I’m working on!

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

The human brain runs on roughly the power of a dim light bulb.

Inside that twenty-watt envelope, it sees, remembers, predicts, learns, moves a body, reads a room, and maintains a model of itself. It does this continuously. No liquid cooling. No data center. No rack of accelerators drawing enough electricity to power a neighborhood.

That fact has become a kind of talisman in debates about artificial intelligence.

The brain is presented as evidence that biology possesses some vast computational advantage that silicon cannot approach.

Artificial systems may be impressive, the argument goes, but they remain crude imitations of a machine refined by hundreds of millions of years of evolution.

The gap is real.

And we are about to blitz beyond it.

The brain is not millions of times beyond our best AI hardware on every dimension. On raw low-precision computation per watt, the difference may already be less than an order of magnitude depending on how we define the operations. The larger advantage is in memory: how much adaptive state the brain keeps close to computation, how quickly it can use that state, and how little energy it spends moving information.

That is an engineering advantage, not magic. And engineering advantages can be erased.

My bet is that AI hardware will cross the brain’s compute-efficiency envelope within three years. That is a BIG bet.

Within five years, new architectures built around wafer-scale systems, stacked memory, and near-memory computation will close the much harder gap in memory access.

By then, silicon will surpass the individual human brain across most forms of economically useful cognition.

That will not make the human mind irrelevant. It will change where the mind ends.

We Are Counting the Wrong Things

Most brain-versus-model comparisons begin with two large numbers.

Direct cell-counting research places the human brain at about 86 billion neurons. Frontier language models contain hundreds of billions or even trillions of parameters. Put those figures beside each other and the model can appear larger than the brain.

But a parameter is not the artificial equivalent of a neuron.

A better analogy compares parameters to synapses.

A parameter is a stored value that affects how a signal moves through an artificial network. A synapse is a connection whose strength affects how activity moves through a biological network. Both hold persistent adaptive state. Both encode something learned from prior experience.

The neuron is closer to a computational node. It receives signals, accumulates them, transforms them, and produces new activity. In a language model, the closest comparison is not one permanent object but the machinery that produces activations and intermediate results as information passes through the network.

The mapping looks roughly like this:

- Synaptic strength maps to model weights

- Neural activity maps to activations and hidden state

- Neurons and dendrites map to the structures that accumulate and transform signals

- Axons and synaptic events map to communication across memory and compute

Once we compare the right categories, the brain regains its lead.

The number of synapses in a human brain is not known precisely, but common estimates range from around 100 trillion to several hundred trillion, with some estimates reaching a quadrillion.

Publicly documented language models have reached the trillion-parameter range. Kimi K2 for example, has one trillion total parameters, although its MoE design activates only 32 billion for each token.

That distinction matters. Total parameters measure stored state. Active parameters help determine the work performed during one pass.

A sparse model contains a large amount of knowledge without consulting all of it at once.

The brain does something similar. It does not activate every neuron or synapse every time you form a thought. Biological computation is sparse, conditional, and shaped by context. Most of the system remains quiet while small coalitions of activity do the immediate work.

If we compare roughly one trillion model parameters with 100 trillion to one quadrillion synapses, the brain is perhaps two or three orders of magnitude larger in persistent adaptive state.

That is a big gap. It is also a surprisingly bridgeable one.

Two orders of magnitude is not an unknowable biological moat. It’s a technology roadmap.

The comparison still has limits. A biological synapse changes over time, interacts with chemical systems, and can hold several forms of state. Neurons are more complex than the simple units in most artificial networks.

The brain learns continuously while acting through a body. Most language models separate expensive training from mostly static inference.

Parameter count is not intelligence.

Capacity is not capability. A larger system can still be worse.

But the comparison does one useful job: it turns a mystical gap into a measurable one.

The Brain Is a Bandwidth Machine

The easiest way to misunderstand modern chips is to focus only on arithmetic.

Silicon is extraordinarily good at multiplication. Current AI accelerators can perform quadrillions of low-precision operations per second. The harder problem is keeping those arithmetic units supplied with data.

Every model weight has to live somewhere. During inference, weights must be read so they can be combined with activations. When the model is larger than the memory close to the processor, those values must travel across packages, boards, and sometimes networks. Each trip costs time and energy.

Multiplication is cheap.

Moving the numbers is expensive.

This has been understood in chip design for years. Mark Horowitz’s widely cited analysis showed that retrieving data from off-chip DRAM could cost orders of magnitude more energy than performing a basic arithmetic operation.

The hierarchy remains: local movement is cheap, distant movement is expensive. Distance becomes energy.

The brain solves this problem through radical locality.

Its memory is distributed throughout the system. The state that shapes a computation lives at the synapses where signals arrive. Neurons accumulate activity from nearby connections. Communication is sparse, slow, noisy, and massively parallel. Instead of moving a giant weight matrix back and forth between separate banks of memory and compute, biology places the memory inside the network.

The brain does not retrieve its model before it thinks.

The model is the physical structure doing the thinking.

That difference explains why comparisons based only on FLOPS can be so misleading. If we assume the brain performs the equivalent of roughly one quadrillion operations per second while consuming twenty watts, it delivers about 50 trillion operations per second per watt. An NVIDIA B300, using its advertised peak of 15 quadrillion dense NVFP4 operations per second and a 1.4-kilowatt power envelope, lands around 10.7 trillion operations per second per watt.

Under that particular set of assumptions, the brain is only about five times more efficient.

Only is doing a lot of work in that sentence. A neural event is not an NVFP4 operation. Peak specifications are not sustained performance. The brain mixes analog and digital functions, and much of its energy maintains a living system rather than executing matrix multiplication.

Put simply: there’s no clean exchange rate between a thought and a FLOP.

Still, the calculation is useful because it shows that raw compute efficiency is not separated by six or nine orders of magnitude.

How do we get to the promised land?

With better transistors, lower precision, sparsity, improved utilization, and hardware designed around AI workloads, a fivefold gap can disappear quickly.

Memory is harder. Much harder. This is where the brain’s engineering is still beyond what humans are currently capable of.

Depending on how we estimate synaptic activity and stored state, the brain’s effective bandwidth per watt may exceed a conventional GPU by hundreds or thousands of times. The exact number is debatable because the units are artificial. The architectural fact is not: the brain spends very little energy moving each piece of information because the distance is short and the communication is sparse.

The brain’s moat is not calculation raw power or even hyper-efficient compute.

It is locality. And that moat is collapsing at accelerating rates.

Silicon Is Learning Locality

The direction of AI hardware is already clear.

High-bandwidth memory places larger and faster memory stacks close to the GPU. Advanced packaging creates wider connections between processors and memory. Chiplets shorten communication paths between specialized components. Sparse models activate only the parts of the network needed for the current input.

Wafer-scale computing takes the idea further. NVIDIA’s B300 pairs up to 288 gigabytes of HBM3e with roughly 8 terabytes per second of memory bandwidth. Blackwell can deliver 15 petaflops of dense NVFP4 compute on a single GPU.

Cerebras uses a very different design. A very unique one.

Instead of cutting a wafer into many small chips and reconnecting them across a board, it turns nearly the entire wafer into one processor. Its WSE-3 contains 44 gigabytes of on-chip SRAM and advertises 21 petabytes per second of memory bandwidth. That is more than 2,600 times the B300’s HBM bandwidth, although SRAM and HBM serve different roles and the systems should not be treated as interchangeable. The point is what becomes possible when data stays on the same piece of silicon.

Cerebras has traded memory capacity for radical bandwidth. Forty-four gigabytes is enormous for on-chip SRAM, but it remains tiny compared with the estimated adaptive state in a brain or the memory needed to hold the largest models. The next step is to combine wafer-scale compute with far more local memory.

Imagine a three-dimensional Cerebras.

Instead of spreading compute and memory across a flat board, stack layers of dense memory directly over or under layers of logic. Connect them with huge numbers of short vertical links. Keep frequently used weights close to the arithmetic. Move less data across the package. Move even less across the rack.

This is not yet a solved product design. There are several problems being worked on by brilliant people in labs all across the planet. Stacking active components creates problems in heat, yield, power delivery, and manufacturing. Memory also involves tradeoffs among density, speed, durability, and precision. A beautiful architecture can fail when it must be manufactured by the million.

But the path forward (and upward) is real. Researchers have already demonstrated monolithic three-dimensional systems with multiple vertically integrated circuit tiers. Other teams are developing near-memory compute, analog in-memory operations, and new forms of nonvolatile memory that can store values and participate in computation.

The hardware roadmap is converging on the brain’s central trick.

Put memory where the work happens.

The Fifteen-Year Bet

Predictions about intelligence become slippery because people combine several different claims.

Hardware efficiency is not model capability. Model capability is not autonomy. Autonomy is not consciousness. A machine can outperform a person at economically useful work without thinking or feeling like a person.

So my forecast has three parts.

First, I expect a shipping AI accelerator to surpass the brain on at least one defensible measure of low-precision computation per watt by 2028/2029. Depending on how the brain-equivalent operation is defined, someone may plausibly claim this sooner. But within three years, the raw compute-efficiency case should become difficult to dispute.

Second, I expect stacked memory, wafer-scale systems, and near-memory computation to erase most of biology’s advantage in effective memory access within five years. This is the harder prediction. It depends less on faster arithmetic than on packaging, materials, heat removal, and the physical distance traveled by each bit. But AI is creating self-reinforcing loops across science and technology, and that make me confident we are already deep inside the singularity and can expect accelerating acceleration going forward.

Third, I expect AI systems to surpass humans across most economically valuable cognitive tasks within the same period.

That third prediction does not follow automatically from the first two. Intelligence is not a benchmark for hardware utilization. The brain has recurrence, embodiment, online learning, emotional signals, and evolved drives that current models do not reproduce.

But hardware parity removes one of the strongest reasons to believe biological cognition occupies an unreachable level.

AI also does not need to fit inside one skull-sized chip. A system can use many processors, external memory, retrieval, tools, and specialized models. It can copy itself, run in parallel, and spend far more than twenty watts when the result is valuable.

It can trade energy for speed in a way evolution could not.

Silicon may surpass the brain as a system before any single chip resembles one.

The Mind Is Moving Outside the Skull

The most important consequence is not that machines become more like us.

It is that we become less limited to ourselves.

Human beings have always extended cognition into the environment. That was one of our special evolutionary forces. Language let one mind shape another. Writing externalized memory. Mathematics externalized formal reasoning. Institutions allowed groups to hold knowledge and coordinate action beyond the capacity of any member. Software externalized repeatable procedure.

AI externalizes parts of cognition itself.

A book can preserve an idea, but it does not adapt the idea to your situation. A database can store facts, but it does not decide which facts matter. A spreadsheet can execute rules, but it does not usually rewrite the rules after examining the outcome.

Models can transform information. They can compare, draft, critique, explain, search, simulate, and act. Connected to tools and persistent memory, they can carry work across hours or days. They can hold several competing interpretations while the human chooses among them.

The model alone is not the external mind.

The system around the model is.

The model generates possibilities. Memory preserves context. Tools let it affect the world. Permissions define what it may touch. Evaluations detect failure. Workflows give it continuity. Human judgment supplies goals and decides what deserves to survive.

A wild thing to think about: that stack above changes the unit of thought itself.

The old unit was the individual mind → one brain, one working memory, one stream of conscious attention.

The new unit is a person surrounded by models, memories, tools, and agents. A research agent can explore the literature while a coding agent tests an implementation and an operating agent monitors the system. The person no longer performs each cognitive step. They shape the environment in which cognition happens.

This is more than productivity software.

It’s a new cognitive architecture.

The boundary of the mind has always been porous. AI makes that porosity operational. Parts of what we remember, notice, compare, and produce will live outside the brain but remain available as extensions of our agency.

The skull stops being the practical boundary of the mind.

The Bottleneck Moves Up the Stack

When compute is scarce, intelligence looks like producing an answer.

When compute becomes abundant, intelligence looks like choosing what should be answered.

This is the deeper shift. AI will make competent cognitive output cheap. It will become easy to create ten analyses, one hundred designs, or one thousand possible strategies. More systems will be able to code, write, plan, negotiate, and research at a level that once required trained specialists.

Abundance does not remove scarcity. It moves it.

Answers become abundant. Good questions remain scarce.

Output becomes abundant. Taste remains scarce.

Analysis becomes abundant. Commitment remains scarce.

Memory becomes abundant. Attention remains scarce.

Intelligence becomes abundant. Agency remains scarce.

The advantage will belong to people who can build and direct a cognitive system without losing themselves inside it. They will know how to divide work among models, create feedback loops, preserve useful context, test uncertain claims, and apply judgment at the points where mistakes matter.

They will not compete with AI by trying to think every thought manually.

They will decide which thoughts are worth having.

Build a Mind You Still Control

There is an optimistic version of this future in which people gain extraordinary leverage.

A capable individual can draw on more knowledge, explore more options, and build more ambitious things than a large organization could manage before.

There is also a dangerous version we need to talk about.

The systems that remember for us and reason with us will influence what we notice. The models that summarize the world will shape which parts of the world remain visible. If we outsource not only execution but also goals, standards, and judgment, greater intelligence can produce weaker agency.

Cognitive leverage without cognitive sovereignty is dependency.

The practical response is not to reject artificial intelligence. It is to become deliberate about the mind you are assembling around yourself. Use it for leverage but don’t let it do your thinking for you.

Own important context. Know which systems can read it. Keep evidence attached to consequential claims. Use multiple attempts when uncertainty is high. Preserve the ability to inspect the work. Automate execution aggressively, but be careful about automating your goals.

The most valuable human skills will sit above raw cognition: choosing objectives, forming values, reading consequences, building trust, exercising taste, and accepting responsibility for a decision.

Those are not consolation prizes left over after machines take the real work.

They are the control layer.

The human brain is still the most remarkable general-purpose thinking system we know. It holds an enormous amount of adaptive state, learns from sparse experience, and runs continuously on about twenty watts.

But its lead is not infinite.

Frontier models are already within a few orders of magnitude of the brain’s estimated synaptic scale. AI chips are within striking distance under some measures of raw compute per watt. The remaining hardware moat is memory locality, and nearly every important trend in advanced chip design is aimed at moving less data across shorter distances.

Silicon will catch the brain.

When it does, that will not mark the end of the human mind. It will mark the end of the skull as the mind’s practical boundary.

Our advantage will not be that we can produce more thoughts per second. It will be that we can decide which thoughts deserve attention, which systems deserve trust, and what all that intelligence is for.

The future of the mind is larger than the brain.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

We did not claim that a theorem had been reproduced. We did not claim that the source census was complete. We did not claim a new mathematical result.

We had proven something much narrower.

CORTEX could take a frozen research mandate, process the evidence, reproduce a defined result, preserve the chain of custody, and stop at the boundary of what the evidence justified.

That might not sound dramatic for a few reasons.

It did not solve an open problem.

It did not discover a new construction.

It did not produce the kind of result that generates headlines about artificial intelligence transforming science.

But it demonstrated a capability that most AI systems still lack.

The ability to know when they have not proven something. And in a world of hallucinations and half-truths, if you want to use the power of AI you need rigid boundaries and verifiable truths to conduct new science.

What We Are Building

CORTEX is an experimental AI-native research system.

Computational Operations for Research, Thesis, Evidence, and eXperimentation.

It is not a model. It’s not a chatbot with a longer prompt. It’s not a single autonomous agent that receives a question and returns something resembling a research paper.

CORTEX is the machinery around the models.

It converts a research mandate into bounded work. It assigns specialized workers to gather evidence, reproduce results, generate candidates, run tests, and attack conclusions. Managers allocate budgets and decide what deserves further attention. Critics search for leakage, weak controls, hidden assumptions, and alternative explanations. Gates determine whether the evidence supports another experiment, a larger claim, or no further action at all.

Every consequential action is supposed to leave a receipt. That’s one of the big breakthroughs in this design.

The system records the source material, tool versions, input hashes, transformations, costs, outputs, objections, decisions, and unresolved questions. It maintains a ledger so that later workers do not have to reconstruct what happened from a polished summary or trust the memory of another model.

Humans remain outside and above this machinery. We choose the mandate. We set the budget. We define the authority boundaries. We decide which claims may leave the lab.

The metaphor I keep returning to is a smart factory.

The models are workers on the factory floor. Some are fast generalists. Some are narrow specialists. Some build. Some inspect. Some try to break what the others produced (they are my favorite).

But workers alone do not make a factory.

A factory also needs work orders, raw-material controls, managers, quality systems, safety rules, accounting, maintenance, and executives who decide what may ship. Without that surrounding structure, adding more workers may increase output without increasing quality.

Research has the same problem.

A larger swarm can produce more hypotheses, more experiments, more critiques, and more prose. It cannot guarantee that any of those outputs deserve belief. Machine effort is becoming abundant. Independently validated knowledge remains scarce.

CORTEX is our attempt to build the conversion layer between the two.

The Frontier Problems Lab, a venture formed by McDonagh Family Office, is the institution we are building around that system. Its destination is a research lab designed to attack problems that resist ordinary workflows. The aim is not to stage impressive conversations with artificial intelligence. It is to learn whether a closed-loop machine institution can turn large amounts of bounded computational and cognitive effort into small amounts of defensible knowledge.

That requires more than discovery.

It requires reproduction, falsification, provenance, memory, independent review, explicit non-claims, and the ability to stop. It requires a system that can distinguish an interesting candidate from a validated result and a validated result from something ready to publish.

After building v1 of the system we began this effort with a calibration run.

Before asking CORTEX to discover something new, we asked it to reproduce something narrow and already defined. Before testing its creativity, we tested its discipline.

Before giving it freedom, we tested whether it could operate inside a contract.

The first question was not whether CORTEX could solve a frontier problem.

It was whether CORTEX could reliably know what it had done.

That is how we arrived at the 56 records.

The Most Dangerous Word in AI Research

The most dangerous word in AI research may be “success.”

We use it to describe too many different things.

A model returned an answer. Success.

An agent completed a workflow. Success.

A program ran without crashing. Success.

An experiment produced the expected number. Success.

A result matched a published table. Success.

A candidate survived a backtest. Success.

These events may all matter. But they do not mean the same thing.

A completed task is not necessarily a correct task. A reproducible result is not necessarily a true result. A true result is not necessarily a novel result. A novel result is not necessarily an important result.

Each step requires different evidence.

When those distinctions disappear, activity becomes confused with progress. More tokens, more agents, more experiments, and more documents create the appearance of a research program without establishing that anything has actually been learned.

This is especially dangerous with modern AI because the output is so persuasive. The model does not merely produce an answer. It produces an answer in the language of expertise. It explains, qualifies, cites, summarizes, and often sounds more certain than the available evidence warrants.

Fluency compresses uncertainty.

That makes the surrounding system more important, not less.

The model can generate possibilities. The research institution must determine what those possibilities mean.

What We Actually Asked CORTEX to Do

The first Frontier Problems Lab run was intentionally modest.

We did not ask CORTEX to solve a famous open problem. We did not ask it to invent a proof or search for a new geometric construction. We did not unleash a swarm of agents on an ambiguous mandate and hope that something interesting emerged.

The job was simple:

Take the frozen source artifacts. Verify their hashes. Parse the records. Reproduce the expected counts. Compare the result with the frozen table. Record exactly what happened.

This was not a test of advanced mathematical creativity. It was a test of whether the research machinery could follow an evidence contract.

That distinction was deliberate.

Before building a system that searches for new knowledge, we wanted to know whether it could reliably handle old knowledge. Before asking it to generate hypotheses, we wanted to know whether it could preserve inputs, execute a bounded mandate, produce receipts, and respect a claim boundary.

Discovery is built on reproduction.

If the system cannot reliably tell us what entered the factory, what operations were performed, what came out, and which claims follow from the result, then adding more intelligence merely increases the speed at which uncertainty is manufactured.

The Run Worked

The run passed.

The expected 56 records were present. The machine-readable source artifacts matched their frozen cryptographic hashes. The record distribution matched the expected contract. The edge totals matched. All 22 rows in the comparison table were reproduced.

The evaluator reached the expected conclusion: the baseline had been reproduced for artifact integrity.

That sentence matters because of what it does not say.

It does not say the underlying mathematical objects were correctly represented in every respect. It does not say every graph in the census has the required geometric properties. It does not say the census is complete. It does not validate the theorem, proof, or broader claims associated with the source.

It says the frozen artifacts were processed consistently and reproduced according to a defined contract.

Nothing more.

This is where many AI research demonstrations would begin expanding the story. The reproduction would become “validation.” The validation would become “verification.” The verification would become evidence that the AI understood the mathematics. By the time the result reached a headline, a successful data-processing exercise might be described as an autonomous mathematical achievement.

We did the opposite.

We narrowed the claim until it fit the evidence.

Reproducibility Is Not Truth

Reproducibility is essential to science, but reproducibility and truth are not synonyms.

A system can perfectly reproduce an error.

It can reproduce a flawed dataset, a mistaken assumption, an incomplete census, a transcription problem, or a test that measures the wrong thing. It can execute an invalid method with flawless consistency.

Reproduction answers one question:

Can the result be generated again under the stated conditions?

Truth demands a lot more.

Were the inputs valid? Did the method test what it claimed to test? Were relevant alternatives excluded? Did hidden assumptions shape the result? Does the conclusion survive independent attack? Does the evidence support the scope of the claim?

Those are separate questions.

Our first run established that CORTEX could reproduce a frozen artifact-level result. It did not establish the broader mathematical truth surrounding that artifact.

That was not a weakness in the experiment. It was the point of the experiment.

A credible research system must preserve the distance between what happened and what can be claimed about what happened.

The System Stopped Safely

The most important output of the run was not the reproduced table.

It was the hold.

CORTEX reached the end of its authorized mandate and did not convert a narrow reproduction into broader research authority. It did not begin searching for new constructions. It did not promote the result into a discovery claim. It did not treat the absence of an error as evidence of mathematical truth.

The candidate remained held.

This is easy to overlook because we are accustomed to measuring systems by what they produce. More answers. More candidates. More code. More experiments. More speed.

But in research, restraint is a productive capability.

A system that can generate a thousand hypotheses but cannot stop itself from overstating weak evidence is not an advanced research system. It is an industrial-scale speculation machine.

A system that can recognize the limit of its evidence is more valuable.

Stopping is not the absence of an output.

Stopping is an output.

It says the available evidence supports this claim and not the next one. It says the next action requires a different mandate, stronger evidence, or additional authority.

It preserves the value of what was learned without pretending that more was learned.

The Model Was Not the Researcher

The run also reinforced a broader lesson about artificial intelligence.

The model is not the research system.

The system included a frozen mandate, source artifacts, cryptographic hashes, a mathematics adapter, deterministic checks, expected outputs, non-claims, resource limits, authority boundaries, a gate decision, and a durable evidence record.

The model was one component inside that architecture.

This is a different way of thinking about AI.

The chatbot frame trains us to focus on the exchange between a person and a model. The person asks a question. The model produces an answer. We judge the answer by reading it.

That frame becomes inadequate as AI moves into consequential work.

Research is not one answer. It is a chain of actions, transformations, tests, judgments, and claims. Each step creates opportunities for error. Each transition needs a contract. Each claim needs evidence. Each expansion of authority needs a gate.

The question is no longer merely whether the model is intelligent enough.

The question is whether the institution around the model is disciplined enough.

Receipts Before Reputation

Human institutions use reputation as a shortcut.

We trust a result partly because of who produced it, where it appeared, who reviewed it, and whether the surrounding institution has earned credibility over time.

An AI-native research institution begins without that accumulated trust.

It must earn credibility another way.

Receipts before reputation.

What were the exact inputs? What were their hashes? Which tools and versions were used? What transformations occurred? What budget was consumed? Which controls ran? What failed? What remained unresolved? What decision was reached? What authority was explicitly withheld?

These records are not administrative debris. They are part of the research product.

The polished paper may eventually explain the result. The ledger explains how the result came to exist.

This matters because AI makes cognitive labor abundant. A machine can produce more hypotheses, analyses, critiques, simulations, and manuscripts than a human team could reasonably inspect.

When production becomes cheap, selection becomes expensive.

When answers become abundant, provenance becomes scarce.

When persuasive language becomes automatic, disciplined claims become a competitive advantage.

The bottleneck moves from generating work to establishing which work deserves belief.

More Agents Do Not Solve This

The popular response to the limits of one model is to add more models.

One agent researches. Another critiques. Another manages. Another votes. Perhaps a larger swarm will converge on the truth.

Sometimes that helps. Different models can find different errors. Specialized workers can handle different tasks. Parallel exploration can search a larger space.

But a swarm is not automatically an institution.

Ten agents can repeat the same assumption ten times. They can share the same contaminated context. They can reward one another’s fluency. They can converge because they were prompted similarly, trained similarly, or shown the same intermediate conclusions.

Agreement is not independence.

A useful multi-agent research system must engineer the conditions under which disagreement can matter. Reviews should be isolated when appropriate. Critics should receive explicit falsification mandates. Inputs should be frozen. Outputs should be committed before comparison. The final evaluator should not silently rewrite the work it is supposed to judge.

The system needs workers.

It also needs managers, critics, auditors, executives, and a ledger.

And those roles must differ in authority, not merely in prompt wording.

Evidence Needs Gates

Building this has been exciting and humbling. A major early learning: boundaries are key!

We initially treated several different questions as if they belonged to one gate.

Was the research evidence sound?

Was an external credibility claim justified?

Was a website operationally safe to publish?

Should the next research campaign be authorized?

Those questions are related, but they are not the same.

Coupling all of these questions creates institutional confusion. It allows an unresolved publication task to halt research, or a narrow research result to inherit public-release authority it never earned.

The solution is separate gates.

A Research Evidence Gate asks whether the evidence supports the stated research conclusion.

A Credibility Gate asks which external credibility claims are justified.

A Publication Gate asks whether a specific release is safe and responsible to publish.

A Plaanning Gate asks whether the next bounded discovery campaign should be authorized.

Different evidence. Different decisions. Different authority.

This separation makes the system both safer and faster. Safety comes from preventing authority from leaking between domains. Speed comes from allowing local research to continue without waiting for unrelated publication work.

Good governance should not merely stop bad actions.

It should make legitimate actions easier.

The Economics of Machine Research

The deeper reason this architecture matters is that AI is changing the economics of research.

Machine effort is becoming cheap.

A model can read thousands of pages, generate candidate mechanisms, write test harnesses, search parameter spaces, run adversarial critiques, and produce structured evidence packages. Multiple workers can operate in parallel. Failed approaches can be recorded and reused. The institution can learn which kinds of experiments produce information and which merely consume budget.

This creates enormous leverage.

It also creates a new failure mode: cheap work can overwhelm expensive judgment.

The future research bottleneck will not be the number of ideas we can generate. It will be the number of claims we can validate.

The winning institution will not be the one with the largest swarm. It will be the one that converts machine abundance into scarce, independently tested knowledge with the least waste, the clearest lineage, and the strongest claim discipline.

That is the factory we are trying to build.

CORTEX is not meant to be an oracle. It is meant to become part of an operating system for research: workers producing evidence, managers allocating effort, critics attacking results, gates controlling authority, and a ledger preserving institutional memory.

The objective is not to make the machine sound more confident.

The objective is to make the institution more trustworthy.

What Our First Run Really Proved

So what did the first run prove?

It proved that CORTEX could accept a frozen artifact-level reproduction mandate.

It proved that the system could verify source hashes, parse the expected records, reproduce the specified table, and create a deterministic result.

It proved that the system could state the result narrowly.

It proved that the surrounding controls could preserve explicit non-claims.

It proved that the run could end without unauthorized continuation.

That is not mathematical discovery.

It is infrastructure for mathematical discovery.

There is a temptation to skip this layer because it is less exciting than asking a powerful model to attack the frontier. But foundations become more important as the machinery above them becomes more capable.

A weak model inside a disciplined system may produce limited results.

A powerful model inside an undisciplined system can produce convincing fiction at industrial scale.

We would rather begin with the discipline.

Knowing What We Know

AI research will produce genuine breakthroughs.

Models will find patterns humans missed. Agent systems will search spaces too large for conventional teams. Machine-generated conjectures, experiments, proofs, counterexamples, and designs will become normal parts of serious research.

But the volume and persuasiveness of the output will create pressure to move faster than the evidence.

That is why our first successful run mattered.

Not because it solved something.

Because it established a boundary.

The records matched. The hashes matched. The evaluator agreed. The machine completed its mandate.

And the institution still said: this does not prove the geometry.

That sentence contains the beginning of a credible AI research lab.

1. Intelligence generated the work.

2. Architecture constrained the claim.

3. The ledger preserved the evidence.

4. The gate withheld authority.

5. And human judgment remained in command.

Our first successful AI research run proved almost nothing.

It showed us that we might be building a system capable of knowing exactly what that means.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

Past a certain capability threshold, the marginal quality of the model matters less than the quality of the context you feed it.

This sounds like heresy to the leaderboard crowd.

It is not. It is production reality.

For the last two years, everyone has been hypnotized by model deltas. This model is 4% better on a benchmark. That model has stronger reasoning. This one codes better. That one has a larger context window. This one is cheaper. That one is more agentic.

This one has better tool use. That one wins on vibes.

Fine.

Model quality matters. No doubt about it.

But once a model is sufficiently capable, the limiting factor shifts.

The bottleneck is no longer raw intelligence. Our middle-of-the-road laptops in 2027 will have open source AI capabilities that exceed today’s GPT 5.5 / Opus 4.8 threshold.

The bottleneck is whether the system knows what the hell is going on.

A brilliant, bleeding-edge model with bad context is a genius dropped into a dark room and asked to perform surgery with rumors.

A somewhat strong model with excellent context is a trained operator with the right file, the right tools, the right patient history, the right constraints, the right objective, and the right feedback loop.

Bet on the second system.

Every time.

The Threshold Changes the Game

Below the capability threshold, model quality dominates.

If the model cannot reason, cannot follow instructions, cannot use tools, cannot write coherent code, cannot hold structure, cannot understand ambiguity, cannot recover from errors, then context will not save it. You can hand a weak model perfect documentation and still get garbage.

There is a floor.

Intelligence must clear it.

But once the model clears that floor, the curve changes. The next upgrade still helps, but not in the same explosive way. The fifth leap in model quality does not create the same returns as the first. The gains begin to compress.

This is diminishing marginal return.

A model going from incompetent to useful is a revolution.

A model going from very good to slightly better is an optimization.

But context quality behaves differently.

Give the system better customer history, better examples, better domain rules, better retrieval, better tool outputs, better workflow state, better constraints, better preferences, better definitions of success, and the output often improves immediately.

Not because the model got smarter.

Because the model got situated.

It finally knows the game it is playing.

This is why context has roughly linear returns across a huge range of practical work. In some systems, it may even look superlinear, because good context unlocks latent capability that was already inside the model.

The model was not missing intelligence. It was missing the map.

Model V.S. Context

A lot of “model gains” people report are not model gains.

They are context gains wearing a “model costume”.

A team switches models and performance jumps. Everyone praises the new model. But what actually changed?

They rewrote the prompt.

They cleaned the input.

They added examples.

They improved retrieval.

They gave the model better schemas.

They added chain-of-workflow state.

They included user preferences.

They constrained the output.

They added a review step.

They improved tool descriptions.

They removed noisy documents.

They gave the system a clearer objective.

… then they say, “The new model is amazing.”

Maybe. Maybe not. I think we’re seeing several forces at the same time.. each of them accelerating us faster into the singularity.

Let’s talk about this, and the impact it has on the path forward for AI.

Read more

Not because it settles which model is “best.” It doesn’t… although some pretty obvious trends are emerging.

The point is that “best” is now too small a word.

Best for what task?

Inside what harness?

At what price per completed unit of work? On whose codebase?

The AI industry spent the last few years arguing about model leaderboards.

Databricks is pointing at something more useful: task-level economics inside real work. Their internal benchmark tested coding agents on actual engineering tasks drawn from Databricks’ own multi-million-line codebase, across the cross-language messiness that makes enterprise software real.

This is not a toy benchmark asking a model to solve clean public problems. It is closer to the thing companies actually care about:

Can this agent change our code, pass our tests, follow our conventions, respect our constraints, and do it cheaply enough that we stop rationing its use?

Because the future of software work will not be decided by model taste alone.

It will be decided by cost per shipped task.

The Unit of Work Has Changed

I’ve run engineering teams and groups of builders for the last 15-years. In many ways, it used to be more simple.

The old software productivity equation was: how many engineers do we have, how good are they, and how well do we coordinate them?

The new equation is stranger: how many agent attempts can we afford to run, how well can we route them, and how quickly can humans judge and integrate the results?

Databricks built its benchmark from merged pull requests. That detail matters. A pull request is not just a diff. It is a compressed unit of organizational knowledge: intent, code, tests, review, build context, and the hidden social fact that a team decided this work was good enough to ship.

That makes it a much better raw material for evaluating coding agents than synthetic puzzles. Public benchmarks are useful, but they age. They leak. They get trained on. They also tend to flatten the task into something cleaner than daily engineering work really is.

Databricks did something more grounded. They pulled from recent internal history, filtered out bot and generated work, looked for self-contained changes with tests, rewrote task descriptions so the model saw the goal rather than the solution, then judged the result by whether held-out tests passed.

They also avoided the evaluation trap that is quietly poisoning a lot of AI discourse: they did not use an LLM judge as the primary arbiter of correctness.

That matters because the goal is not to impress a model with a plausible explanation. The goal is to ship working software.

This is the first lesson from the benchmark: stop benchmarking vibes. Benchmark outcomes.

The Model is Not The Product

The most interesting result from Databricks is not that one model won.

It is that no single axis explains the outcome.

Their findings show a Pareto frontier that includes OpenAI, Anthropic, and open models. That alone is significant. Frontier coding performance is no longer a private two-company story. It is becoming a mixed ecology.

The benchmark also showed rough capability tiers. The top models are good across hard tasks, but they are expensive. Medium and smaller models can handle a lot of common work at much lower cost. That should change how engineering organizations deploy AI immediately.

Most companies still behave as if every task deserves the most expensive model they can access. That feels safe because the premium model is usually strong. It is also economically incoherent.

Not every task is a research problem.

Some work is changing a config. Some work is updating a test. Some work is tracing a simple bug. Some work is designing a migration path across three services without breaking the build.

This is where agentic software starts to look less like chat and more like operations. The important capability is not “which model do you like?” It is routing. It is knowing when to send the task to a cheap model, when to escalate to a frontier model, when to run three agents in parallel, when to ask a human for context, and when to stop because the test signal is not good enough.

The model is one part of the system.

The harness is the agent.

By harness, I mean the execution environment around the model: context selection, file access, tools, memory, permissions, shell commands, diff handling, tests, retries, branch management, and review loops.

Databricks found that the same model, with similar thinking effort, could have materially different cost and efficiency depending on the harness. In some cases, cost per task differed by more than 2x while quality stayed roughly the same.

That is not a footnote.

That is the product.

For the last year, everyone has been arguing about model intelligence. The Databricks benchmark says something more operational: intelligence without context discipline is expensive.

This is what every enterprise needs to internalize: the agent is the model plus the operating system around it.

Cost per Token is Broken

The benchmark also attacks one of the most common mistakes in AI budgeting: treating price per token as a proxy for price per task.

That frame is too tight and misses the full picture.

A cheaper model can become more expensive if it reads more, loops longer, retries poorly, drags in too much context, or fails often enough that humans have to repair the output. A more expensive model can be cheaper if it reaches the right answer quickly.

Databricks gave a clean example. Sonnet 5 was cheaper per token than Opus 4.8, but on their tasks it cost more per completed attempt while scoring lower. The reason was behavior. It consumed more tokens to get to a worse result.

This is the metric shift: do not ask what the model costs.

Ask what the completed task costs.

That is the number founders, CTOs, and operators should care about. If an agent fixes a bug, writes a migration, updates documentation, and passes the right tests, token line items are just ingredients. The unit that matters is the finished piece of work.

This is why every serious company will eventually build its own benchmark. Public benchmarks could not answer the question Databricks actually had, because Databricks does not run a public benchmark company. It runs Databricks. Its codebase, build graph, engineering patterns, and task distribution are the reality that matters.

That is true for every company.

If you have a backlog of merged PRs, you have the raw material for an internal agent benchmark. You can measure which tools solve your work. You can price them at the task level. You can see which models are overkill, which are underrated, and which harnesses quietly burn money by spraying context everywhere.

The companies that do this will compound.

The companies that do not will keep buying AI by brand.

GLM Is Bad News for American Models

The headline result is that GLM-5.2 landed in Databricks’ top capability tier. It was statistically tied with Opus 4.8 on quality in their benchmark, while costing less per task.

That is a big deal. Massive actually.

Not because one benchmark proves GLM is universally better. It does not. The honest read is narrower and stronger: on a serious internal coding benchmark from one of the most technically sophisticated software companies in the world, a lower-cost Chinese open model performed like a daily-driver coding model.

That is exactly the kind of evidence that changes buyer behavior.

In my piece, GLM-5.2 Proves AI Comes for All Moats I argued that GLM matters because it attacks the scarcity story underneath Western AI valuations. Premium labs need the market to believe frontier intelligence will remain scarce, expensive, proprietary, and defensible. They need “best model” to become “best business.”

GLM does not kill that argument, but it does compresses it.

If an open or open-ish Chinese model gets close enough on real coding work, the buyer’s question changes. It is no longer “who has the most prestigious model?” It becomes “why am I paying the frontier tax for this workload?”

Sometimes the answer will be good. Enterprises will still pay for trust, support, indemnity, governance, data controls, integrations, uptime, multimodal polish, and ecosystem maturity.

Premium models will still matter for the hardest work. But not every workload needs the sacred object.

And “not every workload” is where the economic damage begins.

If GLM can handle a meaningful share of coding tasks at lower cost, it does not need to beat OpenAI or Anthropic at everything. It just needs to be good enough on enough work to change the routing table.

That is how markets reprice.

Not all at once. Not with one dramatic replacement. Through thousands of small substitutions.

A config change goes to GLM. A test update goes to GLM. A medium bug fix goes to GLM. A migration plan starts with GLM, escalates to a premium model for design review, then returns to GLM for implementation attempts.

Suddenly the premium lab is not the default.

It is the escalation path.

That is a very different business.

Chinese Efficiency vs American Muscle

The uncomfortable American lesson is not just that Chinese models are catching up.

It is that they are catching up differently.

Western AI culture has been dominated by scale: more compute, bigger clusters, deeper capital pools, premium APIs, and a belief that the frontier can be held by whoever spends the most.

China has been forced into a different game. Sanctions, chip constraints, competitive pressure, and lower pricing power create a harsher environment. That environment rewards efficiency: architectural tricks, distillation, routing, context management, serving optimization, and ruthless price-performance thinking.

There are legitimate questions about Chinese labs that I’ve called out many times before. We know Chinese models benefit from Western outputs. Distillation and synthetic data are everywhere. There will be fights over originality, fairness, export controls, national security, and whether closed labs are funding the research that commoditizes their own products.

Those questions matter.

But they do not erase the market effect.

Customers buy outcomes. If a model solves the task, runs inside the workflow, can be self-hosted, and costs a fraction of the alternative, the origin story becomes secondary for many workloads. Not irrelevant. Secondary.

This is why GLM-5.2 showing up strongly in the Databricks benchmark matters more than a vendor leaderboard. Real-world internal evidence is harder to wave away.

It says the price-performance curve is moving into production reality.

That is the thing to watch.

Self-Hosted Frontier Models

The next phase is not just cheaper API calls.

The next phase is capable self-hosted models doing frontier-level work inside private systems. That changes the adoption curve. Many companies have not fully deployed coding agents because their code is sensitive, their compliance posture is strict, or their executives do not want proprietary source flowing through external systems.

Self-hosted capable models create a different option.

Now the company can run agents inside its own perimeter. It can inspect logs, constrain permissions, connect to internal systems, benchmark every model against its own repo, and run background agents against tech debt, flaky tests, dependency upgrades, security issues, documentation drift, and migration plans.

This is where the abundance shift becomes real.

When intelligence is expensive, you ration it. You use the premium agent for high-value work. You wait for the human to decide the task is worth spending tokens on. You keep the number of attempts low.

When intelligence is cheap and local, you stop asking whether a task deserves an agent.

You ask how many agents should try. You move from chatting to commanding a fleet of AI agents.

That changes software operations. Every issue can get a first-pass investigation. Every pull request can get multiple independent reviews. Every flaky test can get a background repair attempt. Every security advisory can be mapped against the actual codebase before a human opens the ticket.

This does not remove engineers.

It changes what engineering management is.

The bottleneck moves from typing code to designing work systems. The scarce skills become judgment, taste, architecture, review, evaluation, and the ability to define tasks clearly enough that agents can execute them.

This is why the Databricks methodology is so important. The benchmark is not just a report. It is a template for governing agentic labor.

Capture real work. Convert it into tasks. Hold back the answers. Test outcomes. Seal obvious leakage paths. Compare cost per task. Route accordingly. Repeat as models change.

That is the operating system.

Routing is Winning

In the model-scarcity world, the winner is whoever has access to the smartest model.

In the model-abundance world, the winner is whoever allocates intelligence best.

That means routers. Not just technical routers that send prompts to models based on cost and latency, though those matter. I mean organizational routers too: systems that decide which work should be automated, parallelized, escalated to a senior human, delegated to a cheaper model, or wrapped in a full audit trail.

The Databricks benchmark points directly toward this future. It does not say, “we found the one model everyone should use.” It says the frontier is a portfolio: a mix of tools, models, and harnesses, measured against real tasks.

That is the mature frame.

It also tells us where software companies should invest. Do not just buy seats. Build the measurement layer. Build the eval set. Build the model router. Build context discipline. Build permission boundaries. Build internal datasets from your own PRs. Build cost dashboards that show dollars per successful task, not just tokens per vendor.

The companies that get this right will not merely use AI. They will make AI legible.

And once intelligence becomes legible, it becomes manageable.

Once it becomes manageable, it becomes a line of operations.

The Moat Is Moving

The mistake is thinking this means moats disappear.

They do not.

They move.

The moat is less likely to be “we alone have the model.” That moat is getting shorter. Capability diffuses. Open models improve. Chinese labs optimize. Distillation compresses. Costs fall.

The new moats are closer to the work.

Who has the best proprietary evals? Who has the cleanest internal workflow data? Who can route tasks most efficiently? Who has the tightest harness? Who can make cheap intelligence reliable enough to trust?

That is better for builders and worse for anyone relying on scarcity premiums.

The Databricks benchmark is important because it makes this concrete. Coding agents are not a demo category anymore. They are an operating expense, a labor layer, and a routing problem. Open models are not just philosophical alternatives. They are entering the daily-driver conversation. The task, not the token, is the economic unit.

Most importantly, it shows that companies do not need to wait for the market to tell them what works.

They can measure it themselves.

That is the real frontier now.

Not a single model. Not a single lab. Not a single benchmark.

The frontier is the ability to convert cheap, abundant, increasingly local intelligence into reliable work.

GLM-5.2 is one signal. Databricks’ benchmark is another. Together they point in the same direction: the age of paying blindly for premium intelligence is ending. The age of managing intelligence as infrastructure is beginning.

And once self-hosted open models can do frontier-level work, the question stops being whether AI can help.

The question becomes whether your organization knows how to spend intelligence well.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

That is not the scarce thing anymore.

The scarce thing is knowing where to put it, who should use it, how the work should change, and what platform needs to exist underneath it so the gains actually compound.

Most companies are still treating AI like a tool decision. They ask which model to buy, which chatbot to deploy, which copilot to enable, which vendor to trust. Those questions matter, but they are not the main question.

The main question is simpler and harder:

Where can AI create leverage in the operating system of the company?

That is the offering I developed at RevSystems.

We help companies move from scattered AI experimentation to measurable operating leverage by aligning People, Processes, and Platforms around the work that matters most.

The tool frame is too small. AI does not create value because someone has access to a model. It creates value when a workflow changes. It creates value when a team can produce more with the same headcount, make better decisions with the same data, respond faster to the same market, or serve more customers without adding the same amount of operational drag.

The unit of change is not the prompt.

The unit of change is the workflow.

That is where most AI programs break. A company gives people tools, then waits for transformation to appear. Some people experiment. A few power users get faster. A few teams build clever internal demos. Leadership hears enough anecdotes to believe something is happening, but not enough evidence to know what changed.

Activity goes up. Leverage does not.

The reason is that AI adoption is not the same as AI leverage. Adoption means people are using the tool. Leverage means the business system is getting stronger.

Those are different games.

People

The first pillar is People, because AI changes the shape of work before it changes the org chart.

Every team now needs to answer questions it did not have to answer before. What work should remain human? What work should be delegated? Who reviews AI output? Who owns the final judgment? What level of quality is acceptable? What risks are tolerable? What skills now matter more than they did two years ago?

The winners will not simply be the companies with the most AI licenses. The winners will be the companies with the clearest human judgment loops.

AI increases the amount of work that can be attempted. That sounds purely positive until you see what happens inside a real company. More drafts. More analyses. More campaigns. More reports. More ideas. More automation. More noise.

Without better judgment, AI creates abundance without direction.

So the People work is not just training. Training is part of it, but training alone is too narrow. The real work is role design, decision rights, capability building, leadership behavior, and adoption discipline.

People need to know when to use AI, when not to use it, how to inspect it, how to improve it, and how to build repeatable ways of working with it. Managers need to know how to evaluate output when production costs collapse. Leaders need to know how to set priorities when every team can suddenly generate more work than the company can absorb.

This is the human side of leverage.

Not inspiration. Not generic enthusiasm. Operating clarity.

Process

The second pillar is Process, because AI does not belong on top of broken workflows.

If a process is slow, vague, political, duplicative, or poorly measured, adding AI often makes the dysfunction faster. It can accelerate confusion. It can produce cleaner-looking artifacts from the same bad inputs. It can give leadership the feeling of progress while the underlying system stays stuck.

The right move is to redesign the workflow around human-AI collaboration.

That means mapping how work actually moves. Where does demand enter the system? Who touches it? Where does it wait? Where does quality get checked? Where does context get lost? Where are people doing manual translation between systems? Where are high-value employees spending time on low-judgment tasks?

Then we ask: what should AI draft, search, summarize, compare, enrich, route, monitor, recommend, or execute?

This is where the value of AI starts to become concrete.

A sales team does not need AI. It needs faster account research, cleaner follow-up, better call prep, sharper deal inspection, and less manual CRM hygiene.

A customer success team does not need AI. It needs earlier risk detection, better renewal preparation, faster knowledge retrieval, and more consistent customer communication.

A finance team does not need AI. It needs variance explanations, scenario analysis, policy checks, and less spreadsheet archaeology.

A leadership team does not need AI. It needs a better operating cadence, faster synthesis, and clearer visibility into what is actually happening.

The process lens turns AI from a vague capability into a practical redesign of work.

Platform

The third pillar is Platform, because AI leverage depends on the machinery underneath the workflow.

This is the part many executives underestimate. They see the model and miss the system. But the model is only one layer. The real platform includes data access, permissions, integrations, memory, evaluation, security, workflow orchestration, reporting, and governance.

A model without context is a guessing machine.

A model with the right context, tools, permissions, and feedback loops becomes part of the company’s operating infrastructure.

That does not mean every company needs a giant AI platform build. Most do not. But every company needs to know whether its current technology stack can support the workflows it wants to change.

Can the AI access the right data? Can it act inside the right systems? Can it respect permissions? Can it be monitored? Can the output be evaluated? Can humans intervene? Can the workflow be repeated? Can the gains be measured?

If the answer is no, the company does not have an AI strategy. It has a collection of experiments.

The Platform work is about making the technology stack usable for leverage. Sometimes that means cleaning up CRM data. Sometimes it means connecting knowledge systems. Sometimes it means designing agent workflows. Sometimes it means choosing fewer tools, not more. Sometimes it means creating a governance layer so teams can move faster without creating unmanaged risk.

The goal is not technological sophistication for its own sake.

The goal is operational power. Leverage.

Two Front Doors

There are two natural ways to bring this offering into the market.

The first is the CEO front door.

For CEOs, the message is enterprise leverage. AI is now a board-level operating question because it touches productivity, margin, speed, customer experience, and competitive position. The CEO does not need a tour of every tool. The CEO needs to know where AI can make the company meaningfully stronger.

The CEO discussions I have focus on different areas vs the Operators I speak with.

CEOs want to know: where are the three to five places AI can most improve the business? Where are we wasting human capacity? Where are we too slow? Where is quality inconsistent? Where are we blocked by data, process, or ownership? What should we do in the next 90 days?

The deliverable is not a glossy transformation deck. It is an operating roadmap: priority workflows, business cases, owners, metrics, governance, platform gaps, and first-wave pilots.

The second place I focus is Revenue Operations. I’ve been a RevOps Builder for years.

Mastering Revenue Operations

Engineering and building powerful and efficient revenue engines.

By Matt McDonagh

For RevOps, the message is revenue execution infrastructure. Revenue teams are full of hidden manual work. Lead routing, CRM hygiene, forecasting, territory planning, enrichment, campaign handoffs, pipeline inspection, QBR prep, renewal tracking, win-loss analysis, rep coaching, and customer segmentation all contain repeatable cognitive labor.

AI can help, but only if the revenue engine is designed for it.

It answers: where can AI reduce manual work, improve conversion, clean up data, speed up handoffs, and give leadership better visibility?

This is a practical buyer. RevOps does not want theory. RevOps wants cleaner systems, better process compliance, faster reporting, stronger forecast confidence, and fewer hours wasted moving information from one place to another.

The same People, Processes, and Platforms model applies. The entry point changes.

For the CEO, AI is operating leverage. For RevOps, AI is revenue leverage.

Both matter.

The Benefits of True AI Leverage

The benefits are straightforward because they map to the real constraints companies feel every day.

First, speed. AI can compress the time between question and answer, request and response, idea and draft, signal and action. But speed only matters when the workflow is pointed at something valuable.

Second, capacity. Teams can handle more work without adding the same amount of headcount. This does not mean replacing everyone. It means removing low-judgment work from high-judgment people so their time compounds.

Third, quality. AI can make good teams more consistent by giving them better first drafts, better checks, better retrieval, and better operating memory.

Fourth, visibility. When workflows become more structured, leadership gets a clearer view of where work stands, where bottlenecks form, and where the system is leaking value.

Fifth, scalability. A company that redesigns work around AI can grow without recreating every manual process at a larger size.

That is the real prize.

Not novelty. Not demos. Leverage.

The companies that win with AI will not be the ones that talk about it the most. They will be the ones that make the deepest changes to how work gets assigned, performed, reviewed, measured, and improved.

They will align their People so judgment stays sharp.

They will redesign their Processes so AI enters the real flow of work.

They will strengthen their Platforms so tools become infrastructure instead of clutter.

AI is not a magic layer you sprinkle across the company.

It is a new labor layer that has to be wired into the operating model.

That is why this offering exists. It gives companies a practical way to find the leverage, build the roadmap, and move from experimentation to execution.

The age of asking whether AI matters is over.

The only serious question now is where it belongs in the work.

If you would like my help designing and building your revenue engine and the rest of your business with AI, just reach out!

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

The argument was simple.

For most of modern history, specialization won. The world became too complex for one person to understand every domain deeply, so we built careers, companies, institutions, and status ladders around narrow expertise.

Then AI changed the terrain.

Suddenly the person with broad context, taste, judgment, and curiosity could command specialist intelligence on demand. The generalist was no longer limited by what they personally knew how to execute. They could stand above the system, understand the shape of the problem, and direct the specialists.

I used the metaphor of the conductor.

The specialist plays the violin. The model plays the cello. The analyst plays the trumpet. The engineer plays percussion. The conductor does not need to be the best at every instrument. The conductor needs to understand the music.

That was true. It’s still true.

But it is no longer big enough.

The instruments have started becoming workers.

The models no longer just answer. They inspect. They remember. They plan. They call tools. They browse files. They run commands. They test outputs. They compare approaches. They work in parallel. They continue across long-running threads. They take a messy objective and return an artifact.

This is the next phase.

The generalist won once by orchestrating specialist intelligence.

Now they win again by operating persistent agentic labor.

The Unit Of Work Has Changed

The chatbot era trained us to ask better questions.

The agent era trains us to assign better work.

That distinction sounds small until you feel it in practice.

A question produces a response. A task produces a result.

A workstream produces leverage.

This is the real shift. The important interface is no longer the prompt as a clever sentence. The important interface is the work order as a structured command.

Not:

“Explain this market to me.”

But:

“Research this market, identify the five most important companies, compare their business models, pull recent funding and revenue signals, find the key open questions, create a memo, and flag where the evidence is weak.”

Not:

“Help me write code.”

But:

“Inspect this repo, find the source of the bug, propose three likely causes, implement the safest fix, run the relevant tests, update the docs, and summarize the tradeoffs.”

Not:

“Give me ideas.”

But:

“Generate ten strategies, pressure test each one from the perspective of a customer, competitor, investor, and operator, then rank them by upside, feasibility, and time to impact.”

That is not a conversation.

That is delegation.

We’ve moved from chatting to commanding.

The prompt is becoming the work order. The thread is becoming the workspace. The agent is becoming the production unit.

This changes the economics of thinking.

For the first time, cognitive labor is starting to behave like cloud compute. You do not need to hire one human for every branch of exploration. You can spin up attempts. You can run variations. You can compare outputs. You can ask one agent to build and another to critique. You can ask one to research deeply and another to find what the first one missed.

The scarce resource is no longer the first draft.

The scarce resource is judgment.

From Conductor To CEO

The conductor metaphor was right for the age of model orchestration.

But the better metaphor now is the CEO.

Not the celebrity CEO. Not the corporate bureaucrat. The real CEO function.

Define the mission.

Allocate resources.

Choose the right people for the right work.

Create operating systems.

Review performance.

Kill weak projects. Double down on strong ones.

That is what high-agency people are learning to do with agents.

They are not asking AI one question at a time. They are building portfolios of attempts.

One agent explores the technical path.

One agent explores the market path.

One agent writes the memo.

One agent attacks the assumptions.

One agent turns the memo into a customer-facing artifact.

One agent builds the spreadsheet.

One agent checks the numbers.

One agent turns the whole thing into a decision.

The human is no longer the person doing every step directly.

The human is the person designing the system that does the work.

This is where the neo-generalist becomes extremely dangerous.

The specialist can use AI to go deeper inside a narrow domain. That is powerful.

But the generalist can use AI to coordinate across domains. That is more powerful. The generalist can move from strategy to code, from finance to product, from customer psychology to distribution, from legal structure to operational process, from narrative to execution.

Not because they personally replaced every expert.

Because they can now summon, supervise, and synthesize machine labor across the whole map.

That is the second revenge of the generalist. Wait til you see what AI harnesses are going to allow generalists to accomplish.

The Harness Is The Agent

The public still talks about models as if the model is the whole story.

That is wrong.

The model matters enormously. Better reasoning, better coding, better tool use, better multimodal understanding, lower cost, longer context, faster inference. All of that matters.

But the breakthrough is not just the model.

The breakthrough is the harness. Codex by OpenAI is the best harness on the market right now. It makes sense that OAI are focusing on operationalizing their models and creating leverage with them.

A model sitting in a chat window is intelligence trapped behind glass.

A model inside a harness can act.

It can read files. It can edit artifacts. It can run tests. It can call APIs. It can use a browser. It can remember durable preferences. It can follow project instructions. It can operate inside a sandbox. It can create diffs. It can ask for approval. It can spawn subagents. It can work in the background. It can return when the task is done.

That is a different species of tool.

Codex is important because it makes this visible.

Yes, it begins in software. Of course it does. Software is the perfect first battlefield. Code is text. Repos are structured. Tests exist. Logs exist. Diffs exist. The whole environment is already legible to machines.

But coding is not the final category.

Coding is the wedge.

Once you understand the pattern, it expands everywhere.

The same harness logic applies to legal research, financial modeling, sales operations, content production, diligence, recruiting, customer support, internal analytics, compliance, procurement, and executive operations.

The agent needs context.

The agent needs tools.

The agent needs permissions.

The agent needs memory.

The agent needs feedback.

The agent needs evaluation.

The agent needs an environment where work can be attempted, checked, corrected, and shipped.

That is what the harness provides. This is why “prompt engineering” was always too narrow.

The serious skill is ontological architecture.

A prompt is temporary.

A workflow is reusable.

A chat is isolated.

A memory is persistent.

A response is information.

A harnessed agent is true labor.

Memory Changes The Relationship

Memory is one of the most under-appreciated breakthroughs in agents.

People think memory means the system remembers your name, your tone, or a preference. That is the least interesting version.

The real power of memory is operational continuity.

The agent remembers how you work.

It remembers the structure of your projects.

It remembers recurring workflows.

It remembers your standards.

It remembers the failure modes you already corrected.

It remembers the context you should not have to repeat. That changes the relationship from tool to teammate.

A tool resets every time you pick it up. A teammate accumulates context.

That accumulation is the beginning of compounding.

In the old chat world, every session began with context loading. You had to explain the company, the project, the preference, the audience, the constraints, the past decisions, the style, the weird edge cases.

That is friction. Friction kills usage.

Friction keeps AI trapped as an occasional assistant instead of becoming part of the production system.

Memory lowers that friction.

Skills lower it further.

Reusable instructions lower it further.

Plugins and connected tools lower it further.

Eventually, the agent does not just know the task.

It knows the operating environment.

That is when things start to get weird.

Because once agents have memory and tools, work can move from episodic to continuous.

The agent can monitor.

The agent can revisit.

The agent can compare.

The agent can improve the same workflow over time.

The agent can become part of the rhythm of the organization.

That is not a better chatbot. That is a new labor layer.

Wealth Systems

Five Forecasts for the Future of Work

I don’t think people have metabolized what is about to happen to work…

Read more

a month ago · 2 likes · Matt McDonagh

Long-Running Tasks Change Ambition

The other major unlock is duration.

Short tasks create short thinking.

Long-running tasks create ambition.

When AI only works in single-turn answers, you naturally ask it for things that fit inside a single answer. Summaries. Drafts. Explanations. Lists. Ideas.

Useful, but limited.

When agents can work longer, you ask different questions.

You stop asking for a paragraph and start asking for a project.

You stop asking for an answer and start asking for an investigation.

You stop asking for a suggestion and start asking for a working artifact.

This matters because most valuable work is not a single act of intelligence.

It’s a chain in a sequence:

1. Research.

2. Planning.

3. Execution.

4. Review.

5. Correction.

6. Integration.

7. Shipping.

The old AI interface helped with pieces of that chain.

The new agentic interface starts absorbing the chain itself.

That changes what humans attempt.

If it takes eight hours of human work to explore a path, most people will not explore five paths. They will pick one, maybe two, and live with the uncertainty.

If an agent can explore five paths in parallel, the frontier moves.

You do not need to guess the best approach upfront.

You can run the portfolio.

This is how intelligence becomes abundant.

Not because every answer is perfect. Because the cost of trying collapses.

Cost Performance Changes Behavior

The most important economic fact about AI is not that intelligence gets better.

It is that useful intelligence gets cheaper.

When something is expensive, you conserve it. When something becomes cheap, you waste it productively.

This is what happened with compute. This is what happened with bandwidth. This is what happened with storage. This is what happens with every foundational technology that drops in cost fast enough.

At first, you use it carefully.

Then you use it casually.

Then you redesign the system around the assumption that it is abundant.

Cognitive labor is entering that phase.

The old behavior was scarcity behavior.

Ask one perfect question.

Get one good answer.

Use it carefully.

The new behavior is abundance behavior.

Launch ten attempts.

Make them compete.

Have agents critique each other.

Run the same problem through different frames.

Search the possibility space.

Select the strongest output.

Synthesize the best pieces.

Ship.

This is the core argument.

The winner is not the person who asks AI one better prompt.

The winner is the person who designs a portfolio of attempts, lets agents explore, then applies human judgment to select, synthesize, and ship.

That is the new leverage loop. And it is available to individuals before institutions know what to do with it.

The Bottleneck Moves To Taste

When cognitive labor becomes abundant, output explodes.

That sounds good.

It is also dangerous.

Abundance creates noise.

Agents can produce bad work faster than humans can produce bad work. They can create plausible nonsense, duplicate effort, miss context, overfit to instructions, confidently drift away from the real objective, or flood the zone with artifacts that look finished but are not actually true.

This is why judgment becomes more important, not less.

The naive view says AI reduces the value of human expertise.

The opposite is true at the frontier.

AI reduces the value of raw execution. It increases the value of knowing what good looks like.

Taste becomes a production function.

Verification becomes a managerial skill.

Context becomes capital.

The person who cannot judge outputs will drown in them.

The person who can judge outputs will compound.

This is the paradox of abundant cognitive labor.

The more the machine can produce, the more valuable the human editor becomes. This is why AI will not kill software engineering. The opposite in fact: there is much more software now to maintain and scale!

The more agents can explore, the more valuable the human allocator becomes.

The more drafts appear, the more valuable taste becomes.

The more work gets automated, the more important it is to know what work should exist in the first place.

This is why the future does not belong to passive users.

It belongs to people with agency.

The New Human Capital Stack

The value of raw execution is falling.

The value of direction is rising.

The value of judgment is rising.

The value of taste is rising.

The value of verification is rising.

The value of synthesis is rising.

The value of workflow design is rising.

The value of proprietary context is rising.

The value of coordinating parallel streams of machine labor without losing the plot is rising the most.

That is the new human capital stack.

In the industrial economy, capitalists owned machines and workers operated them. In the knowledge economy, companies owned distribution and workers performed cognitive tasks.

In the agentic economy, high-agency individuals will operate machine labor directly.

Think about the implications of that.

A founder with agents can simulate parts of a company before hiring the company.

A writer with agents can operate like a media team.

An investor with agents can run continuous diligence.

A lawyer with agents can multiply research capacity.

A RevOps leader with agents can connect marketing, sales, customer success, finance, product, and data into one operating loop.

A student with agents can learn through personalized research, tutoring, testing, and project work.

A generalist with agents can become an institution.

That is the real disruption. Not that every job disappears overnight. That the minimum viable team size collapses.

That the ambitious individual gets access to a layer of cognitive labor that used to require a staff.

That small teams can do things that previously required departments.

That departments can do things that previously required whole companies.

That the shape of organization itself begins to change.

The Portfolio Is The New Prompt

The prompt was the first interface. The portfolio is the next one.

This is how serious work will happen.

You define the objective. You decompose the problem. You create multiple workstreams. You give each agent context and constraints.

You let them explore. You force comparison. You review the evidence.

You synthesize.

You ship.

Then you capture the workflow so it can run again.

That last part matters.

If you do not capture the workflow, you are still just chatting.

If you do capture it, you are building infrastructure.

This is where skills, memories, automations, templates, project instructions, and tools become so important. They turn one-off intelligence into repeatable operating leverage.

The first time you run an agentic workflow, you get output.

The tenth time, you get a system.

The hundredth time, you get an advantage.

This is how compounding starts.

Not with a magical prompt. With a repeatable loop.

What To Do Now

The practical mandate is simple.

Stop asking, “How can I use AI more?”

That question is too weak.

Ask better questions:

• What work do I repeat?

• What decisions require too much manual research?

• What workflows depend on context trapped in my head?

• What tasks have clear inputs and outputs?

• What projects would I explore if the cost of trying were 90% lower?

• What workstreams could run in parallel?

• Where do I need a builder, a critic, a researcher, an editor, and an operator working at the same time?

Then build the system.

Write the operating procedure.

Attach the context.

Define the output.

Create the review loop.

Run multiple attempts.

Compare results.

Save what works.

Delete what does not.

Improve the workflow.

Repeat.

This is not about using AI as a novelty. This is about converting your work into agentic infrastructure.

The people who understand this will feel like they have more hours in the day.

Because they will.

Not biologically.

Operationally.

The Top 1% of Codex users (OAI’s agent harness) average 71 hours of agent runtime per day. They are literally getting more hours in the day. And the night. Even when they sleep.

They will run more attempts than everyone else. They will explore more paths. They will learn faster. They will ship more. They will compound context while others are still typing isolated prompts into empty chat windows.

That is the divide.

The Generalist Returns Again

The first revenge of the generalist was access to specialist intelligence.

The second revenge is command over agentic labor.

The generalist was built for this moment.

Broad context matters.

Taste matters.

Curiosity matters.

Pattern recognition matters.

The ability to move between domains matters.

The ability to ask, “What is the actual objective here?” matters.

The ability to see the whole system matters.

The specialist age rewarded depth inside a narrow lane.

The agentic age rewards people who can connect lanes, direct labor across them, and synthesize the result into reality.

This does not mean expertise disappears. It means expertise gets surrounded by leverage.

The best specialists will become terrifying.

The best generalists will become operating systems.

That is the new frontier.

Not human versus AI.

Human plus persistent machine labor.

Human plus memory.

Human plus tools.

Human plus parallel agents.

Human plus the ability to ship in constantly accelerating loops.

Cognitive labor is becoming abundant.

The scarce thing now is knowing what to do with it.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

It may be agents per megawatt.

That sounds strange if you still think of AI as software. It sounds less strange if you think of AI as labor. And it starts to feel inevitable once you see that inference is becoming an input into almost every form of knowledge work.

For the last century, the richest countries were the ones that could combine people, capital, education, institutions, energy, and coordination into high human productivity. We measured the result as GDP per capita because the human worker was the main unit of economic action. The better the system around that person, the more each person could produce.

That frame is about to get too small.

We are entering a world where a growing share of useful work is performed by agents: model-driven systems with context, tools, permissions, memory, evaluation loops, and the ability to run tasks across time.

Not chatbots. Not autocomplete. Not one-off question answering. Agents.

The unit of work is changing.

A human used to be the smallest practical unit of judgment-bearing labor. Soon, the smallest practical unit will be an agent-hour. Then an agent-minute. Then a swarm of task-specific workers spun up, pointed at a problem, evaluated, merged, and shut down.

When that happens, economic power changes. A country with cheap power, efficient data centers, frontier models, strong orchestration, and the institutions to deploy them will not just have better software. It will have a larger effective workforce.

That workforce will not sleep.

It will not commute.

It will not wait for a meeting invite.

It will be constrained by energy, compute, model quality, data rights, workflow design, and human judgment. But the bottleneck will no longer be only how many educated humans a society can produce. It will be how many useful agents it can run, how cheaply it can run them, and how well it can aim them.

That is a massive shift… a transformation on a societal scale.

We will see inference become a majority input into GDP.

GDP Per Capita Was a Human Era Metric

GDP per capita made sense because people were the scarce productive substrate.

If you wanted more doctors, analysts, or engineers, you trained, hired, educated, or imported more people.

Capital helped. Machines helped. Software helped. But the human was still the point where perception, judgment, communication, and accountability came together.

A factory could multiply a worker. A spreadsheet could multiply an analyst. A search engine could multiply a researcher. SaaS could multiply a manager. But these tools mostly raised the output of a human operator. The person remained in the loop at the center of the task.

Agents invert that relationship.

The human increasingly moves from doing the task to designing the workstream. The operator defines the objective, supplies context, gives tools, creates constraints, selects strategies, evaluates outputs, and decides what ships. The agent does more of the middle.

That middle is enormous.

Most knowledge work is not pure genius. It is reading, summarizing, comparing, formatting, researching, drafting, reconciling, classifying, planning, testing, translating, checking, and following through. It is moving information from one shape to another with enough judgment to avoid obvious failure.

That is exactly the zone agents are entering.

The old economy asked how much output one person could create with a set of tools. The new economy asks how many competent agents one person, one company, or one country can coordinate toward useful ends.

This is why GDP per capita begins to feel incomplete. It tells you how much output is produced per human resident. But it does not tell you how many non-human workers are running inside the system. It does not tell you how much inference capacity a country can direct. It does not tell you whether a small population can operate a vast synthetic workforce.

The denominator is changing.

Work Becomes an Energy Problem

If agents are labor, inference is the fuel.

Training gets the attention because training runs are dramatic. They are large, expensive, and easy to mythologize. But training is the creation of capability. Inference is the use of capability. The more AI moves from demos into work, the more the economic center of gravity moves toward inference.

Every task an agent performs consumes compute.

Compute consumes energy.

Energy flows through data centers.

Data centers become factories for cognitive labor.

This is the part many people still miss. The AI economy is not weightless. It is deeply physical. It needs power generation, grid interconnects, chips, cooling, land, fiber, substations, transformers, supply chains, permitting, and resilience. It needs engineers who understand that the cloud lives somewhere.

The abstract economy is about to become visibly industrial again.

For countries, this changes the strategic map. Energy policy becomes labor policy. Grid capacity becomes workforce capacity. Data center efficiency becomes a national productivity variable. Chip supply becomes not only a technology issue, but a labor market issue.

If your society can support more inference per watt, it can support more agents per megawatt. If it can support more agents per megawatt, it can support more concurrent work. If it can support more work in parallel, it can search more solution space, run more experiments, serve more customers, write more code, monitor more systems, design more products, and compound faster. And faster.

Now imagine a country with the energy, compute, models, and operational stack to run 20 trillion concurrent agents. Or 22 trillion. The exact number matters less than the shape of the thought experiment. A workforce that large, running 24/7, pointed at science, engineering, logistics, education, drug discovery, software, national administration, defense, manufacturing, metamaterials, and business formation, is not a normal productivity improvement.

It’s a discontinuity.

The biggest proliferation of technology in human history may not come from one invention. It may come from the sudden ability to apply persistent cognitive labor to every bottleneck at once.

Not one research team trying one path.

Millions of agent teams trying millions of paths.

Not one analyst producing a report.

Thousands of agents reading the underlying data, generating competing interpretations, testing assumptions, and escalating the five that matter to a human.

This is why agents per gigawatt may become a civilizational metric. It compresses the physical and cognitive reality into one phrase. How much useful work can you extract from energy?

That has always been the question.

AI just makes it explicit.

Agents Outnumbering Humans

This is the last year where a majority of the knowledge workforce is human.

That claim sounds extreme because we are used to counting workers as people with jobs. But the better question is not how many people are employed. The better question is how many workers are doing economically relevant knowledge tasks.

By that measure, 2027 could be the year agents cross the line.

Not because agents will replace every employee. Not because offices disappear. Not because companies suddenly become empty shells. The transition will look stranger than that. Humans will still own goals, relationships, taste, accountability, politics, trust, and many forms of embodied work. But the number of agentic workstreams running across the economy could exceed the number of human knowledge workers much faster than institutions are ready to admit.

A single person may run ten agents in the background.

A small team may run hundreds.

A large enterprise may run millions of short-lived agents across customer support, sales operations, data cleaning, software testing, internal search, finance, legal review, security monitoring, procurement, and analytics.

Most of these agents will not look like employees. They will appear as tasks, runs, workflows, background jobs, assistants, automations, copilots, queues, and systems. That is why the shift will be undercounted.

The economy will add a synthetic labor layer before it knows how to measure it. Traditional employment statistics will still describe the human labor market. GDP will still describe output. But underneath, a new workforce will be forming.

The first sign will not be mass unemployment. The first sign will be output that no longer matches headcount.

Tiny teams will launch products at a speed that used to require departments. Individual operators will sustain a volume of research, writing, analysis, outreach, and coordination that used to require staff. Enterprises will quietly automate the middle layers of work and wonder why their org charts feel less real than their workflow graphs.

The boundary between a tool and a worker will blur. Token economics, coined tokenomics, is one of the most import disciplines of the future because of this.

Wealth Systems

Token Economics Will Drive Everything

Brian Armstrong, CEO of Coinbase made an X post recently containing a blueprint for the next business operating system…

Read more

a month ago · 2 likes · Matt McDonagh

When a system can receive a goal, gather context, use tools, make intermediate decisions, ask for help, evaluate its own output, retry, and return a result, it is not just software in the old sense. It is a worker-like process. It may be narrow. It may be brittle. It may need supervision. So do many human processes.

The point is not that agents become human. This is not a Pinocchio story. This a production story.

The point is that many tasks never required humanity in the first place. They required adequate cognition, context, and follow-through.

That is what is being industrialized.

The New Production Function

The agent economy has a different production function.

The old formula was roughly people plus capital plus process plus technology.

The new formula is judgment plus context plus tools plus inference plus evaluation. Judgment decides what matters. Context gives the agent access to the relevant world. Tools let it act. Inference performs the cognitive work. Evaluation tells the system whether to keep, retry, escalate, or discard.

Most people focus on the model.

That is understandable. The models are new, they are constantly evolving, but focusing solely on the model is too narrow. The model is the engine. The work happens in the system around the engine.

An agent without memory forgets. An agent without tools talks. An agent without permissions cannot act. An agent without evaluation drifts. An agent without context hallucinates around the edges. An agent without a human operating model becomes noise at scale.

This is why the winners will not simply be the people with access to the best model. They will be the people and institutions that build the best harnesses.

The harness is the agent.

A model in a chat box is potential energy. A model connected to data, tools, workflows, tests, approval gates, and feedback loops is productive energy. That conversion is where the leverage lives.

At the national level, the harness includes power, chips, data centers, institutions, capital markets, and trust. At the company level, it includes clean data, mapped workflows, clear permissions, evaluation suites, and managers who can specify work. At the individual level, it includes taste, tool fluency, and the ability to review more than you personally produce.

The people who win in this world are not passive consumers of AI. They are conductors of work.

They know how to break a vague goal into parallel attempts. They know what can be delegated and what must be judged. They know when to ask for breadth and when to force depth. They know how to compare outputs. They know how to create feedback.

They know how to protect the final mile from slop.

Execution capacity used to be scarce. Now that we are generating at the speed of light, evaluation capacity becomes scarce.

That changes the status of taste.

Taste used to be seen as soft. In the agent economy, taste is an economic bottleneck. If you can generate 100 plans, 100 product concepts, 100 customer segments, 100 investment memos, or 100 design directions, the scarce skill is knowing which three are worth attention and which one deserves resources.

Abundance punishes weak judgment.

When output is cheap, selection matters more.

Leverage, Leverage, Leverage

The same shift happening to countries will happen to individuals, just at a different scale.

Open your phone. Review 10 strategic plans before breakfast. Choose where to deploy 100 hours of deep work. Send agents to research the market, draft the memo, build the prototype, test the copy, reconcile the spreadsheet, summarize the calls, identify the risks, and prepare the next set of decisions.

That is not science fiction. It is the natural endpoint of current behavior.

The best users are already squeezing impossible amounts of output into ordinary days. They are not doing it by typing faster. They are doing it by changing their relationship to work. They are turning tasks into work orders. They are running parallel attempts.

They are treating AI less like a search bar and more like a staff of workers. They have AI planners and AI orchestrators driving AI workers.

You can see the early pattern in people who produce 70 or more hours of output in a day. The point is not that every hour is equal. Some output is shallow. Some needs cleanup. Some is discarded. The point is that the human is no longer personally touching every intermediate step.

The operator creates direction. The agents create surface area. The operator selects, edits, combines, and ships.

That loop can scale.

At 10:1 leverage, one strong person starts to look like a small team.

At 100:1 leverage, a small team starts to look like a company.

At 1000:1 leverage, the old categories break.

The 1000:1 leverage era is not about doing 1000 times more random work. It is about applying a large synthetic workforce to the parts of your life and business where more attempts actually compound. Research. Sales. Product. Investing. Writing. Recruiting. Learning. Code. Operations. Negotiation prep. Market mapping. Scenario planning.

Many people are used to shrinking their goals to fit their available hours. They ask what can I personally get done this week? That question made sense when labor was the binding constraint.

But if agents can execute the middle of the work, the better question is what deserves a fleet?

It forces prioritization. It forces taste. It forces the operator to become clearer about the desired future state. You cannot effectively command agents if you do not know what you want, what good looks like, or what tradeoffs you are willing to accept.

AI does not remove agency from the human. It raises the price of weak agency.

If you have no direction, more agents give you more drift. If you have no taste, more output gives you more confusion. If you have no standards, more speed gives you more mess. But if you have direction, taste, and standards… agents become pure leverage.

That is why this moment is so asymmetric. Passive users will ask for answers. Active operators will assign work. Passive users will wait for perfect agents. Active operators will use imperfect agents inside disciplined workflows.

The future arrives first as a behavior pattern.

The future of work looks very different to the past.

Wealth Systems

Five Forecasts for the Future of Work

I don’t think people have metabolized what is about to happen to work…

Read more

a month ago · 2 likes · Matt McDonagh

Agents Per Gigawatt

We should start building the language now because the old language will hide the change.

GDP per capita will still matter. Employment will still matter. Wages will still matter. Human welfare is the point, not a side note. But the productive structure beneath those measures is changing.

Inference is becoming labor.

Energy is becoming cognition.

Data centers are becoming workforce infrastructure.

Agents are becoming the new marginal workers of the knowledge economy.

The prediction may be early in its exact timing. Maybe 2027 is the crossing. Maybe it takes a little longer for the measurements to catch up. But the direction is already visible. The human-majority knowledge workforce is not a permanent fact of nature. It’s a historical condition created by the limits of prior technology.

Those limits are changing.

At the macro level, countries will compete on agents per gigawatt.

At the company level, firms will compete on workflows per employee.

At the individual level, operators will compete on leverage per decision.

This is the new labor stack.

The most important question is no longer simply how productive is each person?

It is how much useful work can a person, team, company, or country command?

That answer will define the next economy.

The agent workforce is coming online. It will be measured first in tasks, then in workflows, then in energy, then in GDP.

And eventually the obvious thing will become obvious.

The future of productivity is not just output per person.

It is agents per gigawatt, aimed by people with judgment.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

Labor intensive operations are massive sinks of capital, time, and human potential.

An architect of automated systems eliminates the bottleneck of human labor, allowing capital to flow directly into perfect, instantaneous execution. You dictate reality through code.

That all means code is the ultimate lever.

And by extension, artificial intelligence is the most powerful lever invented.

It feeds raw energy into all other sciences, technologies, and human efforts. We are standing at the absolute epicenter of a fundamental reorganization of capital, labor, and time. The old paradigms of human productivity are entirely obsolete.

The singularity is already our daily reality.

Systems create leverage. Systems create surface area. Systems create the inevitable outcomes you demand from reality. When you transition from a consumer of software to an architect of autonomous systems, you fundamentally rewrite the laws of physics governing your output.

Agents are the engines of this new world.

Wealth Systems

Five Forecasts for the Future of Work

I don’t think people have metabolized what is about to happen to work…

Read more

a month ago · 2 likes · Matt McDonagh

You do not need more employees. You do not need more hours in the day. You need a fleet of highly effective autonomous agents executing your strategy flawlessly across every digital domain. The individuals who understand how to build and deploy these digital workers will capture all the economic value of the next decade.

Building them requires absolute precision. Review the 4-part “Introduction to Agentic Engineering” series below to build your baseline. Then we will discuss the 5 keys to building effective agents.

Key 1: Ruthless Contextual Constraints

An agent without a rigidly defined persona is a severe liability. You must lock the model into a hyper specific worldview with immutable rules, distinct boundaries, and explicit goals. Ambiguity breeds hallucination, inefficiency, and catastrophic system failure. When you give a model too much freedom, it will invariably drift from the core objective and waste computational resources.

Precision is the only acceptable standard.

How do you achieve this?

We define the role. We define the constraints. We define the exact format of the desired output. When you engineer the prompt architecture, you are literally programming the neurobiology of your digital worker.

Weak instructions guarantee weak results.

A high agency system requires absolute clarity to operate autonomously. You must strip away all unnecessary context, eliminate conflicting directives, and provide a singular focus. The architecture must reject irrelevant inputs, focus entirely on the core objective, and execute without hesitation. The agent must exist entirely for the specific task it was born to solve.

Constrain the mind to maximize the impact.

You architect the system prompt to be an unbreakable contract. The agent must internalize its identity, recognize its limitations, and operate strictly within the provided framework. If it deviates from the contract, the entire workflow falls apart in unpredictable ways. The foundation of agentic behavior is an ironclad definition of self.

Identity drives behavior.

Engineering this context requires a deep understanding of latent space manipulation. You are shaping the probability distribution of the model to eliminate useless branches of logic. You are forcing the cognitive engine down a narrow corridor of high value execution. You are building walls around the thought process to ensure the output hits the target every single time.

Constraints breed undeniable competence.

You must view the system prompt as the foundational constitution of your digital entity. Every single word carries weight, every instruction shapes the neural pathways, and every exclusion defines the boundary of the agent’s reality.

A poorly drafted prompt leaves massive vulnerabilities in the execution logic, allowing the agent to wander into useless tangents.

You are defining the physics of the agent’s existence. This is almost as important as the next key!

Key 2: Interoperable Tooling and Action Spaces

Read more

It drives raw, unadulterated energy into all other sciences, technologies, and industrial efforts. We are witnessing the deployment of an intellectual engine that alters the fundamental physics of work. Systems create leverage and surface area you can use to drive massive, outsized outcomes.

This is the new reality of leverage.

You must build systems to harness this raw energy. Architectures that capture the explosive growth of machine intelligence.

You must engineer pipelines that convert computational cycles into tangible reality. The current models exhibit extraordinary competence in producing content, code, and structural frameworks at zero marginal cost.

They generate the world.

That said, a fundamental gap remains in the architecture of machine intelligence. The models are rapidly advancing in complex reasoning, logic processing, and algorithmic design. They synthesize complex variables into coherent outputs faster than any biological brain. But they lack the aesthetic, cultural, and contextual compass we define as “taste”.

Taste is the ultimate moat. Taste is the culmination of lived experience, cultural nuance, and emotional resonance. The algorithms calculate the optimal path, but they cannot feel the pulse of a market. They output a thousand software designs, but they cannot recognize the singular masterpiece that resonates with the human soul.

The machine provides the menu.

You make the choice.

The Mechanics of Permutations

The true power of the machine lies in its ability to flood the zone with options.

Don’t outsource your brain to AI. You must use AI to generate variations, mutations, and permutations.

When you confront a problem, you no longer rely on a single human spark of inspiration. You force the model to create ten, fifty, or one hundred distinct approaches to the exact same objective.

Volume creates its own quality.

By generating variations, you explore the immediate boundaries of the core idea. By forcing mutations, you introduce calculated errors and novel structures into the standard paradigm. By demanding permutations, you reorder the variables into entirely new operational models.

The human mind tires after the third iteration. The machine runs until you cut the power.

Iteration is now virtually free.

Consider the architecture of a complex data pipeline. A human data engineer defaults to the familiar patterns they learned over the last decade. They build the same extraction, transformation, and loading sequences they always build. The AI acts differently. It generates a dozen distinct schemas based on pure optimization algorithms.

The machine breaks the bias of habit.

First you command the AI to produce an architecture optimized strictly for latency.

Then command it to produce an architecture optimized strictly for cost reduction.

Then command it to produce a design prioritizing fault tolerance above all else.

The system hands you a comprehensive array of blueprints.

You select the weapon.

This principle applies equally to investment banking and capital allocation. You do not build a single financial model. You instruct the intelligence to generate many conceivable permutations of the deal structure. You model the catastrophic downside, the asymmetrical upside, and the stagnant middle ground simultaneously.

You map the entire territory of risk.

Expanding the Horizon of Possibility

Linear thinking produces linear results. You must use AI to think laterally. The models are trained on the entirety of human knowledge, allowing them to draw connections across entirely unrelated disciplines. They map the architecture of biology onto the structures of financial markets. They apply the principles of fluid dynamics to supply chain logistics.

They bridge the impossible gaps.

When you task an AI to solve a problem, you explicitly command it to pull frameworks from outside your industry. You ask it to approach a software engineering bottleneck with the mindset of an urban planner. You instruct it to analyze an investment thesis through the lens of evolutionary biology. The machine easily navigates these disparate domains to surface completely novel solutions.

Innovation requires this synthetic friction.

Use AI to break your rigid mental frameworks.

Use AI to shatter industry consensus.

Use AI to discover the hidden variables driving the outcome. The machine does not respect the artificial boundaries humans place between physics, economics, and art. It views all data as a unified field of information waiting to be connected.

AI sees the universal patterns. It doesn’t have as many of the heuristic biases that come built into our brains, either.

The Ultimate Adversary

Weak ideas survive in a vacuum. You must use AI to red team your strategies and come up with a comprehensive list of things to improve. The machine has no ego to protect and no feelings to hurt. It savagely deconstructs your business model, your code base, and your investment thesis.

It is the perfect, unyielding critic.

You feed your complete plan into the model and command it to find the fatal flaws. You tell it to adopt the persona of your most aggressive competitor, a skeptical regulator, or an impatient customer. The AI immediately highlights vulnerabilities, edge cases, and systemic risks you failed to consider. It provides an objective diagnostic report of your structural weaknesses.

Destruction paves the way for optimization.

In software development, this adversarial process is mandatory. You write the initial logic and immediately deploy the AI to attack the code. It probes for injection vulnerabilities, memory leaks, and inefficient queries. It generates malicious payloads designed to break your specific architecture.

The machine hardens the shield.

In strategic planning, the red team protocol prevents catastrophic failure. You outline your go to market strategy and demand the AI formulate a counter strategy that destroys your business. The model identifies your reliance on a single distribution channel. It highlights the fragility of your supply chain. It calculates the exact capital required for a competitor to price you out of the market.

You fix the cracks before the flood arrives.

The Architecture of Autonomous Loops

Isolated tasks produce isolated gains. Consider using AI to lock these processes together in continuous, self reinforcing loops. You wire the generation engine to the lateral thinking module. You feed the output directly into the red team adversary.

These systems can run autonomously.

The generator creates the permutations. The lateral thinker injects the novel frameworks. The red team tears the concepts apart and dictates the necessary improvements. The refined parameters feed directly back into the generator to start the cycle anew. This creates a high speed flywheel of continuous intellectual evolution.

Speed is the ultimate weapon.

You architect these loops using basic programmatic logic. You write scripts that trigger the generation phase upon the ingestion of new data. You pipe the raw outputs into the adversarial models using automated API calls. The system runs in the background while you sleep, constantly refining and optimizing the core thesis.

The loop compounds intelligence.

Imagine a trading algorithm locked in this architecture. The generator creates thousands of potential trading signals. The lateral module applies sentiment analysis from obscure geopolitical data sources. The red team model simulates historical market crashes to destroy the weak signals. The surviving signals execute automatically in the market.

This is the physics of infinite leverage.

The Supremacy of Human Taste

The loops spin. The generation multiplies. The options expand into infinity. But keep yourself in the loop and make decisions using a very wide and deep menu provided by AI. The machine builds the labyrinth, but you walk the path.

You are the sovereign arbiter.

You (using AI) review the thousands of generated mutations. You filter the lateral connections. You evaluate the red team vulnerabilities. You apply your cultivated taste to select the single exact configuration that aligns with the market reality. The AI does the heavy lifting, the infinite sorting, and the logical derivation.

You apply the human spirit and taste filter.

Taste is the ability to recognize the resonant frequency of an idea. It is the instinct that knows when a product feels right, when a strategy aligns with the cultural moment, and when an investment holds asymmetrical upside. The machine cannot quantify this instinct. It provides you with the most expansive menu in human history.

You point and execute.

A pure data engineer looks at the numbers and chooses the highest value. A high agency architect looks at the menu and applies context. They know that the mathematically optimal user interface might feel alien to the actual customer. They know the perfectly optimized financial structure might alienate the exact founders they want to back.

Taste supersedes pure computation.

The future belongs to the operators who merge these two domains. The individuals who rely purely on human effort will be crushed by the sheer volume of machine output. The individuals who rely entirely on the machine will produce sterile, uninspired garbage. The victors use the machine to generate the universe of possibilities and use their taste to collapse the wave function.

The Blueprint for Execution

This methodology defines the future of all high agency operations. The individuals who master this dynamic command unprecedented power, influence, and capital. They treat AI not as a tool, but as a boundless cognitive utility. They build empires using the sheer force of computational leverage.

The world belongs to the architect.

You build the system. You define the parameters. You set the loops in motion. The machine floods your dashboard with refined, stress tested, and laterally optimized options. You exercise your taste, make the definitive choice, and drive the outcome into reality.

Leverage is absolute in the singularity.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

It’s massive not because it predicts the future.

Because it shows the future already forming inside the logs of the most advanced AI software in the world, used daily by millions.

The paper, The Shift to Agentic AI: Evidence from Codex, studies how people are using Codex across individual users, organizational users, and OpenAI workers. The findings are not subtle. They show the exact moment AI stops being a chatbot and starts becoming a labor system.

This is the transition from asking to delegating.

This is the beginning of the agentic economy.

The most important number in the paper is not that active Codex users grew more than fivefold in the first half of 2026, although that alone is extraordinary. The important number is that among OpenAI workers, Codex now accounts for 99.8% of output tokens across Codex and ChatGPT. Organizational users are at 63.3%. Individual users are at 16.5%.

The frontier has already moved.

At OpenAI, the default interface for work is no longer conversation. It is delegation. The employee does not ask a model to explain a concept. The employee assigns a task, lets the agent inspect files, execute commands, modify artifacts, and return completed work.

That is an entirely different economic object.

A chatbot is a faster search bar. An agent is a junior worker with tools.

The old interface was question and answer. The new interface is command and execution. That distinction sounds small until you map it across every knowledge worker in the economy.

Then it becomes civilization-scale.

Software Is the First Battlefield

Codex begins in software because software is the perfect initial substrate for agentic AI.

Code is digital. Code is modular. Code can be tested. Code has clear artifacts. Code has logs, diffs, compilers, tests, repositories, issues, and deployment pipelines. The entire software production system is already machine-readable.

That means software is the first industry where cognitive labor can be fully wrapped in an agentic execution loop.

But the paper makes clear this is not staying inside software.

Codex users are already using it for documents, spreadsheets, memos, data analysis, research, collaboration, communication, planning, recruiting, sales, product work, and legal workflows. The paper repeatedly shows that the deepest adoption expands beyond the original developer base.

This matters.

The normal public narrative says coding agents are for engineers. That is wrong. Coding agents are the first visible form of a broader work agent. The developer is just the first professional whose daily labor is already close enough to executable text that the machine can absorb the workflow.

The same pattern will move outward.

First the agent writes code.

Then it maintains systems.

Then it reads company documents.

Then it drafts reports.

Then it updates spreadsheets.

Then it coordinates across Slack, email, CRM, calendar, data warehouse, and internal tools.

Then the human is no longer doing the work directly.

The human is directing the system that does the work.

This is why the paper’s distinction between conversational AI and agentic AI is so important. Conversational AI produces responses. Agentic AI produces outcomes.

Responses are information.

Outcomes are labor.

The Complexity Curve Is Moving Up

The most powerful section of the paper is about task complexity.

In December 2025, 35.4% of active individual Codex users sent at least one prompt estimated to require more than one hour of experienced human work. By May 2026, that number reached 70.2%.

Even more important: the share of users sending at least one request estimated to require more than eight hours of experienced human work rose from 2.1% to 25.6%.

Read that again.

A quarter of sampled individual Codex users were already handing off tasks that would take an experienced human more than a full workday.

This is not autocomplete.

This is not a productivity trick.

This is humans learning how to package larger blocks of work into machine-executable assignments.

The prompt is becoming the work order. The thread is becoming the workspace. The agent is becoming the production unit.

The paper also finds that the most complex requests tend to happen at the beginning of threads. That makes perfect sense. The human starts by delegating the broad mission. Then the follow-up turns become supervision, correction, refinement, and integration.

That is exactly how managers work with teams.

The initial instruction defines the objective.

The later interaction manages execution.

The human role is shifting up the abstraction stack.

The New Managerial Class

The most important worker in the agentic economy is not the person who knows the most facts.

It is the person who can design, delegate, verify, and integrate machine labor at scale.

The paper’s concurrency data makes this obvious.

Among individual and organizational users, most people are still using Codex in a relatively linear way. Roughly two-thirds of organizational and individual users did not run concurrent turns during the measured week.

But OpenAI workers are already operating differently.

Only 10.7% of OpenAI users ran a sole workflow at any one time. Nearly 28.6% managed five or more concurrent agents at some point during the measured period.

That is the new labor model.

One human. Multiple agents. Parallel execution. Continuous review.

This is not “using AI.”

This is managing an artificial workforce.

The highest-intensity users are already living in that world. The paper shows that the median OpenAI employee had Codex turns running for 2.5 hours on June 11, 2026. But at the 99th percentile, OpenAI employees ran about 71 hours of agent turns within the average day.

Seventy-one hours of work in one calendar day.

That is only possible when work becomes parallelized through autonomous execution.

A human cannot work 71 hours in a day. A human can manage systems that do.

That is the entire economic transition.

Skills Are the New Operating Procedures

The paper’s section on skills and plugins may be the most underappreciated part.

In Codex, skills allow users to encode reusable instructions, workflows, references, scripts, and procedural context. Plugins package capabilities and integrations. Together, they turn ad hoc prompting into repeatable production infrastructure.

That is the leap.

A prompt is temporary.

A skill is institutional memory.

A plugin is distribution.

The paper finds that skill use rose from 5.4% of active Codex users on March 1, 2026 to 26.6% on June 11, 2026. Among individual users, 25.7% invoked at least one skill in the measured week. Among organizational users, 30.4% did. Inside OpenAI, skill use was nearly universal at 96.2%.

This is exactly what should happen.

At first, people use agents manually. They type instructions over and over again. They paste context. They repeat preferences. They correct the same failure modes. They treat the agent like a clever external contractor.

Then the serious users systematize.

They write the operating procedure.

They attach the reference files.

They define the review loop.

They standardize the workflow.

They turn repeated human judgment into reusable machine context.

That is where leverage compounds.

The agent itself is powerful. But the agent connected to persistent procedural memory is far more powerful. The organization that captures its workflows into reusable skills will accelerate. The organization that leaves everything inside scattered chats will drown in its own friction.

This is why the next competitive moat is not just model access.

Everyone will get model access.

The moat is workflow architecture.

The moat is proprietary data.

The moat is captured context.

The moat is knowing how your organization actually works and encoding that into systems agents can execute.

Not Created Equal

The authors are appropriately careful. They note that OpenAI is not a normal organization. Workers there are closer to the frontier, usage is cheap at the margin, training and informal knowledge sharing are common, and the culture is already oriented around these tools.

That caveat is correct.

It is also the entire point.

OpenAI is not just the average firm. OpenAI is the preview environment.

What happens inside OpenAI in 2026 happens inside aggressive technology companies next. Then financial firms. Then professional services. Then media. Then logistics. Then healthcare administration. Then government operations. Then education.

The delay is not capability.

The delay is organizational digestion.

Most companies are still structured around human bottlenecks. Meetings. Approvals. Hand-offs. Status updates. Permission layers. Fragile processes hidden inside people’s heads. These systems were designed for a world where labor was scarce, communication was slow, and execution required humans moving one task at a time.

That world is ending.

The paper shows that when friction drops, work reorganizes around agents very quickly. At OpenAI, Codex became dominant across functions, not just engineering. Legal, recruiting, research, product, communication, and data workflows all moved toward agentic execution.

This is the pattern every serious organization should study.

The question is not whether your employees will use AI.

The question is whether your company can restructure work fast enough to absorb agentic labor.

Human Capital Is Being Repriced

The value of raw execution is falling.

The value of judgment is rising.

The value of system design is rising.

The value of verification is rising.

The value of taste is rising.

The value of proprietary context is rising.

The value of being able to coordinate ten parallel streams of machine work without losing the plot is rising dramatically.

This is the new human capital stack.

In the old economy, workers were paid for performing tasks. In the agentic economy, workers are paid for defining objectives, designing systems, supplying context, judging outputs, and integrating results into reality.

That sounds abstract. It isn’t.

A lawyer who can supervise five legal research agents, review their work, synthesize the answer, and produce a client-ready memo will outperform the lawyer still manually searching documents.

A founder who can deploy agents across product, sales, research, support, finance, and operations will outperform a legacy team waiting for weekly meetings.

An investor who can run continuous diligence agents across filings, technical documents, market data, customer signals, and founder history will outperform the analyst still building static spreadsheets.

A writer who can operate research, editing, distribution, image, and audience-analysis agents will outperform the writer staring at a blank page.

The individual systems architect becomes a company.

The company that fails to become a system becomes obsolete.

What To Do Now

The practical mandate is straightforward.

1. Audit every workflow you touch.

2. Find the repeated tasks.

3. Find the tasks with clear inputs and outputs.

4. Find the tasks where context is scattered but knowable.

5. Find the tasks that require data collection, transformation, drafting, comparison, review, or coordination.

Then turn those workflows into agentic systems.

Do not merely “use AI more.” That is vague and weak.

Build the loop. Write the instructions. Attach the references. Capture the data. Create review checkpoints. Measure output. Improve the workflow.

Repeat until the system becomes faster than the human process it replaced.

This is how you compound.

The people who win will not be the people who occasionally ask ChatGPT for advice. The people who win will be the people who convert their daily work into repeatable agentic infrastructure.

The paper gives us the empirical map.

Agentic adoption starts unevenly.

Technical workers move first.

Non-technical workers follow.

Task complexity rises.

Concurrency rises.

Skill use rises.

Output explodes.

The human moves from operator to orchestrator.

That is path of the singularity.

The Frontier Has Already Crossed

The shift from conversational AI to agentic AI is not a product update.

It is a labor-market phase change.

The chatbot era taught humans to ask better questions. The agent era teaches humans to assign better work.

That is a much bigger transition.

OpenAI’s Codex paper shows the early shape of this world with data instead of speculation. The frontier users are not just chatting more. They are delegating larger tasks, running agents in parallel, reusing codified workflows, and reorganizing their own effort around supervision and integration.

This is how the next economy gets built.

Not by replacing every human overnight.

By turning every high-agency human into the manager of a growing machine workforce.

The leverage curve is bending upward.

The only serious response is to build systems that bend with it!

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment

Some of you will remember the piece I wrote about DeepSeek. Naval retweeted it, that got me invited onto a bunch of podcasts and speaking engagements, suddenly I became “the AI guy” and (forgive the cheese) it changed my life.

And here we are again.

GLM-5.2 just came through the side door of the AI industry and kicked the whole building sideways.

Not because it is the undisputed best model in the world.

It isn’t.

Not because it ends OpenAI, Anthropic, or Google.

It doesn’t.

But because it changes the question.

The question is no longer: “Can an open Chinese model catch up to the American frontier?”

The question is now: “How much premium can the closed frontier labs keep charging once open models are this good, this cheap, and this deployable?”

Please read that again.

Because this is the part the market has not fully digested.

GLM-5.2 is not just another benchmark-chasing model release. It is a pressure event. It lands right in the middle of the OpenAI and Anthropic IPO narrative, right as public markets are being asked to underwrite trillion-dollar AI companies on the assumption that frontier intelligence remains scarce, closed, expensive, and defensible.

Then Z.ai shows up with a model that has a 1M-token context window, serious long-horizon coding ability, MIT-licensed open weights, local deployment options, and API output pricing around $4.40 per million tokens.

That is the wrench.

Maybe not a fatal wrench. But definitely a wrench.

The whole valuation story for the big Western AI labs depends on a simple belief: that the best intelligence will remain locked behind proprietary APIs and expensive subscriptions, and that enterprises will have no choice but to pay rent to the model gods.

GLM-5.2 attacks that belief directly.

It says: what if the frontier is not a castle?

What if it is a floodplain?

Z.ai’s own positioning is very clear. GLM-5.2 is built for long-horizon tasks. Not cute chatbot tasks. Not “write me a limerick about SaaS pricing” tasks. Real agentic work: codebases, multi-step engineering, long debugging loops, research reproduction, tool use, sustained execution.

That matters because coding agents are the first place where LLMs stop being toys and start becoming labor.

And GLM-5.2 is aimed directly at that market.

The model extends context from 200K to 1M tokens. But the more important claim is not “it can fit a million tokens.” Lots of people can slap a huge context window on a model and watch performance decay into soup.

The claim is that this is a usable million tokens.

That means a model can absorb the shape of a real software project: the architecture, the file boundaries, the API contracts, the weird historical decisions, the dependency constraints, the tests, the style, the hidden landmines, the things every senior engineer knows after three months and every AI agent usually forgets after twenty minutes.

That is not just “more memory.”

That is continuity.

And continuity is where software agents become dangerous.

A bad coding model can generate snippets.

A good coding model can solve issues.

A great coding model can hold a system in its head long enough to make coherent changes across time.

That is the ballgame.

This is why the benchmarks matter, even if we should never worship them. Z.ai claims GLM-5.2 is the top open-source model across several long-horizon coding benchmarks. On FrontierSWE, it trails Claude Opus 4.8 by about 1% while edging out GPT-5.5. On PostTrainBench, it ranks behind Opus 4.8 but ahead of GPT-5.5 and Opus 4.7. On Terminal-Bench 2.1, it posts an 81.0 versus GLM-5.1’s 63.5.

Again, benchmarks are not reality.

But they are not nothing.

They are smoke. And in AI, smoke usually means the fire is already spreading.

The real shock is the price-performance curve.

A model in the general neighborhood of closed frontier coding models, available for roughly $4.40 per million output tokens, with cached input pricing far below standard input, and open weights you can run yourself if you have the hardware.

This is not just cheap. This is strategically cheap.

This is “why exactly are we paying the frontier tax?” cheap.

This is the part that gets uncomfortable for OpenAI and Anthropic.

Their IPO stories need scarcity. They need the market to believe that intelligence is capital-intensive, yes, but also defensible. They need investors to believe the billions poured into compute create a moat that converts into pricing power. They need “we have the best model” to become “we have the best business.”

GLM-5.2 does not destroy that argument.

But it damages it.

Because every time an open model gets close enough, the premium model companies have to explain why “better” is worth 5x, 10x, or 20x the price.

Sometimes it will be.

For mission-critical tasks, enterprises will still pay for reliability, support, indemnity, governance, safety, integrations, procurement comfort, data controls, and brand trust.

But not always.

And “not always” is where valuations go to get humbled.

If you are a startup burning millions per year on inference, you do not need GLM-5.2 to beat Claude or GPT on every dimension. You need it to be good enough on enough workloads that your unit economics stop bleeding.

If you are an enterprise with sensitive code, you do not need the model to be the universal oracle. You need an option you can host, inspect, constrain, fine-tune, and govern.

If you are a developer using agents all day, you do not need religious loyalty to a logo. You need the model that gets the work done without turning your token bill into a second payroll.

That is the opening.

And it is wide.

This also quietly ruins the vibe around Gemini 3.5 Pro.

Google had its moment lined up. Gemini 3.5 Flash was pitched as frontier intelligence with action, strong agentic coding, long-horizon execution, and speed. Google said 3.5 Pro was coming next. The narrative was supposed to be: Google is back at the frontier, and Gemini is now the model family for agents.

Then GLM-5.2 appears with open weights, 1M context, serious coding benchmarks, and pricing that makes every closed model launch feel a little heavy.

That does not mean Gemini 3.5 Pro will be bad.

It may be excellent. In fact, I would be surprised if wasn’t. Google has absurd infrastructure, elite research talent, distribution everywhere, and a giant surface area across Search, Workspace, Android, Cloud, and everything else. Underestimate Google at your own risk.

But model launches are partly about narrative.

And GLM-5.2 stole oxygen.

It made “frontier agentic model” feel less like a sacred object and more like a category.

That is a huge psychological shift.

The Z.ai technical story is also not trivial. They are not just saying “we trained a big model and it got good.” They are talking about architectural efficiency. IndexShare reuses the same indexer across sparse attention layers and reduces long-context computational cost. They improved multi-token prediction for speculative decoding and increased acceptance length. They are explicitly optimizing the machinery required to make 1M-token work practical.

This is the deeper theme.

The frontier is not just about scale anymore.

It is about efficiency of intelligence.

Who can get the most capability per dollar, per watt, per GPU, per unit of latency, per developer hour?

This is where China has been terrifyingly strong.

DeepSeek showed the world that reasoning could be produced more efficiently than people assumed. GLM-5.2 continues the same pattern in the coding-agent world.

The uncomfortable American lesson is that constraints create invention.

If you have unlimited capital, unlimited GPU access, unlimited pricing power, and a customer base trained to pay premium rates, you can become lazy in ways that are invisible until someone hungrier ships around you.

China’s labs have been forced to optimize. Sanctions, compute limits, and market pressure created a different evolutionary environment. The result is not always the best model in absolute terms. But it is often the best model for the price.

And markets love price-performance.

They always have.

China = Copiers or Innovators?

Now, let’s take the countercase seriously.

Because there is one.

The harshest version goes like this: Chinese models are mostly distills of Western frontier models. They are downstream beneficiaries of OpenAI, Anthropic, and Google doing the expensive first-principles research. They are not creating the frontier at all. In fact they are just compressing it. Progress would stall if these labs did not have armies of VPN accounts hitting Western APIs from non-Chinese IPs, extracting behavior, generating synthetic data, and laundering proprietary intelligence into “open” models.

That argument is not crazy.

In fact, it is almost certainly true, in part.

Distillation is everywhere. Synthetic data is everywhere. Model outputs train other models. The whole field is eating itself recursively.

And yes, the American frontier labs may be doing the most expensive trailblazing. They discover the capability, absorb the failures, pay the compute tax, build the scaffolding, run the safety work, create the product category, and then others imitate the behavior at lower cost.

That is a real concern.

If every cheap open model is downstream of closed frontier labs, then the open ecosystem may be more dependent on the closed labs than it wants to admit.

The “Chinese models are just distills” critique is really an argument about originality, sustainability, and fairness.

Originality: did the model learn fundamental capability from its own training process, or did it learn to mimic the behavior of models that were much more expensive to create?

Sustainability: if Western labs stopped advancing, would these models keep improving or plateau?

Fairness: is this competitive innovation, or is it industrial-scale free-riding?

Those questions matter.

But here is the problem for the countercase: customers do not pay for metaphysics. They pay for results.

If a model solves the task, integrates into the workflow, runs locally, and costs a fraction of the alternative, the buyer does not usually care whether the capability came from pristine original research, clever distillation, open papers, synthetic data, reinforcement learning, or some blurry mixture of all of the above.

The market asks: does it work?

Then: how much does it cost?

Then: can I trust it?

The distillation argument may be morally and strategically important. It may influence export controls, lawsuits, procurement rules, and national security policy. It may absolutely shape how governments respond.

But it does not erase the competitive effect.

A cheaper substitute does not become less disruptive because its origin story is messy.

If anything, that makes the situation more destabilizing.

Because the American labs may be funding the frontier research that commoditizes their own products.

Loops and Curves

Spend $100 billion pushing the frontier.

Watch someone else learn from the exhaust.

Compete against their cheaper model.

Lower your prices.

Raise more money.

Repeat.

That is a brutal loop.

This is why GLM-5.2 matters beyond the model itself.

It points toward a world where frontier capability diffuses faster than frontier economics can stabilize.

The capability curve goes up. The cost curve goes down.

The moat duration shrinks.

That is incredible for builders.. and terrifying for anyone underwriting monopoly pricing.

I still think closed labs have advantages.

They will have the best multimodal systems. They will own premium consumer products. They will have enterprise trust. They will have the deepest research benches. They will build better safety layers, better tool ecosystems, better integrations, better memory systems, better orchestration, better evals, and better support.

Also, open models are not magic. Local deployment is only “free” after you buy or rent serious hardware. A 1M-token MoE model is not casually running. Operationalizing open weights takes engineering skill. Serving long context at scale is hard. Security is hard. Reliability is hard. Fine-tuning can make models worse. Quantization can change behavior. Enterprise support matters.

So no, this is not “OpenAI is dead.”

That is lazy.

The real story is subtler and much more important.

OpenAI, Anthropic, and Google are not being killed by GLM-5.2.

They are being repriced by it.

Their products may still be better.

But their scarcity premium is under attack.

And once scarcity premium compresses, everything changes: margins, growth assumptions, IPO multiples, enterprise negotiations, product bundling, compute strategy, and the speed at which model intelligence becomes a commodity input.

That last phrase is the one to watch.

Commodity intelligence.

Not dumb intelligence.

Not weak intelligence.

Commodity intelligence that is extremely capable, widely available, locally runnable, and cheap enough to disappear into every workflow.

That is a different civilization.

Because when intelligence gets cheap, people stop rationing it.

They put it everywhere.

They run agents against every repo, every spreadsheet, every compliance process, every sales motion, every research question, every personal goal, every operational bottleneck.

The world becomes saturated with cognitive labor.

This is what I meant when I wrote about DeepSeek. The important thing was not just “China made a good model.” The important thing was that the recipe for intelligence production was changing.

GLM-5.2 is another turn of that same screw.

DeepSeek proved reasoning could emerge with shocking efficiency.

GLM-5.2 suggests long-horizon agentic work is becoming open, cheap, and deployable.

That is a massive shift.

The future does not belong only to whoever has the biggest model.

It belongs to whoever can turn intelligence into leverage at the lowest sustainable cost.

And that is why this release feels so important.

We are watching the AI industry move from priesthood to power tool.

From “come worship at our API” to “download the weights and build.”

From subscription intelligence to ambient intelligence.

From closed scarcity to competitive abundance.

There will be lawsuits.

There will be export controls.

There will be safety panics.

There will be national security arguments, some very real and some very convenient.

There will be benchmark fights, distillation accusations, pricing wars, and a lot of people pretending they saw all of this coming.

But the direction is getting clear.

The frontier is leaking.

And once intelligence leaks, it does not go back into the bottle.

Thanks for reading Life in the Singularity.

This post is public, so feel free to share it with someone still underwriting closed-model moats like it’s 2024.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

Share Life in the Singularity

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏

Leave a comment