AI agents are coming into their own, and they’re coming up with approaches and techniques that are surprising even their human creators.
Matej Balog, a researcher at DeepMind, has shared an insightful anecdote about AlphaEvolve, a system capable of discovering new algorithms. His description highlights the AI’s capacity to devise solutions that diverge significantly from human intuition. It demonstrates the rapidly increasing sophistication of AI in problem-solving.
Balog recounts a specific instance: “One particular example I can talk about is we applied AlphaEvolve to discovering algorithms for matrix multiplication, and we did it by asking it to actually design a search algorithm, a gradient-based search algorithm, that looks for matrix multiplication algorithms in turn.”
He goes on to describe the meta-level approach taken: “So it’s a bit of a meta thing, you look for an algorithm that finds an algorithm, but within that first algorithm, the search algorithm, we started from a really simple code skeleton, giving it basically nothing. We just told it, ‘Use gradients basically.'”
The results were both unexpected and impressive. “Then it was able to write these complex loss functions and update functions, which had all sorts of tricks about penalizing various behaviors and introducing randomness in completely unexpected ways.”
Balog concludes with his personal reaction to this feat of AI: “Okay, well, this is the type of code that maybe a human could possibly write, but would they have actually thought of writing this particular piece of code? That was really an a-ha moment, at least for me, that wow, this is doing something kind of human-like, but not something that obviously a human would try.”
Just today, Google had announced its AlphaEvolve AI agent, which is Google’s newly-released coding agent for algorithm discovery. The agent is powered by Gemini, and Google says it is able to design faster matrix multiplication algorithms, find new solutions to math problems, and make data centers, chip design and AI training more efficient across the company.
And the implications of AlphaEvolve writing code that humans wouldn’t have thought of are immense. If AI can design more efficient and effective algorithms than humans, it opens up the possibility of solving problems that are currently intractable. From optimizing complex systems to accelerating scientific discovery, the potential applications are vast. Furthermore, the “unexpected ways” in which AlphaEvolve approached the problem suggest that humans might have created a sort of intelligence that’s different from their own — and if harnessed correctly, can move human civilization forward in ways that we hadn’t thought possible.