Last week we covered the announcement of the artificial intelligence (AI) system AlphaEvolve, by Google DeepMind. The first reactions to the announcement are coming in.
Experts praise AlphaEvolve as a major step, Nature News notes (unpaywalled version). It’s the first system using general-purpose AI to make new discoveries.
AlphaEvolve uses DeepMind’s Gemini models to create code for solutions. Users give a problem, a way to judge answers, and a starting idea. The system suggests thousands of tentative solutions. An algorithm checks these against goals, like using fewer resources. Over time, AlphaEvolve explores many ways to solve a problem and builds better solutions.
“We explore this diverse set of possibilities of how the problem can be solved,” says research co-leader Matej Balog as reported by Nature News.
AlphaEvolve builds on FunSearch, a 2023 tool that solved math problems. It found a faster way to do matrix multiplication, a math process used in AI training. It also outperformed AlphaTensor, a DeepMind tool made just for matrix math.
AlphaEvolve could help design scientific instruments and materials if clear goals or simulations exist. But it may only work for problems that can be coded. Sceintsts want more testing outside DeepMind and would prefer an open-source version over DeepMind’s private system.
DeepMind states that AlphaEvolve takes less power than AlphaTensor but still too much to share freely, and hopes scientists suggest new uses.
AlphaEvolve shows signs of being self-improving AI
Popular AI show host Matthew Berman enthusiastically comments that AlphaEvolve shows signs of being self-improving AI. "One thing that is absolutely critical for achieving self-improving AI is the ability for AI to discover new knowledge," he says. "Sam Alman and the Open AI team said that's coming next year and it seems like Google may have beaten them."
Google has already deployed AlphaEvolve across its data centers, chip designs, and AI training systems. AlphaEvolve "demonstrates an intriguing evolution in AI itself," notes VentureBeat, "starting within the digital confines of Google’s servers, optimizing the very hardware and software that gives it life."
