AlphaEvolve: Using LLMs to solve Scientific and Engineering Challenges | AlphaEvolve explained

Key Takeaways

[Music] What if you could take a rough idea for solving a problem and instead of solving it yourself, you just let an AI evolve the solution over time? That's exactly what Google Deep Minds Alpha Evolve does. In this video, we're diving into the system that blends code generation, evolutionary strategies, and LLM to discover brand new algorithms and optimize realworld infrastructure. Grab a cup of coffee because in this AI coffee break, we'll explain how Alpha Evolve works. Alpha Evolve starts with a basic code base and improves it automatically. It uses large language models to suggest small edits and tests how well those edits perform using userprovided evaluation code. The best candidates are used to further do edits on. Over time, Alpha Evolve evolves code that solves problems far beyond what a human or even a single LLM could create from scratch. In more detail, at the heart of Alpha Evolve is an evolutionary loop. The user provides three things. First, a starting program. This can be very basic or even wrong. important is that this program should be structured well enough to run even if it's just a skeleton with placeholder logic and functions that return constants. Second, the user must provide the evaluation code, a function that gives numerical scores to how well a program performs. Third, the user provides markers to flag which parts of the code Ali Vuloveve is allowed to change. And that was the human work. Then the user can lean back and make the GPUs go br to find some cool solution for the problem they just described. Alpha Evolve starts by using the provided user code and wrap prompts around it for an ensemble of LLMs like Gemini Flash for fast diverse ideas and Gemini Pro for more thoughtful highquality suggestions. Each LLM proposes diffs small edits to the code. These edited programs are run and scored with the user provided evaluation function and the best performing programs are added to a growing database of ideas. Then new prompts are built from those winners and the loop continues with generating a prompt now with the best prior program in the prompt and so on. But wait, I said that the best performing programs appear in the prompts for building the next program. But to choose the best program is not always ideal. The authors complicate things with fancy strategies like map elites or islandbased population models. Because here's the problem. If you only chase the top scoring last solutions, you risk getting stuck in a local optimum where an evolutionary niche turns out to be a dead end or not the most optimal and you'd miss out on very different and potentially better ones simply because the system didn't explore widely enough. That's where MAP Elites comes in. Instead of starting from a single solution and gradually improve it, you begin with a wide range of random solutions and over time you keep the best performing ones across different behavior niches. These are your elites. New solutions are then generated by building on a random subset of such diverse elites. So evolution happens from many strong and different starting points which helps cover a broader landscape of possibilities. Islandbased population methods take a different angle. Here the idea is to keep evolution isolated for a while to allow each subpopul or island to specialize without interference. Each island explores its own path in parallel. Occasionally one island might receive a solution from another, like borrowing an idea from a distant lab. This cross-pollination introduces fresh perspectives and can help a stuck island escape local optima by importing innovations discovered elsewhere. Together, these strategies help Alpha Evolve balance exploitation, so improving on good ideas with exploration, trying new ones. And that's critical when you're searching for genuinely new algorithms because breakthroughs often come from the unexpected. So all of this put together gives a process which doesn't just blindly mutate code. Alpha Evolve remembers what worked in the past and builds smarter prompts that inspire better solutions. It can even optimize for multiple goals at once like speed and accuracy or evolve entire code bases, not just isolated functions. Results speak for themselves. In one of its most striking achievements, Alpha Evolve discovered a new way to multiply 4* 4 complex valued matrices using only 48 scalar multiplications. This beats the famous Strassen algorithm, a method that hadn't been improved for over half a century. The authors also took the effort to formulate in code over 50 open mathematical problems and let Alpha Evolve tackle them. In 75% of cases, it matched the best known solutions. But in 20% of them, it went beyond state-of-the-art, solving problems like the minimum overlap and kissing numbers in 11 dimensions. And don't ask me what happens to the rest 5%. This paper doesn't say it. And in the engineering world, Alpha Evolved helped optimize Google's own infrastructure. It improved scheduling in data centers, made TPU circuit design more efficient, and also interesting is that it could speed up the LLM underlying Alpha Evolve. It sped up attention in transformer models underlying LLMs and even cut training time for Gemini itself by 1%. Which does not sound like much because let's imagine that Gemini trained over 3 months. Then that would almost save 22 hours. But here's how I interpret these results. The problems tackled by Alpha Evolve were already highly optimized to begin with if Alpha Fold couldn't surpass existing solutions by much. But the other way to look at it, and this is the exciting part, it's that there's still so much room for LLMs and the Alpha Evolve system to improve because who knows how much code must be touched to speed up Gemini's training beyond 1%. And that's what makes this exciting. It's not just about writing code. It's about discovering new knowledge. But before we start worrying that AI will steal all the science jobs, there's an important catch. A five evolve only works when results can be automatically measured. If a task relies on human judgment, nuanced interpretation, or real world experimentation, this system cannot help, at least not yet. And let's not forget, Alpha Evolve doesn't find or formulate problems on its own. It still needs a human to define the challenge, highlight what's interesting, design the code scaffolding, and write the evaluation function. And in that sense, the scientist isn't being replaced, they're being augmented. Now, if Alpha Evolve sounds familiar, it's because it builds on a previous project from Deep Mind called Fun Search, which we talked about in a past video. That system also used LLM and evolutionary search to solve math problems, but it was much more limited. The idea and logic for fun search is the same one that you see here for Alpha Evolve. But fun search could only evolve a single Python function, typically around 10 to 20 lines of code. Alpha Evolve, by contrast, can edit entire code bases across any programming language. Fun search relied on small language models and had to generate millions of samples to get results. Alpha Evolve is far more efficient. It uses not small but state-of-the-art language models and only needs a thousand of samples and not millions to find solutions. Fun search could only optimize one metric at a time, Alpha Evolve can juggle multiple objectives like accuracy, speed, and simplicity all at once. And while Fun Search needed very fast evaluations, things seconds per program, Alpha Evolve runs in parallel on accelerators and can afford hours of compute per evaluation if needed. In short, Alpha Evolve continues what punarch started and scales it up massively in power, scope, and generality. And that was it for today's AI coffee break. Thanks for watching. Please give it a like, leave your thoughts in the comments, and don't forget to subscribe for our upcoming deep dives into AI research. Also, check out our merch store if you want to have some cool mugs or hoodies. Links in the description. See you next time, and until then, keep evolving. Okay, bye. [Music]