Big Announcement
We are switching to 1 free post with at least 5 premium posts each month.
The pace of technology is accelerating, and the alpha from these research paper breakdowns has prompted retweets from Naval and other big exposure online (podcasts with Robert Breedlove, speaking at national events, etc..). All of this combined has lead to me meeting smart, passionate and very talented people and ultimately, to me joining several technology companies including a few stealth ventures that will be announced at the right time.
All that said, I need to post less often but with much, much deeper quality in order to bring powerful and useful information to my business partners.
That’s why this post (and most posts going forward) will be premium for paying subscribers only. The plan is to use NotebookLM, Gemini 2.5 Pro Deep Research and other AI tools to deeply interrogate and integrate the most recent research papers in machine learning and related domains. Our posts will aim for a 20-minute reading time but sometimes we’ll need to go for 45-minutes or more to get a full picture.
I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration across science and technology. I’m a trusted tester. I’ve been in ML since 2009. Currently testing Google’s Diffusion model for them. Doing even more that I can’t talk about!
I started Life in the Singularity in May 2023 to track all the changes in AI/ML, robotics, blockchain, quantum computing and the rest of the technologies accelerating humanity forward into the future. Our brilliant audience includes Fortune-500 board members, engineers and executives, builders including truly talented technologists, angel investors + VCs and other pockets of the capital community, and most importantly, thousands of people who want to use technology to maximize the utility in their lives.
P.s. WealthSystems.ai will remain largely free because the mission of educating and exciting people about taking wealth into their own hands by building wealth systems is too important to paywall + evangelizing is literally the mission. Tough to do that behind walls!
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!
Onto the research…
Imagine a world where the most complex scientific puzzles and intricate engineering challenges could be solved not just by human ingenuity alone, but by a powerful AI collaborator capable of discovering entirely new solutions?
This is the promise of AlphaEvolve, a groundbreaking coding agent developed by Google DeepMind. The primary challenge AlphaEvolve tackles is the automation of discovery for highly complex problems, from cracking open scientific enigmas to optimizing critical computational infrastructure.
At its heart, AlphaEvolve employs an ingenious evolutionary approach. It orchestrates a team of Large Language Models to iteratively refine algorithms.
Think of it as a highly accelerated, AI-driven version of natural selection, but for code.
AlphaEvolve starts with an initial algorithm (which can be very basic) and continuously tasks LLMs with proposing modifications to improve it. These new algorithmic "species" are then rigorously tested by one or more automated "evaluators" that provide feedback on their performance. This continuous loop of generation, testing, and selection allows AlphaEvolve to progressively discover better and more efficient solutions.
Once I explain how this works later in the piece you will truly be amazed at the pace of progress. AlphaEvolve has demonstrated its prowess by discovering novel, provably correct algorithms that surpass state-of-the-art solutions in mathematics and computer science. For instance, it found a more efficient method for multiplying 4x4 complex-valued matrices, the first such improvement in 56 years, outperforming Strassen's renowned algorithm in this specific setting. Beyond theoretical breakthroughs, AlphaEvolve has delivered tangible optimizations for Google's own large-scale systems, developing more efficient data center scheduling algorithms, simplifying hardware accelerator circuit designs, and even accelerating the training of the LLM that powers AlphaEvolve itself.
The most significant takeaway is that AlphaEvolve represents a major leap in using AI for discovery. It’s not just about making existing things faster; it’s about the AI autonomously finding new ways of doing things, potentially leading to scientific and practical discoveries that were previously beyond our reach. It’s a powerful demonstration of how AI can become a true partner in innovation.
Faster and Faster We Go…
AlphaEvolve doesn't just incrementally improve upon past methods; it fundamentally expands the horizon of what automated discovery systems can achieve, particularly when compared to its predecessors and related approaches.
Traditional evolutionary programming, while powerful, often relied on human-designed "mutation operators”.
Think of these as the rules for how an algorithm could change. Designing effective operators for complex problems is incredibly difficult and can limit the creativity of the evolutionary process. More recent efforts to combine LLMs with evolution, such as FunSearch (also from Google DeepMind), were significant steps forward but had notable limitations.
FunSearch, for example, primarily focused on evolving a single Python function, typically consisting of only 10-20 lines of code. It was geared towards problems with very fast evaluation times, used smaller LLMs, optimized for a single metric, and had minimal context provided to the LLM. While groundbreaking for mathematical discovery, its architecture wasn't designed for evolving large, multi-component algorithms or complex engineering systems. Other LLM-based code generation or optimization techniques often lacked the sustained, feedback-driven evolutionary loop that allows for deep, iterative improvement.
Instead of just a single function, AlphaEvolve can evolve entire code files, potentially spanning hundreds of lines and multiple interacting components, across various programming languages. This is crucial for tackling real-world engineering problems where solutions are rarely confined to a small snippet of code. Users can simply mark blocks of code within an existing codebase for AlphaEvolve to target and improve.
AlphaEvolve harnesses the power of cutting-edge LLMs like Gemini 2.0 Pro and Flash. It feeds these models rich contextual information, including previously successful (and unsuccessful) code variations, problem descriptions, relevant literature (even PDFs), and detailed feedback from evaluations. It can even evolve the prompts given to the LLMs ("meta prompt evolution") to improve the quality of generated code over time.
AlphaEvolve uses an ensemble of LLMs. Think of a fleet of ships vs a single vessel. Faster models like Gemini 2.0 Flash for high-throughput idea generation, and more powerful models like Gemini 2.0 Pro for occasional, higher-quality suggestions that can lead to breakthroughs. It also requests changes in a structured "diff" format, allowing for precise, targeted modifications to large codebases. The evaluation process is also enhanced, supporting cascades (testing on progressively harder cases to quickly prune bad ideas) and even using LLMs to provide feedback on qualities like code simplicity.
Unlike systems that optimize for a single score, AlphaEvolve can simultaneously optimize for multiple objectives. Please, read that again. This is vital for real-world problems where, for example, speed, memory usage, and energy consumption might all be important. Interestingly, optimizing for multiple metrics can even lead to better results on a single target metric by encouraging more diverse solutions.
The system maintains a "program database" of evolving solutions, implementing an algorithm inspired by MAP-Elites and island-based models to balance improving the best solutions (exploitation) with maintaining diversity to explore new avenues (exploration).
So, are we beyond AGI already?
The core assumption underpinning AlphaEvolve is that the problem at hand must allow for automated evaluation. In other words, there needs to be a way for a computer program to automatically assess how good a proposed solution is. This is readily achievable in domains like mathematics (checking if a proof is valid or a construction meets certain criteria) and software optimization (measuring speed or resource usage). While FunSearch shared this reliance on automated evaluators, AlphaEvolve applies it to problems of far greater scale and complexity.
AlphaEvolve contributes significantly to several rapidly advancing fields. It's a landmark in AI for scientific discovery, demonstrating that AI can not only assist humans but also generate novel, SOTA-beating insights. It pushes the boundaries of code superoptimization, showing how LLMs can iteratively refine complex code far beyond simple bug-fixing. More broadly, it’s a powerful example of agent-based AI, where an autonomous system orchestrates other AI models and tools to achieve complex goals.
It fuels the ongoing discussion about the potential for AI to accelerate the pace of innovation across science and engineering, transforming how we approach and solve the world's most challenging problems.
Imagine you have a grand challenge: to design the fastest, most fuel-efficient racing car ever.
The User is like the team owner. They define the goal ("win the championship"), provide the initial car blueprints (the starting algorithm, maybe just a basic chassis), and set up the racetrack with sophisticated timing and telemetry systems (the automated evaluator to measure speed, fuel use, etc.).
Meanwhile, the LLMs are a diverse team of brilliant but inexperienced automotive engineers. Each has access to vast knowledge of car design and physics.
AlphaEvolve is the seasoned Chief Engineer and Team Manager.
The Chief Engineer (AlphaEvolve) takes the initial blueprints and shows them to small groups of engineers (LLMs), along with data from past test runs and design ideas that seemed promising. It asks them: "How can we make this specific part of the car (a block of code) better? Or how can we combine ideas from these different successful designs?"
The engineers (LLMs) sketch out modifications—a new aerodynamic fin, a tweaked engine component, a different gear ratio (code changes, often as "diffs" or precise edits).
Each modified car design is then immediately built (the code is updated) and rigorously tested on the track (run through the evaluator). The performance data (scores on speed, efficiency) is fed back.
The Chief Engineer (AlphaEvolve) keeps a detailed logbook of every design tried and its performance (the program database). It strategically selects the most promising designs. These are not just the fastest overall, but also those that excel in specific areas (great cornering, amazing fuel economy) to inspire the next round of innovation.
This cycle repeats relentlessly.
Forever. At faster and faster speeds as our hardware and software improve. See how massive this is getting?
Some engineering ideas are dead ends, but others lead to small gains, and occasionally, a radical new design element emerges that gives a huge leap in performance. Over time, through this AI-driven evolutionary process, the team doesn't just tweak the initial car; they might discover entirely new engine types or aerodynamic principles, resulting in a championship-winning vehicle far beyond what any single engineer could have initially conceived. That's AlphaEvolve: constantly learning, evolving, and discovering.
Think of what this means for the future, when AI systems are able to continuously improve their logic, integration with tools and the internet, UI… everything?