David Garone
2.6K posts

David Garone
@DavidGaron14458
Exploring the nexus of art & AI. From Broadway to digital innovation, I'm on a quest to blend creativity with tech. #ArtTechFusion



[Responding to my friend @edfrenkel’s requests for comment.] Happy to lend a thought or two. It very much depends as to what we are calling “Real Mathematics”. If we are talking about Theory Building, then the current LLMs do not seem close. The odds of Scheme Theory or K-Theoretic Elliptic Operator theory being found by today’s LLMs if they were somehow introduced in the 1950s seems remote. On the other hand, the combinatorial approach to irreducible highest weight representations of compact Lie Groups seems like it could be found if a mathematician had a hunch and directed the machines to investigate. Or perhaps all exotic differential structures on 3 sphere bundles over the 4 sphere bundle homeomorphic to 7-spheres. So how to think about what is within reach? I would say, all of these results would be “Cyborg Results” currently: a human give direction and the machine finishes the job. So, what is going to happen going forward? I think it is pretty clear. Three things. A) The Cyborg Results (CRs) will get more profound. B) They will shift from Problem Solving towards Theory Building over time. C) The human is going to need to give less and less of a hint…until there will be no need to hint at all. And I think that it will be a while until all three happen at once. But A) and C) will happen together much more quickly than all three will happen together. B) is going to be the sticking point. But it too will give. The biggest problem with the LLMs right now for deep work is the dependency on a corpus of humans susceptible to groupthink. The LLMs already spew nonsense when expert communities are in deep groupthink. Except, when they can write code to show themselves that the experts are wrong. Which @grok heavy can do now, already with SymPy. The others seem more polutable. I think @elonmusk has a different orientation on this. More irreverent, expert-skeptical and less consensus oriented.









BTW, I've basically stopped using Opus entirely and I now have several Codex tabs with GPT-5-high working on different tasks across the 3 codebases (HVM, Bend, Kolmo). Progress has never been so intense. My job now is basically passing well-specified tasks to Codex, and reviewing its outputs. OpenAI isn't paying me and couldn't care less about me. This model is just very good and the fact people can't see it made me realize most of you are probably using chatbots as girlfriends or something other than assisting with complex coding tasks


Optical WiFi is vital for 1000s of reasons. Unfortunately the understanding will be common in 2125 but we will have to live through this period. Important to support this work. Thank you @rfsafe .








