OpenAI's Math Breakthroughs Aren't Ready for Prime Time

OpenAI recently announced that its AI had solved hundreds of the world's hardest math problems. But a group of top mathematicians, called AGMAI, says OpenAI didn't follow their new guidelines. These guidelines are meant to ensure that AI math solutions are trustworthy and that humans can understand them. AGMAI's first rule was to stop testing advanced math problems on private AI models, but OpenAI did exactly that.

The guidelines also ask for detailed explanations of how the AI reached its answers. OpenAI only provided these for 10 out of 719 solutions. And while AGMAI suggested formalizing proofs (writing them in a computer language that can be checked), only 42% of OpenAI's proofs were formalized. Even worse, a new paper found mistakes in how OpenAI translated its solutions from plain English into computer code for a famous fluid dynamics problem. These errors don't necessarily mean the solutions are wrong, but they show that AI can't be fully trusted without human review.

Famous mathematician Terence Tao criticized OpenAI's approach, saying that AI is solving problems without anyone truly understanding them. He points out that the people prompting the AI often don't care about the math itself and can't explain the results to others. This is a problem because math is about building knowledge, not just getting answers. If no one understands the solutions, they're not very useful.

So, while it's impressive that AI can tackle hard math, it's not ready to replace human mathematicians. The math community needs to review AI's work just like they do for human-written proofs. Until then, we should be cautious about trusting AI's math solutions for anything important.