— AlphaEvolve is a Gemini-powered AI coding agent that is able to make new discoveries in computing and mathematics, says Matej Balog, a researcher at Google DeepMind, to VentureBeat.
Google doesn’t say when the model will be available to the general public, but is planning on working with certain scientists and is «exploring» a broader release.
Picks the best LLM outputs
The model works by combining two of Google large language models. Gemini Flash for speed, and Pro for more granular details. It then evaluates the results and picks the best one for further, well, evolvement.
Once results are in from the LLMs, “evolutionary coding agents” improves on them and can produce hundreds of lines of complex code or algorithms, for example by cutting LLM training time by 1 per cent — which might seem small, but is a massive leap forward in costs and energy.
Improved its own hardware
It also made a change to the hardware that drives it, identifying a software error that was dropping unnecessary bits and improving efficiency in Google’s Tensor Processing Units, which is expected to be included in upcoming production chips.
When focused on a collection of over 50 open math problems, AlphaEvolve also improved on the best known solutions in about 20 per cent of cases, and matched them in 70 per cent of cases.
Google thinks AlphaEvolve will be useful for material science, drug discovery and other algorithm heavy applications.
It is currently developing a user interface for the model, and will launch a program for «selected academic users,» but offers no timeline. They also say they are «exploring possibilities to make AlphaEvolve more broadly available.»
Read more: DeepMind’s launch post, detailed at VentureBeat, explained at Ars Technica.