[yawnxyz]

I think knowing what part of the knowledge base to delete — to get to adequately small reasoning model — is the hard part.

Doesn't "reasoning" rise from the knowledge? How much of a brain can you cut away before you affect the reasoning? When do you know what you've cut away, and what aspects did you miss / forget about?

We can probably train / fine-tune, w/ synthetic data, and we'll get reasonably close, but the "reasoning" will always hit rough patches, bc our training didn't include that kind of reasoning... and if we had to give it examples of every single kind of reasoning, then it can't move past all the already-established kinds of reasoning, so it's still pattern matching