This is a special post for quick takes by Prometheus. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
In all the interactions with AI, there has been one recurring problem that doesn’t seem to go away: they don’t get it. I don’t know how to explain this in any better language than that. I don’t know how to create a “Get It” benchmark. But whenever I talk to Claude, ChatGPT, Gemini, or any other model about a concept, the longer the interaction lasts, the more I get the sense that it doesn’t really “get it”. In this way, I think AI skeptics are actually pointing to something real when they say they’re not “real intelligence”. A lot of their arguments are poor, but I think there is something truly missing. And it doesn’t seem to change with improvements in other capabilities. GPT 5.3 seems just as bad at “getting it” as GPT 3.5. I don’t think Claude “gets” simpler concepts, but struggles with more complex ones. It doesn’t seem to “get” any concepts at all, simple or otherwise. I don’t know what this means going forward, or if “getting it” is even needed to start an intelligence explosion. I can imagine there are cases where an AI could be a capable AI researcher, without having to really “get it”, just as they can be capable coders. But it could be another obstacle to alignment, where the AI creating a smarter AI will fail to align it, not because it is misaligned, but just that it is too stupid to really grasp its own values.
Since writing this, my perception has changed drastically. I don't know if it was Fable 5, or Astra, or Opus 5.5, but there's a much greater sense of them "getting it" now. It's felt like a categorical change in the past six months, where I went from constant hand-holding to these models actually being truly useful. Everyone has their own ceiling, where the cognitive work we do is within the reach of AI. I think I might be reaching mine. Assessing first-hand their true ability may be more difficult in the future, as it will likely overtake my personal capacity to determine what a certain threshold above my capability is.
I definitely feel like Claude suddenly "gets it" a lot more when it comes to code. I'll ask it to do something and it'll come back and mention that it noticed a problem with how the code was organized and fixed it. Or the approach I suggested doesn't work so it wrote code that does something else (in a good way -- it was right, and it honestly reported what it was doing and asked me to check its work).
It's still very inconsistent about this though. I'm working on a game where fun is a well-defined formula (it's a Cookie Clicker style game), but it decided ... (read more)