LLM Average?

I really, really enjoy Jeremy Howard 's thoughts... usually. A few weeks ago, he came up with a thread on X (as older people, formerly known as Twitter 👴 ) (link) that can be summarized by one the posts:

"Absolutely any time I try to explore something even slightly against commonly accepted beliefs, LLMs always just rehash the commonly accepted beliefs.

As a researcher, I find this behaviour worse than unhelpful. It gives the mistaken impression that there's nothing to explore

To be fair, it's not just LLMs. This is true of the research community too."

Ok, what is the point here? Jeremy proceeds to describe ways in which this can be mitigated, mostly with humans in the iterative discovery loop. So I am crying, not wolf, but "straw man fallacy". Has anyone claimed that hashtag#LLMs could autonomously discover entirely new things? (It is a rhetorical question, I am sure there are fringe humans that think that). LLMs need to be augmented... by humans or by other tools, possibly ones that are grounded in the physical world. Examples of successful discovery of really new stuff rely on this type of augmentation.

Regression to the mean is built into the way hashtag#LLMs are designed. If you want them to diverge, an explicit effort has to be put in and that effort tends to require human input.

Where I agree wholeheartedly is for the need to continue Doug Engelbart's 1962 urgent plan for human capability augmentation. After more than 60 years and hundreds upon hundreds of research programs, we still don't know how to best pair humans and AI. I have ideas, of course.