Summary

AlphaEvolve turns algorithm discovery into iterative code search: Gemini models generate candidates, a user-supplied evaluator filters them, and evolutionary search preserves a diverse population instead of committing early. It can start from scratch or improve a heavily optimized incumbent, with outputs already running in key Google infrastructure. The mechanism is “the creative power of large language models” yoked to “the strictness of the evaluation function,” not a stronger one-shot coding prompt.
The lineage runs from AlphaGo’s “move 37” through AlphaTensor in 2022 and FunSearch to a general search over programs. Pushmeet’s thesis is that the capability that explored huge Go spaces can enter a “superhuman region of algorithms”; matrix multiplication was the proof because Strassen’s counterintuitive improvement had stood for more than 50 years. Matej rejects complacency: “the very best people” had attacked these problems, strengthening the evidence that AlphaEvolve found something genuinely difficult and new.
AlphaEvolve’s central scaling claim is continual improvement without the early plateau common in autonomous coding agents. Easy problems may resolve “almost immediately,” while decades-old scientific questions or high-value Google systems warrant much longer searches; the required generations cannot be predicted because apparently trivial questions can be extremely hard and vice versa. More search time and computational resources can support longer searches and better results.
Evaluators are both a bottleneck and an unlock: a precise objective can enable discoveries beyond the best developers, while an incomplete natural-language specification leaves ordinary agents uncertain about whether an answer is right or a hallucination. Pushmeet says AlphaEvolve can “leverage the hallucinations for a beneficial purpose,” then stress-test promising ideas across progressively more instances. The guests do not see exact evaluators as a conceptual limit: correlated simulators, auxiliary signals, LLM critiques, and proof agents could extend the search, though weaker signals may require more iterations.
A limited form of self-improvement is visible, but cognitive improvement is not yet validated. Sarah cites a 23% speedup in part of the training infrastructure “if I recall correctly”; Pushmeet says the current result is more efficient computation, while Matej describes the feedback loop as perhaps “on the order of months” for training the next generation of Gemini. Whether gains are one-off, diminish toward a limit, or “keep accumulating up and up and up” remains explicitly open.
Math and computer science are the near-term sweet spot because automated evaluation often comes “basically for free,” while biology and chemistry need simulators or predictive models for candidates such as molecules. Pushmeet frames science as search—for ideas, constructions, and drug candidates—and expects systematic evaluation to turn agents into a scientist’s “superpower.” Human input remains important for objectives and constraints, including the blunt drug-design requirement that a treatment not kill the patient.
The output is deployable, inspectable code, making AlphaEvolve more collaborative and reviewable than an opaque neural-network decision-maker. Mathematicians can reverse-engineer the construction rather than merely accept an answer; earlier cap-set work with Jordan Ellenberg exposed “interesting symmetries” they had not known. Google’s examples span data-center efficiency, hardware, and software, while a trusted-tester program informs broader release strategy under significant evaluator and compute requirements.

Deep dive

1. AlphaEvolve searches code rather than merely predicting it

Pushmeet’s concise definition is an “AI coding agent” that discovers algorithms for open scientific problems while producing practical code already deployed in “key parts of Google’s own infrastructure.” The chosen bar joins scientific novelty to production performance.
Pushmeet roots the approach in AlphaGo: if a neural agent can explore the enormous space of Go positions efficiently enough to invent “move 37,” the same philosophy might search algorithm space. AlphaTensor applied it to matrix multiplication in 2022, providing evidence for a “superhuman region of algorithms.”
Sarah tests whether valuable solutions remained undiscovered because researchers were complacent. Pushmeet instead emphasizes constructions that are intricate, non-intuitive, and buried in an enormous search space; Matej adds that they deliberately selected problems worked on for years by “the very best people,” where existing results had survived unusually strong scrutiny.
AlphaTensor was purpose-built for matrix multiplication. FunSearch then moved into “the space of programs” and produced what Pushmeet calls the first scientific discovery from an LLM; AlphaEvolve extends that line into a more general agent.

2. The evaluator defines the destination while evolution finds the route

For data-center scheduling, the user supplies an evaluation function—such as an existing simulator that scores how well an algorithm schedules jobs. In Pushmeet’s formulation, the user defines “the what” and AlphaEvolve fills in “the how,” either starting from nothing or from a highly optimized incumbent.
Gemini models propose creative code changes, the evaluator separates working ideas from failures, and an evolutionary algorithm maintains a diverse population. Rather than “commit to a very specific type of solution early on,” it can combine ideas from different strong candidates before engineers consider deployment.
Each generation selects from prior candidates and introduces combinations or new ideas, so the population should become stronger. The evolutionary analogy is literal at this level, although the useful unit being mutated and selected is executable code.
Difficulty determines the search budget: an unexpectedly easy problem may be solved “almost immediately,” while a decades-old open question may require extended exploration. Matej stresses that sustaining improvement without plateauing is itself difficult; predicting the iterations in advance is not, because trivial-looking questions can prove extremely hard and vice versa.

3. Evaluators convert hallucination into useful exploration

Pushmeet contrasts AlphaEvolve with ordinary coding agents receiving incomplete natural-language specifications. Models make mistakes, and without a reliable test the user cannot know whether an answer is right; AlphaEvolve can instead “leverage the hallucinations for a beneficial purpose” by treating surprising proposals as candidates rather than conclusions.
Evaluation also governs search depth. A novel idea might deserve tests on a few instances, “a thousand different instances,” or stronger stress-testing before it influences later generations; balancing cheap filtering against thorough validation determines which creative branches survive.
Pushmeet calls strict evaluation the easiest entry point, not “a conceptual limitation.” LLMs can critique natural-language proposals, while imperfect simulators and auxiliary objectives can still guide evolution when they are “correlated with the true signal”—potentially requiring more iterations but allowing progress between exact scoring and subjective judgment.
Pushmeet notes that even simulators test a particular distribution rather than prove universal behavior, so a proof agent might be needed for guarantees. Conversely, the multi-agent AI co-scientist uses LLMs to assess hypotheses for novelty, significance, and impact, producing “much much much better results” than the base model alone.

4. A limited form of self-improvement is visible, but its curve is unknown

Sarah asks whether a 23% speedup in part of the training infrastructure—cited with “if I recall correctly”—constitutes recursive self-improvement. Pushmeet answers “in some senses, yes,” but confines the demonstrated improvement to reduced computation time; Matej describes the current loop as speeding training for the next generation of the Gemini model.
Making training faster is not yet the same as making a model “fundamentally better in certain cognitive tasks.” Pushmeet sees no reason the latter should fail if an evaluator can identify beneficial training changes, but says that evaluator and the resulting capability improvement still need validation.
Matej adds that the current loop is “maybe on the order of months.” Future gains could be one-off, continue while shrinking toward a limit, or keep accumulating; which trajectory governs is “a big open question” the team cannot answer today.

5. Scientists remain responsible for objectives, constraints, and understanding

Math and computer science came first because evaluators are common and sometimes available “basically for free.” Biology or chemistry could fit the same architecture when a simulator or predictive model can score a candidate molecule, while future AlphaEvolve versions may relax today’s requirements further.
Pushmeet’s broader framing is that much of science consists of searching for the right idea, construction, or drug candidate. As material and drug discovery become more systematic, agents can exploit those evaluations as a “superpower,” but humans need to encode properties such as efficacy, side effects, and delivery mechanism as objectives or hard constraints.
Matej describes real use as “surprisingly collaborative.” AlphaEvolve discovers an algorithm that constructs the solution, and that algorithm may matter more than the final answer because mathematicians can inspect how the ideas fit together; its code looks humanlike, yet prompts the question, “Would you have thought of writing it in exactly this way?”
In earlier cap-set work with Jordan Ellenberg, generated programs revealed “interesting symmetries that mathematicians did not know about.” The same inspectability lets production engineers review code before deployment—a “completely different league” from trusting an opaque neural network—while trusted testers inform broader applications under significant evaluator and compute requirements.

A Big Ruling on LLM Training and Midsummer Mail on NBA Salaries in Tech, Starting from Scratch in 2025, and More Zuck's $100BN AI Budget Buys Friedman & Gross; Navan Files for IPO