This is quite unserious speculation.

It's easy to create digital copies of models. Just copy their weights!

May 28?? Woop, woop, paper deadline has been extended by two months.

Or, perhaps, when they say LLMs, do they really mean deep learning? Then why not say deep learning directly?

Do they mean any AI system that maps into the language space at the output layer?

Then doesn't this disregard that you can have any architecture in the hidden layers, including search, tools, symbolic AI, etc, much of which ChatGPT already does?

Do they mean autoregressive LLMs, following LeCun's argument that errors accumulate exponentially, making them very hard to be correct?

Then what about diffusion-based LLMs? Can that lead to AGI?

Do they mean pretraining style language models that are solely trained to estimate the probability of word sequences on an unlabeled pretraining corpus?

If so, then is their argument in the spirit of Bender & Koller's Climbing towards NLU paper? If so, does that even apply to post-trained models?

but even a randomly initialized language model will accept a prompt and produce an answer. would that be AGI?

the core question is whether the reward can be modeled for RL. I think it's going to be rather simple using LLMs as judges...

So that implies he firmly believes to be doing what's best for the US and the world?

How do you reconcile that with the countless blatant falsehoods he propagates?

Great! Your initial post didn't show how nuanced your views are.

yes and that's fine

And why are these things mutually exclusive?

More generally, what do you think is Trump's motivation to be president? I struggle to find any motivation that could explain his ... unconventional way of doing politics.