[kbrkbr]

I have also a hard time understanding how AGI will magically appear.

LLMs have their name for a reason: they model human language (output given an input) from human text (and other artifacts).

And now the idea seems to be that when we do more of it, or make it even larger, it will stop to be a model of human language generation? Or that human language generation is all there is to AGI?

I wish someone could explain the claim to me...

[Gerardo1]

Because the first couple major iterations looked like exponential improvements, and, because VC/private money is stupid, they assumed the trend must continue on the same curve.

And because there's something in the human mind that has a very strong reaction to being talked to, and because LLMs are specifically good at mimicking plausible human speech patterns, chatGPT really, really hooked a lot of people (including said VC/private money people).

[hackinthebochs]

LLMs aren't language models, but are a general purpose computing paradigm. LLMs are circuit builders, the converged parameters define pathways through the architecture that pick out specific programs. Or as Karpathy puts it, LLMs are a differentiable computer[1]. Training LLMs discovers programs that well reproduce the input sequence. Roughly the same architecture can generate passable images, music, or even video.

It's not that language generation is all there is to AGI, but that to sufficiently model text that is about the wide range of human experiences, we need to model those experiences. LLMs model the world to varying degrees, and perhaps in the limit of unbounded training data, they can model the human's perspective in it as well.

[1] https://x.com/karpathy/status/1582807367988654081