[ExI] A sad day for the US Weather Bureau

[Ben Zaiboc]

On 05/10/2026 03:23, Keith Henson wrote:
> John, I largely agree with your assessment of the state of AIs.
>
> But I somewhat wonder if they are hitting the limits of knowledge? They know virtually everything humans know, but they simply cannot know beyond that. Any thoughts on where this will go?


Yes.

I think we probably are near to hitting the limits of what stochastic parrots can do.

In order to go beyond existing knowledge, they will need different architectures. We need to go beyond LLMs if we want proper AIs. I think LLMs are just one, fairly small, part of the picture, and we need to add other ways of processing information to get fully human-equivalent intelligence.

LLMs are extremely good at collecting and sifting through existing knowledge (and making nonsense up), but useless at actually understanding things (which is probably why they invent nonsense). More and more people seem to be realising this now.

Unless the AI researchers can come up with something new that achieves the same thing, I reckon they will need some version of embodiment, plus the kinds of neural architecture that we have, or things that perform the same functions. We have a fair idea of the minimum requirements for producing the kinds of minds that we have, from neuroscience. I don't know that anyone has made a serious, large-scale effort to replicate that. I suspect not. So much money can be made from LLMs that all the focus currently seems to be on them.
-- 
Ben