0xMo0nR3kt3r

If your AI agent is making you angry every day, you've actually discovered the central unsolved problem in AI research. You just might not know it yet You set up an agent. It works. Then it doesn't. Constant regressions. It sends your coffee to your house for 25 days straight, then to your neighbour on day 26. You’re frustrated at your slop-assistant all the time because it can’t do the same thing twice in a row So you build a task pipeline with verification steps. Still breaks. You add memory. Better, but now your agent can't properly navigate its own memory. So you think, "I just need to break reality down into clean categories.” Tasks, goals, messages, meetings, values, etc… then it will finally stop being so probabilistic” Ok great! You need a system of ontology. But the moment you try to build one, you've walked into a problem philosophers have failed to solve for 2,500 years. There is no perfect ontological system describable in language This is because humans don't think in language. We think in fuzzy concepts that map to many words or resist description entirely. Sensations, spatial reasoning, vibes. Language is a lossy compression of reality It turns out the losses are much smaller than AI researchers expected, which is why LLMs blindsided most of the field. But the losses are real, and agents expose the fact that language never perfectly models reality LLMs will keep improving. But to fully replace human economic output, we probably need world models, not just better language prediction