Skip to content
No PriorsNo Priors

No Priors Ep. 120 | With Google DeepMind’s Pushmeet Kohli and Matej Balog

Much of the scientific process involves searching. But rather than continue to rely on the luck of discovery, Google DeepMind has engineered a more efficient AI agent that mines complex spaces to facilitate scientific breakthroughs. Sarah Guo speaks with Pushmeet Kohli, VP of Science and Strategic Initiatives, and research scientist Matej Balog at Google DeepMind about AlphaEvolve, an autonomous coding agent they developed that finds new algorithms through evolutionary search. Pushmeet and Matej talk about how AlphaEvolve tackles the problem of matrix multiplication efficiency, scaling and iteration in problem solving, and whether or not this means we are at self-improving AI. Together, they also explore the implications AlphaEvolve has to other sciences beyond mathematics and computer science. Sign up for new podcasts every week. Email feedback to show@no-priors.com Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @pushmeet | @matejbalog Chapters: 00:00 Pushmeet Kohli and Matej Balog Introduction 0:48 Origin of AlphaEvolve 02:31 AlphaEvolve’s Progression from AlphaGo and AlphaTensor 08:02 The Open Problem of Matrix Multiplication Efficiency 11:18 How AlphaEvolve Evolves Code 14:43 Scaling and Predicting Iterations 16:52 Implications for Coding Agents 19:42 Overcoming Limits of Automated Evaluators 25:21 Are We At Self-Improving AI? 28:10 Effects on Scientific Discovery and Mathematics 31:50 Role of Human Scientists with AlphaEvolve 38:30 Making AlphaEvolve Broadly Accessible 40:18 Applying AlphaEvolve Within Google 41:39 Conclusion

Sarah GuohostMatej BalogguestPushmeet Kohliguest
Jun 25, 202542mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

AlphaEvolve: Google DeepMind’s AI Agent That Discovers New Algorithms

  1. The episode explores AlphaEvolve, a Google DeepMind autonomous coding agent that uses Gemini models plus evolutionary search to invent new algorithms for hard scientific and engineering problems.
  2. Pushmeet Kohli and Matej Balog explain how AlphaEvolve builds on AlphaTensor and FunSearch, moving from single-problem systems to a general agent that searches directly in program space.
  3. They detail how users define evaluation functions, how evolutionary loops refine candidate programs, and how AlphaEvolve has already produced deployable improvements across Google’s infrastructure and on long‑standing math/computer science problems.
  4. The conversation touches on technical creativity, the importance and limits of automated evaluators, early signs of self‑improving AI, and how such agents could transform scientific discovery while remaining collaborative tools for human experts.

IDEAS WORTH REMEMBERING

5 ideas

Define precise evaluation functions to unlock AI-driven discovery.

AlphaEvolve only needs a clear way to score candidate solutions; once a simulator or evaluator is in place, it can search the space of programs and uncover algorithms that surpass long-standing human-designed methods.

Treat LLM “hallucinations” as fuel for structured search, not just errors.

Within AlphaEvolve, the model’s creative, sometimes wrong ideas are aggressively filtered by evaluators and evolutionary selection, turning what is usually a liability into a systematic source of novel algorithmic candidates.

Maintain diverse candidate solutions to avoid premature convergence.

The evolutionary loop is explicitly designed to preserve a population of different approaches, recombining and refining them over many generations so the system doesn’t lock into a suboptimal early idea.

Use AI agents where search spaces are huge and intuition fails.

Problems like matrix multiplication or data center scheduling involve astronomically large, non-intuitive search spaces; AlphaEvolve excels here by systematically exploring regions humans are unlikely to stumble upon.

Prioritize interpretable code outputs for safe deployment.

Unlike opaque neural policies, AlphaEvolve outputs human-readable code, allowing engineers and mathematicians to inspect, understand, and gatekeep solutions before integrating them into critical systems.

WORDS WORTH SAVING

5 quotes

AlphaEvolve is an AI coding agent that is able to discover new algorithms that are able to make new discoveries on open scientific problems.

— Matej Balog

We are able to leverage the hallucinations for a beneficial purpose.

— Pushmeet Kohli

It is a tool that is already available inside Google and it is being used for many, many problems.

— Matej Balog

We are maybe seeing the first sign of self-improvement, but one also needs to be very specific about what we have shown so far.

— Matej Balog

You can just sort of see this as a tool that will give scientists a superpower in their ability to search over very complex and sometimes counterintuitive solution spaces.

— Pushmeet Kohli

Origins and evolution of algorithm-discovering systems: AlphaTensor, FunSearch, and AlphaEvolveHow AlphaEvolve works: evaluation functions, evolutionary search, and LLM-based code generationWhy humans missed these algorithms: vast search spaces and non-intuitive constructionsThe central role and future of evaluators, simulators, and AI-based critiquesEarly signs and constraints of recursive self-improvement in AI systemsApplications in Google’s infrastructure: data centers, training pipelines, hardware and software optimizationImplications for science, mathematics, and the evolving role of human scientists and engineers

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome