That's what I am thinking when commenters compare LLMs with humans. Humans got the experimental loop going through the real world, pure LLMs are brains in vats. They need experimental confirmation to become truly creative.

But on the other hand are the commenters who think AGI to ASI is going to take hours. When is there time to experiment anything? Even ASI needs real world confirmation to do research.

The claim is that an AGI can instruct humans to do the experimentation (knowingly or not).

AGI to ASI might take hours if all that's needed is more compute. Such an intelligence might exponentially spread across the internet. Though I don't think it's the most likely scenario.

Well put. I'd add that theories can produce predictions, and if the predictions can be proved experimentally, all the better for the theory.

Or perhaps even stronger - theories _should_ produce predictions.