Announcing AlphaEvolve, our new LLM coding agent that has
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
White paper:
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
White paper:
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf
Comments
Would be nice to see it benchmarked vs something where the human effort can be easily quantified
http://ciar.org/h/notes.cmc.txt
Maybe Gemma3-27B is good enough to implement CMC even without fine-tuning? I should find time to dork with it.
https://bsky.app/profile/did:plc:i4jevytmqw4yg2vplnuyz6e3/lists/3kcykc7b3v62r
(I'm not at all a fan of crypto, and I am not interested in promoting "AI" just understanding its impact)
It found the first improvement after 56 years to how fast we can multiply 4x4 matrices 🤯 (with complex number entries)
And it did so by discovering a sophisticated search algorithm
I’m looking forward to many more applications across the sciences and computational problems
On BlueSky: @emidup.bsky.social @swarat.bsky.social
I implemented something inspired by it to optimise agents within my jax "swarm" environment. It works with only 3x1080 Ti running Qwen3-14B, amazing!
Repo https://github.com/JakeForsey/swarm
Best https://github.com/JakeForsey/swarm/blob/main/swarm/agents/vibevolve_v5.py