[visarga]

> There's a stark difference between coming up with an an hypothesis and just roll with it and scrutinising a hypothesis by conducting experiments and tests.

Like the difference between LLM hallucinations and truth. At pure language level we can come up with any explanation (LLMs can generate any hallucination), it is validation in the real world that makes it true. Neither brains or LLMs can skip the validation step, we are not THAT smart.

[rafaelmn]

That's the part that makes current LLMs fundamentally flawed as AGI agents - unless you incorporate experience based knowledge validation/learning - the book knowledge you have is useless. Most of published research out there isn't reproducible, and that's the highest standard we have for published knowledge.

[visarga]

I mean, there is plenty of "experience" codified in the raw internet scrape they used. But not enough for LLMs, and not in-domain. So it won't contain the kind of errors LLMs make and corrective feedback.