You may be right, and I actually agree with you: I think that in this case the most likely outcome is that Fable becomes available again at some point, albeit possibly only to a restricted set of users within the US.
But I think my larger points stands: even if we do see Fable access again, this is the beginning of government restriction of LLMs and we are going to see more and more of it. In fact, I would be very surprised if we ever see an open weight model with Mythos capabilities. Chinese labs have been consistently releasing open models 6-12 months behind the frontier. In 6 months we may see them go dark.
Similarly, in the US I think we can expect more and more government restrictions on the strongest LLMs, in ways that may go beyond flimsy checks like uploading a valid US passport. It may not happen this year but I think it will happen eventually.
It still surprises me sometimes that LLMs are just available for _anyone_ to use. Isn't it odd that it turned out this way? When I grew up reading sci-fi I thought AI, if I ever saw it in my lifetime, would be something locked up behind the walls of big corporations and governments. But instead we have all been able to use it for an infinity of banal purposes for $100 a month. This is a strange situation but we have got used to it. But it may not continue that way.
The Chinese labs would only go dark if they believe they’ve surpassed the American labs, otherwise what benefit is there to them to refrain from sharing the models? Better to have all of their allies able to use the same models by making them public
" if I ever saw it in my lifetime, would be something locked up behind the walls of big corporations and governments."
Who knows, maybe the good stuff is locked up. If one of these corporations had something very special they may very well find it more profitable to enjoy the competitive advantage of using it for themselves than marketing it.
> It still surprises me sometimes that LLMs are just available for _anyone_ to use. Isn't it odd that it turned out this way?
I assume it's some of their best training data.
To me, it is obvious that what we are going to have is KYC/AML style compliance from US banking.
We already have the rails for automated customer identification from US banking.
I think there is a larger "AGI" category error with all this too that is akin to the old futurist idea of driving nuclear powered cars in the "future". The Ford Nucleon.
Nuclear power comes to us in a mix of electricity from the utility company but is far too dangerous for an individual to posses nuclear material for personal nuclear reactors.
An electric car does run on nuclear energy in some sense but not the way the Ford Nucleon was envisioned.
The error of the AI bubble is that we are pricing these companies with SaaS multiples when they are eventually going to be public utilities. There is really no other way to handle the dual use nature of anything close to "AGI".