Google DeepMind has just done something amazing: it’s solved a longstanding math problem with a brand-new solution.

It's a huge achievement, not just for AI's capabilities in abstract problem-solving but also for the broader implications in fields like data infrastructure.

DeepMind used its large language model, FunSearch (fun meaning ‘function’, sadly), to solve the cap set problem. FunSearch combines a language model, fine-tuned on computer code, with systems that refine and filter solutions.

That this AI produced a hitherto unknown solution is incredibly exciting. It offers a counter-example to LLMs’ growing reputation for hallucinating information: now we know AI can not only handle scientific data but also contribute genuinely new solutions to science that would otherwise require human ingenuity.

It's also fascinating - and slightly unnerving! - that Alhussein FawziDeepMind research scientist, seemed unsure about the exact reasons behind their success: ‘We have hypotheses, but we don’t know exactly why this works.’ Comments like this underline just how experimental and quickly advancing some of this AI research is.

It’s a milestone to celebrate, for sure. But it's also worth considering the long-term impact these kinds of breakthroughs are going to have on the supporting infrastructure. With AI tackling increasingly complex tasks, the volume and sophistication of data infrastructure are going to need to keep pace.

Here's the original article: Google DeepMind used a large language model to solve an unsolved math problem | MIT Technology Review