An AI That Took On One of the World's Unsolved Problems — for $2,000

This episode looks at a recent headline — an AI reportedly made meaningful advances on long-standing unsolved problems in mathematics and theoretical computer science, at a compute cost estimated around $2,000.It unpacks what that number actually means: not a prize, not the cost of building the model, but an estimate of what the compute used by Astra, OpenAI's next model, would translate to in API costs as it worked through a set of research-level problems across areas like sphere packing, group theory, and cryptography.There's a quiet observation at the center of this episode — that the cost of thinking itself seems to be falling. The hundreds of billions spent building such a model is one thing; what it costs to then put that model to work on problems human researchers have spent years on is apparently something else entirely.It also touches on what this might mean for researchers: less a story about AI replacing human judgment, and more one about a shift in role — from people who search for answers to people who evaluate them.A small reflection on a number that could easily seem large or small depending on how you look at it, and what it might signal about where AI stands in relation to the frontier of human knowledge.