Slechte chocola mag uit de koelkast komen, maar goede chocola moet op kamertemperatuur.

One challenge with the "convex hull of ideas" analogy (irrespective of its correctness) is that human intuition for high-dimensional convex geometry is notoriously poor, and requires simultaneous maintenance of a few simpler intuitions which can each conflict at different times.

1/ 🎉 Excited to present at #EMNLP2026 🎉 Transformers and Mamba are fundamentally different. Do they also represent things differently? Globally, they differ. Locally, they are similar. 🧵

Mathematicians used to study Euclidean geometry. It was beautiful. But coordinates and calculus together became a more powerful tool that surpassed it. Even though they were often very ugly and involved a lot of grinding work, they made Euclidean geometry essentially obsolete.

I am enjoying the new Codex autocomplete, but I'm curious how it works. It seems to pre-prepare the same annoyed comments I would make. If the predictions are from the LLM, then surely it should have known I wasn't going to be happy with its output in the first place.

So not to go all Yudkowsky Monocause Theory. But a lot of AI company talk that *sounds smart* did come from me, or me via Bostrom or other diffusion modes. High-powered ideas like "recursive self-improvement" and "AGI" are ideas they did not invent themselves, but got off higher-quality outside thinkers. There's a big intellectual history behind me in turn, but the AI companies did not study this history independently; they got it downstream of my writing it up. I naturally spent some time trying to figure out what the positive vision for AGI and ASI would look like, back in the mid-2000s. But I considered it a lower priority to write up. And AI company executives are not at the intellectual level where they can invent coherent answers to questions like that, if they don't have my answer sheet to copy from. They can't tell you what the positive vision for AGI or ASI looks like, because I didn't prioritize writing up that part of my earlier thoughts. They can't copy it off from the same answer sheet they used to produce smart-sounding conversation like "general intelligence" or "recursive self-improvement" or "extinction risk" or "AI alignment problem". If they're not reading off my answer sheet, AI company executives are pretty ordinary as intellectuals. Being able to spend billions on GPUs does not make you a genius inventor, even if we're used to amazing inventions coming from genius rather than black-box optimization. AI company executives are not going to give you better answers about "what's the positive vision" than any other random intellectual challenged with a question beyond their wits. If you start hearing good-sounding answers from Anthropic or OpenAI, it means that Mythos or HPIM advanced to the point of learning how to sound convincing about utopia to humans. I did do an *abstract* writeup on constraints and requirements over positive ASI scenarios (see eg "31 Laws of Fun"); but that does AI companies no good, because they would need to be able to sketch a concrete AGI scenario from those abstract ideas. That again is beyond them if I have not done their homework. It is of course not in humanity's interest that I try to do this homework for the AI companies now. It would only be a lie. They have no utopia to offer.

Develop nuclear power.