Keep pulling the thread on Pushmeet Kohli and Matej Balog.
AlphaEvolve is an AI coding agent that uses Gemini models and evolutionary search to discover new algorithms for open scientific problems.
The self-improvement feedback loop for AlphaEvolve speeding up the training of the next generation of Gemini models is currently on the order of months.
In 2022, the AlphaTensor system was the first AI agent to discover algorithms for multiplying matrices that were better than those created by humans.
It remains an open question whether AI self-improvement will result in a one-off benefit, diminishing returns, or continuously accumulating gains.
Algorithms discovered by the AlphaEvolve agent have been deployed in key parts of Google's infrastructure.
AlphaEvolve has demonstrated its versatility by improving efficiency across Google's full technology stack, including data centers, hardware design, and core software.
AlphaEvolve has been used to improve the efficiency of its own training infrastructure, demonstrating a form of recursive self-improvement.
The FunSearch agent was the first LLM-based system to make a scientific discovery by searching the space of programs.
An earlier version of AlphaEvolve discovered programs for the cap set problem that contained interesting mathematical symmetries previously unknown to mathematicians.
AlphaEvolve is designed to sustain improvement over time on difficult problems, avoiding the early plateaus common in other automated agents.
The output of AlphaEvolve is human-interpretable code, which allows engineers to inspect, understand, and verify the solution before deployment.
Solving difficult problems with AlphaEvolve requires a significant amount of computational resources beyond a single LLM call.