one year on
Google DeepMind unveils AlphaEvolve, a Gemini-powered agent that designs algorithms for math and computing
The system has already been deployed inside Google's data centers and chip design pipeline, and claims improvements on open mathematical problems including the kissing number.
Google DeepMind today announced AlphaEvolve, an AI agent that combines Gemini large language models with automated evaluators to discover and optimize algorithms for mathematics and computing. The system uses an ensemble of Gemini Flash and Gemini Pro models in an evolutionary loop: generate candidate code, test it against automated metrics, and refine the best performers.
DeepMind reports that AlphaEvolve has already been deployed across Google’s infrastructure. A scheduling heuristic it discovered has been running in Google’s Borg data center orchestrator for over a year, recovering on average 0.7% of worldwide compute resources. In hardware design, the system proposed a Verilog rewrite that removed unnecessary bits from a matrix multiplication circuit, which has been integrated into an upcoming Tensor Processing Unit. The agent also achieved a 23% speedup in a matrix multiplication kernel used in Gemini’s own training, reducing total training time by 1%, and up to a 32.5% speedup in the FlashAttention kernel.
On the research front, the system tackled over 50 open problems in mathematics. It rediscovered state-of-the-art solutions in roughly 75% of cases and improved the best known solution in 20% of cases, including a configuration of 593 outer spheres that establishes a new lower bound in 11 dimensions. The team also highlights a new algorithm to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving on Strassen’s 1969 algorithm which required 49.
On Hacker News, where the announcement quickly rose to the front page with over 1,000 points, commenters dug into the matrix multiplication claim. Commenters pointed to Winograd’s 1967 algorithm and Waksman’s 1970 algorithm as possible precedents, while others argued the rank-48 result may still be new. Other commenters questioned whether the reported kernel speedups might suffer from the same benchmarking issues seen in recent AI-generated CUDA kernel claims, while others argued that the combination of LLM-based search with automated validation is a genuine step forward for combinatorial optimization.
The record
Debated whether the 48-multiplication 4x4 complex matrix algorithm is truly novel, citing Winograd's 1967 algorithm and Waksman's 1970 algorithm as possible precedents, while others argued the rank-48 result may still be new. Some expressed skepticism about claimed speedups, recalling the Sakana AI CUDA kernel hoax. Others praised the system's potential for automating tedious optimization.
One year later — open only if you can handle spoilers
Over the following months, the community remained divided on the novelty of the matrix multiplication result, though the system's practical impact inside Google's infrastructure was quietly noted. The broader approach of using LLMs for algorithm discovery saw continued investment from both DeepMind and other labs, but no rival system matched AlphaEvolve's breadth of results in its first year.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy