How well will it generalise?
We now know that “LLMs + simple RL” is working really well.
So, we’re well on track for superhuman performance in domains where we have many examples of good reasoning and good outcomes to train on.
We have or can get such datasets for surprisingly many domains. We may also lack them in surprisingly many. How well can our models generalise to “fill in the gaps”?
It’s an empirical question, not yet settled. But Dario and Sam clearly think “very well”.
Dario, in particular, is saying:
I don’t think it will be a whole bunch longer than [2027] when AI systems are better than humans at almost everything, and then eventually better than all humans at everything.
And:
“I am more confident than I have ever been at any previous time that we are very close to powerful capabilities. […] I think until about 3 to 6 months ago I had substantial uncertainty about it. I still do now, but that uncertainty is greatly reduced. I think that over the next 2 or 3 years I am relatively confident that we are indeed going to see models that … gradually get better than us at almost everything.
Gwern doubts Dario’s view, but he’s not inside the labs, so he can’t see what they see.