OpenAI’s recent progress on Navier–Stokes made me rethink one of the most common arguments against LLMs. For years, people have said that LLMs are “just” probabilistic fitting of their training data. Then they started solving IMO-level problems. Then research problems. Each time, the response was the same: this is not hard enough, still just recombination of existing knowledge.
But if a system can combine existing ideas in a way no human has found before, explore a huge search space, and eventually produce a genuinely new and correct solution to a problem that resisted experts for decades, what exactly is missing for us to call that intelligence or creativity?
Maybe creativity was never about creating something from nothing. Maybe it is mostly search: finding a tiny number of meaningful structures in an enormous space of possibilities. Human mathematicians also work with existing definitions, theorems, and proof techniques. The hard part is not whether the ingredients are old. The hard part is discovering a composition that is new, useful, and correct.
If that is true, then many forms of “deep intellectual creativity” may be far more mechanizable than we assumed, especially in fully digital domains like mathematics.
Saying “AI is just recombining existing knowledge” may be technically true. But so is saying “life is just atoms arranged in different ways.”
True, but so what?
@jacobli99 That's the common problem for many harness or prompt-tuning papers: you don't know what is transferable and what is model / dataset specific
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models
Tech Blog: https://t.co/ekQkO83jCv
RIP 😢 Yifan tragically took her own life shortly after leaving ETH, and, like many others on my timeline, I've been hearing some unsettling rumors her experiencing mistreatment while at ETH. I hope ETH can look into this and provide justice for Yifan and all parties involved.