From 49 Multiplications to 48: DeepMind’s AI Rewrites a Fifty-Year Math Record

Pushmeet Kohli, who leads AI for Science at Google DeepMind, likes to open with a story from a biology conference. After his talk, a scientist came up and said he had spent ten years on a single protein, collecting lab data to pin down its structure, and still had nothing that would yield. He had plenty of measurements. If he had known the shape, he could have checked it immediately. He ran AlphaFold 2. It returned a structure that fit the data perfectly. “I’ve been working on this for 10 years,” the biologist said. “What do I do next?”

On Sequoia Capital’s Training Data podcast, hosts Sonya Huang and Pat Grady put the obvious question to him. The field has been waiting for the moment when AI stops answering questions and starts making novel scientific discoveries. Is AlphaEvolve that watershed? Kohli calls it a key milestone rather than a finish line. What the team has shown, he says, is that a large language model, once coupled with a harness, can discover new algorithms and prove mathematical results that researchers have studied for many years.

That word “harness” is doing a lot of work. Kohli places AlphaEvolve in a longer history of AI for science, with AlphaFold as the prototypical example: AlphaFold 2 came out at the end of 2021 and won the Nobel Prize in 2024, making the impact of specialist models hard to argue with. The open question was whether large language models and foundation models could matter in the same way. FunSearch, released about two years before this conversation, was DeepMind’s first serious attempt to find out.

FunSearch took an LLM and coupled it with an evaluator whose job was to say whether a new conjecture was a hallucination or a genuine insight. In this setup, Kohli says, hallucinations were actually useful, because some of them were brilliant ideas nobody had thought of. The harness is that combination of an evaluation function and a search protocol sitting around the language model. Together they can produce discoveries that are new and, in his view, genuinely impactful.

The first version still had a human in the loop in a different sense. FunSearch asked the model to complete a small function inside a template the researcher had to provide, and even with weaker models it found algorithms mathematicians had been chasing. AlphaEvolve drops that restriction. Instead of filling in a few lines, it searches over whole algorithms—large pieces of code—and keeps optimizing them over a long period. It also needs far fewer function evaluations than FunSearch: by looking at fewer proposals, it discovers new algorithms much more quickly.

The engine is Gemini, and Kohli says each generation’s better grasp of code is why search gets more sample-efficient. When the proposal generator understands code, the candidates it emits are not just syntactically valid; they are semantically trying to solve the task. Searching a large space, he notes, is a race on two clocks: how fast you can generate a new candidate algorithm, and how fast you can tell whether it is any good. Gemini Flash matters here because it can propose candidates efficiently and quickly.

AlphaEvolve is also broader than its predecessors in language and domain. It can search in C++ and Python, and in Verilog, the hardware-description language used to design chips. The only requirement is a function evaluator that can score a proposal quickly. If you can say, for a given program, how good it is, the system is in scope. That generate-evaluate-keep loop, one of the hosts observes, roughly mirrors the scientific method.

Kohli points to another DeepMind system released earlier in 2025: AI co-scientist. There, Gemini plays the entire academic process. One instance generates hypotheses, another critiques, another reviews and ranks ideas, another edits. They are the same underlying model, prompted to take different roles in a multi-agent setup. The combined system, he says, produced behavior well beyond a single Gemini’s answer, with much better proposals and new ideas.

Why that works is still being studied. Kohli’s observation is that the first answer co-scientist gives is often not very different from baseline Gemini. Leave the multi-agent system running not for minutes or hours but for days, ranking and refining, and the answers get much better. His guess is that deep insights sit in the tail of the distribution, and Gemini is better at judging which idea is stronger than at inventing a strong idea from scratch. It is the old computer-science split: verifying a solution is often easier than finding one.

The hosts note that generators paired with verifiers now show up across the industry, in general models and in narrow systems. Kohli agrees the field is only at the start of agent research. In AlphaEvolve the generator is a foundation model and the evaluator can even be hand-coded, tied together with evolutionary search. In co-scientist, several agents share a memory. What the optimal agent configuration looks like, he says, is still an open research problem.

The motivation goes back to DeepMind’s talent for searching huge spaces. Reinforcement-learning agents had already handled Go and StarCraft. The team asked whether the AlphaZero family could discover new algorithms, and built AlphaTensor to attack matrix multiplication. It improved on results that had stood for fifty years. Two questions remained: could they do better, and could they produce a solution that was more interpretable?

A practical problem inside Google made the interpretability question urgent: how to schedule jobs in a data center. Heuristics designed by some of Google’s best engineers already had a huge effect on compute utilization. A typical RL agent might beat those heuristics, but then a neural network would be deciding which workloads go to which machines, and if something broke, nobody would know how to debug it. What engineers really wanted, Kohli says, was a piece of code they could read and run.

So instead of searching only in the space of matrix-multiplication algorithms, as AlphaTensor had, or training a neural-network policy to solve the problem directly, they built an agent that searches in the space of programs. The benefit is interpretability: you can inspect the code and its properties. Experts who looked at the resulting programs could recover insights from them. That is the path that led from FunSearch to AlphaEvolve.

One of the math problems FunSearch took on was the cap set problem, a question Terence Tao is known to care about—he has called related questions among his favorites. DeepMind collaborated with mathematician Jordan Ellenberg, and when they inspected the program FunSearch had written, Ellenberg found symmetries in the problem that had not been recognized before. The agent had discovered those structures and used them to get a better solution, without anyone teaching it the trick. It is the algorithmic cousin of AlphaGo’s Move 37: a path humans did not take, written down in code they can still read.

Asked whether math is the gold standard for testing whether models generate novel scientific results, Kohli says it has a useful property: it is precise. You know whether you have found the property you were looking for. For 4×4 matrix multiplication, the best known method, due to Strassen, used 49 multiplications; AlphaEvolve showed you can do it with 48. There is no arguing about that number. You do not need RLHF to decide whether the output was “nice,” and you do not need another language model to score it. You just know you are better.

That crispness is harder to find in the physical world, but Kohli says AlphaEvolve is already useful wherever you have an evaluator you trust. Data-center scheduling and Verilog chip design are the examples the hosts raise. Unlike a human programmer who might try ten, a hundred, or a thousand things, AlphaEvolve can go on and on, inventing counterintuitive strategies no one would have imagined. Humans can serve as evaluators, he adds, but then the limits are scale and accuracy: how many programs can you score, and how well?

The team has not used a human in the loop for AlphaEvolve; most evaluators were programmatic. Kohli still sketches a hypothetical: if the system produced several equally strong algorithms for a math problem, the “best” one would be the most elegant or the simplest to understand, and elegance is grounded in a human observer. Connecting digital results to the lab, as in materials science, depends on how much you trust the computational evaluator. If you fully trust it, the job is done. If you do not, you still have to validate the result in the real world.

As the models get better at emitting Verilog, one host asks, does the chip designer go away? Kohli answers with what happened after AlphaFold. Proteins are the Lego blocks of life; their shapes explain function and unlock drugs and enzymes. Before AlphaFold 2, a single structure could take one to five years and a million dollars, and some proteins had resisted decades of work. Only about 37 percent of human protein structures were known. After release, the team could predict structures for essentially all proteins on the planet and made those predictions available to everyone.

That is the conference story in full: the biologist who had spent ten years on one protein ran AlphaFold 2, watched the prediction match his lab data, and asked what to do next. Kohli says AlphaFold then did three things. It advanced structural biology: work that used to need a synchrotron, six months, and a million dollars now takes a second. It accelerated the field. And it democratized it. A scientist in Latin America, South Asia, or Africa working on a neglected tropical disease, with no funds and no instruments, suddenly had structures for whatever parasite they were studying.

Those scientists did not stop working. Structures were no longer scarce, so they moved to the next questions: how to use that knowledge to treat disease and design better drugs. Kohli expects the same shift with AlphaEvolve. Once agents can solve problems beyond human ability, the live question becomes which problems to solve, and which chip properties to push: efficiency, cooling, cheaper construction, fault tolerance. You can make the objective more sophisticated because you now have more sophisticated systems to optimize it.

Has AlphaFold actually produced an inflection in new drugs? It has sped things up, Kohli says, but drug discovery is a long chain. First you pick a target protein; then you design a molecule that binds it, a step that used to take as long as two years and can now take weeks or months. After that come phase one, two, and three trials, plus toxicity. AlphaFold removed one blocker and shortened the overall timeline. Other blockers remain, and a new generation of AI-for-biology models is aimed at those. The team has taken a big step, he says, but it still needs to take a few more.

Asked which domains will be most lucrative for this family of models, Kohli flips the question: which domains matter to society? AI, he thinks, will accelerate healthcare, materials science, and smart systems across the board. He notes that we name eras of civilization by their materials—stone, bronze, iron, and now, depending on your mood, silicon or plastic. What sets humans apart from other species is the ability to transform energy. A room-temperature superconductor, or cheap fusion that drives the price of energy toward zero, would rewrite geopolitics and economics in ways that are hard to predict.

The same is true of coding agents: if everyone can write software, intelligence becomes ubiquitous. Everything from materials to energy to healthcare will be affected. On whether scientific discovery will have a fast takeoff, Kohli does not point to a future cliff. “I think we are living through the middle of it,” he says. When you are in the middle you do not really see it, but the era of AI-accelerated scientific discovery has, in his view, already begun.

He names two bottlenecks. The first is validation: bridging the gap between the digital and the real world, and capturing what actually matters for the problem. The second is accessibility. You can build the most sophisticated technology, and if people do not know how to use it, you will not get the impact you want. AlphaFold is used not only because it is accurate—Kohli puts it around 90 to 95 percent, not 99—but because it knows its limits. When it is going to be wrong, it “holds its hand” and says so. Today’s language models, he argues, still lack that calibrated uncertainty. The scientist who spends two years chasing a bad prediction will never come back.

In a rapid-fire close, Kohli names co-scientist as the must-read paper of the year and the wake-sleep algorithm—work from MIT’s Kevin Ellis and Josh Tenenbaum—as the under-discussed idea he cares about: you should not only write programs, you should build libraries of common modules that make every future program easier. He only “somewhat” agrees that inference-time compute will be the next major leg of scaling; training-time compute and distillation still matter because better internal representations make models more accurate and more efficient. He is bullish on robotics in the medium to long term, and on humanoids because the built world was designed for human form. Future science Nobels, he says, will not all go to AI teams: humans are still winning them, but there will come a point when AI is indispensable and the prizes go to humans and AI working together.