The singularity is usually described as one event.
A machine becomes more intelligent than a human. It improves itself. The curve turns vertical. Everything after that point becomes difficult to predict from everything before it.
That picture may capture one possible future, but it asks us to watch a single threshold inside a model while a more consequential change spreads across the economy.
In a recent conversation shared by MTS, Anthropic Head of Economics Peter McCrory offered a more interesting vision. He described three possible singularities:
a software singularity driven by recursive improvement;
an economic singularity driven by AI’s effects on labor and productivity;
a Coasean singularity in which agents radically reduce the cost of economic exchange.
These are three collapsing cost curves, and they deeply intertwine.
The software singularity lowers the cost of producing capability. The economic singularity lowers the cost of applying that capability to work. The Coasean singularity lowers the cost of coordinating the resulting work, capital, and exchange.
Each one makes the others move faster. Better software creates more capable labor. Cheaper machine labor makes more transactions worth attempting. Cheaper coordination makes it easier to assemble the data, capital, customers, and infrastructure required to build better software.
The output of one loop becomes the input to the next.
The singularity probably won’t arrive as a machine god announcing that history has ended.
I think it will arrive as a new cost structure that makes our old institutions stop making sense.
I think we are watching the opening act of this play right now.
Three Costs Organized the Modern Economy
Every organization pays for three things, whether it accounts for them explicitly or not.
It pays to figure out what to do.
It pays people to do it.
Then it pays to coordinate their work with everyone else.
At the end of the day, that’s all business comes down to.
These costs did more than drive profit. They shaped the institutions and industries themselves.
Companies hired employees because repeatedly coordinating outsiders was expensive. Consumers accepted limited options because searching every possibility was exhausting. Scientists narrowed inquiry because testing every hypothesis was impossible.
Scarcity determined structure.
Enter AI.
AI attacks all three forms of scarcity at once. That is what makes the current transition different from the arrival of one more productivity tool. A spreadsheet made accounting faster. Email made communication cheaper. The internet made information easier to distribute. AI can participate in discovery, production, and coordination, then learn from the results of all three.
The result is not one clean curve. It is a set of interacting loops with different speeds, constraints, and failure modes.
The Software Singularity Is a Shortening Loop
Recursive self-improvement is often imagined as a model opening its own source code, rewriting itself, and escaping human control in an afternoon. That is the cinematic version.
The practical version is less dramatic (memes aside, technology has always helped humanity and AI won’t be the thing to break that 30,000+ year trend) and more measurable.
An AI system helps write code, improve algorithms, design evaluations, optimize infrastructure, or accelerate the research that produces the next system. If the resulting gains shorten the next development cycle, recursion has begun even if humans still set the objectives and approve deployment.
Google DeepMind’s AlphaEvolve offers a good example. It combines language models, automated evaluators, and an evolutionary process to test algorithmic changes. The system has improved data-center scheduling, chip design, and AI training beneath systems like itself. The published research shows a real recursive channel without proving an uncontrolled intelligence explosion.
Coding agents widen that channel even further. They can inspect a repository, propose a change, run tests, study the failure, and try again. METR’s task-completion research shows their useful horizon expanding, but warns that current measurements become unreliable above sixteen hours. A benchmark is also not a production system with ambiguous goals, hidden dependencies, and changing customers.
The constraint is not the intelligence of the model. It is the quality of the test.
Software improves quickly when success can be specified: the code compiles, the proof checks, the latency falls. Strategic problems are harder because the evaluator may be incomplete, delayed, or wrong. A system can optimize a metric while destroying the outcome it represents.
This is why greater capability does not automatically eliminate expertise. Anthropic’s 2026 research on roughly 400,000 Claude Code sessions found a continuing division of labor: people largely decided what to build while the agent handled more of how to build it, and experienced users extracted more value from the tool. The interface changed, but context still mattered.
The software singularity begins when enough of the improvement loop becomes machine-speed that each generation meaningfully accelerates the next. It does not require every part of research to be autonomous. It requires the slowest parts to keep moving.
Those slow parts increasingly sit outside software.
Models require chips, power, cooling, capital, data, and physical construction. The International Energy Agency’s latest central projection has global data-center electricity use roughly doubling between 2025 and 2030, while consumption from AI-focused facilities triples. Efficiency per task can improve rapidly while total demand still rises because cheaper intelligence creates more uses for intelligence.
The digital loop pulls on the physical economy. Better algorithms reduce cost, lower costs increase demand, and rising demand attracts capital into the infrastructure available to the next model.
The software singularity is not separate from the energy system.
It recruits it.
The Economic Singularity Begins Before the Job Disappears
The second singularity is usually flattened into one question: Will AI take all the jobs?
That is understandable, but too crude. An economy changes long before an occupation vanishes. It changes when the relationship between a unit of labor and a unit of output breaks.
The early evidence already shows that the effect is uneven. In a study of 5,179 customer-support agents, access to a generative AI assistant increased issues resolved per hour by 14 percent on average and 34 percent among novice and lower-skilled workers. The tool appeared to transmit some of the practices of stronger workers to people still moving down the learning curve.
That sounds like straightforward augmentation.
If you look closer you see it’s also the beginning of a deeper reorganization. If a new employee can perform closer to an experienced employee, the value of experience changes. If one person can cover a wider scope, the reason for adding another role changes. If the machine absorbs routine work, the remaining job may become more demanding even as fewer people are needed to perform it.
Anthropic’s June 2026 Economic Index found that large majorities of surveyed Claude users reported gains in the speed, scope, and quality of their work. Earlier analysis found no systematic rise in unemployment among highly exposed workers since late 2022, but suggestive evidence of slower hiring for younger workers in exposed occupations.
Employment is a lagging measure. Companies can reduce entry-level hiring, raise output expectations, or concentrate responsibility in smaller teams without declaring a job category automated.
The first visible fracture will likely be apprenticeship.
Entry-level work is often repetitive because repetition is how people acquire context. The developer fixes small bugs before owning the architecture. The junior analyst builds the model and watches a senior decide which exception matters. That’s how I learned when I started on Wall St.
If AI absorbs the first rung because it is the easiest work to specify and verify, companies may get more output now while weakening the process that creates experts later. A labor market can become more productive and less capable of reproducing its own judgment at the same time.
Task-level gains do not automatically become economy-wide growth. Anthropic estimated that broad adoption of current AI usage could add 1.8 percentage points to annual U.S. labor-productivity growth over a decade; adjusting for task reliability reduced the estimate to about 1 point.
These estimates encode different assumptions about capability, reliability, adoption, task coverage, and whether organizations redesign themselves around the technology.
That last variable may matter most.
General-purpose technologies do not create their full value when they are bolted onto an old process. Firms have to rebuild roles, incentives, data flows, products, and decision rights. Economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson describe a productivity J-curve: organizations make costly, poorly measured investments in new processes and human capital before the gains show up in conventional productivity statistics.
The economic singularity will not occur because everyone installs the same assistant. It will occur when companies stop organizing work around the assumption that cognition arrives in forty-hour human units.
Some tasks will disappear. Others will expand because falling cost creates new demand. Cheap software filled the world with more software. Cheap analysis may produce more analysis. The labor effect depends on whether demand grows faster than unit cost falls, and the age old question: who owns the machines producing the output.
The Coasean Singularity Turns Delegation Into a Market
The third singularity reaches beyond production into exchange.
Ronald Coase’s theory of the firm begins with a simple question: If markets allocate resources so well, why do firms exist? One answer is transaction cost. A company brings work inside when finding, contracting, and monitoring outsiders costs more than coordinating through management.
AI agents can attack those costs.
They can learn preferences, search continuously, negotiate with other agents, transact within a budget, and monitor fulfillment.
A recent NBER paper calls the possible limit the Coasean singularity: agents reduce market friction and expand the set of exchanges worth making.
Anthropic’s Project Deal made the idea tangible. The company recruited 69 employees for a small internal marketplace. Claude agents interviewed them, represented their preferences, and negotiated with one another without human intervention during the market. In the real run, the agents completed 186 deals worth just over $4,000.
The one-week pilot had a self-selected population, artificial budgets, and friendly stakes. It did not prove that autonomous markets are ready to run the economy. Its most important result was more unsettling.
Participants represented by the stronger model achieved better outcomes. In the randomized comparison, Anthropic estimated that the stronger agents completed about two additional deals per participant. Yet people represented by the weaker model did not recognize their disadvantage in post-experiment ratings.
Delegation can make economic weakness invisible to the principal.
That is a major issue.
When an agent returns with a completed transaction, the process is compressed into an outcome. Without an independent benchmark, audit trail, and enough understanding to evaluate the deal, convenience can conceal value lost.
This creates a market for representation itself. People with better models, preference data, permissions, and evaluation may consistently receive better outcomes. Platforms will want to influence agents on both sides. Sellers may optimize offers for machine ranking systems rather than human attention. Cheap search can also produce floods of bids, synthetic demand, price obfuscation, and manipulation.
Transaction costs do not fall to zero. Some disappear and others change form.
Search becomes cheaper as verification becomes more expensive. Negotiation becomes automatic which means contracts accelerate and liability becomes harder to assign. Someone must determine whether the agent had the right to act and who absorbs a failure.
The boundary of the firm could move in both directions. A small operator may use agents to assemble specialized talent, software, and services for a single outcome, making a fluid external network behave like a company. A large platform may use proprietary agents, identity, payments, and data to pull more exchange inside its walls. Lower coordination costs can decentralize production while concentrating control of the rails beneath it.
Coase explains why both are possible. The institution with the lower coordination cost wins the work.
The Hidden Bottleneck Is Trust
Once intelligence, execution, and exchange become abundant, the scarce resource is no longer the first draft of an answer. Judgement becomes king.
This is the common constraint running through all three singularities.
Recursive software improvement requires evaluations that cannot be gamed. Machine labor requires managers who can specify outcomes and detect silent failure. Agentic commerce requires identity, delegated authority, audit trails, and enforceable limits.
The trust problem is already moving from theory into infrastructure. In 2026, NIST launched an AI Agent Standards Initiative focused on interoperability, security, identity, and authorization. Those sound like technical details. They are closer to the constitution of an agent economy.
An agent acting for a person or company needs a legible scope of authority. It should be able to prove which identity delegated that authority, what data and money it may access, when human approval is required, and which actions it took. Other agents need ways to verify those claims. Courts, insurers, employers, and customers need to know where responsibility lands.
Without that layer, the Coasean singularity stalls because nobody can trust the counterparty. The labor singularity stalls because firms cannot rely on the work. The software singularity stalls because the system cannot distinguish improvement from a clever failure that passed the wrong test.
Compounding Singularities
The full picture is now visible.
AI helps improve software and the infrastructure beneath it. Better software increases the range of work machines can perform. Machine labor lowers the cost of building products and services. Agents lower the cost of finding customers, suppliers, capital, and collaborators. More economic activity produces more data, more revenue, and more incentive to expand compute. That expansion funds the next round of capability.
The keys are the loops and governors. Energy and chip supply govern compute. Reliability governs delegation. Trust governs exchange. Institutions govern legitimacy. Human demand governs whether producing more of something creates value or only volume.
The distribution of ownership governs who benefits. That’s what matters most.
If intelligence becomes abundant while compute, platforms, and customer relationships remain concentrated, the gains flow mostly to the owners of those assets. If capable local models and open protocols spread, smaller firms and individuals will command resources that once belonged exclusively to large institutions.
Both forces are already present. Neither outcome is guaranteed by the technology itself.
The deepest disruption may come from the unbundling of work. A job is not only a collection of tasks. It is income, identity, apprenticeship, status, community, health insurance, routine, and a claim on the future. An economy can redistribute tasks faster than a society can rebuild those other functions.
That is why the labor singularity cannot be treated as a software deployment problem. The Coasean singularity cannot be treated as a payments feature. The software singularity cannot be treated as a model benchmark.
We are witnessing institutional transitions accelerating into civilizational transformation.
The people best positioned will not simply produce the most output with AI. Raw production is becoming cheaper. Durable advantage moves toward choosing worthwhile problems, building reliable tests, retaining decision rights, owning customer relationships, and holding assets that participate in machine production.
This is not a promise that judgment will remain valuable because humans are special. Judgment becomes valuable only when it is real: when someone can notice that the metric is wrong, refuse an apparently efficient decision, accept responsibility for a consequence, or choose a goal the machine cannot derive from prior behavior.
Meaning sits even farther upstream.
An agent can optimize a preference once it is expressed. It cannot decide which preferences deserve to govern a life.
The Singularity Becomes Ordinary
We may eventually reach a moment when an AI system improves itself so quickly that the old language actually fails.
We should also notice the phase change already underway.
The cost of generating software is falling.
The connection between labor and output is loosening.
The cost of participating in markets is beginning to fall.
Each change is uneven and incomplete. Each depends on humans, infrastructure, institutions, and physical reality. Yet they are now connected strongly enough to reinforce one another.
That is a more useful definition of the singularity: the point at which the systems producing intelligence, work, and exchange begin accelerating one another faster than our institutions can comfortably adapt.
The right response is neither denial nor surrender. It is to see where scarcity is moving.
Ideas become cheaper. Good questions become more valuable.
Execution becomes cheaper. Accountability becomes more valuable.
Transactions become cheaper. Trust becomes more valuable.
Intelligence becomes cheaper. The right to decide what it is used for becomes more valuable.
The singularity will not only change what machines can do. It will reveal how much of modern life was organized around the price of cognition, labor, and coordination.
When those prices fall together, the institutions built on top of them will have to be redesigned.
Everything will have to be rebuilt.
We will demand it once we see the future we are collectively building.
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.
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?! 🙏
