Ari

An LLM fintuned on a book does not act like a human who read a book; it doesn't have consistent episodic memories about the experience, etc. What does the finetuning corpus that produces an LLM that acts like a human who read book look like?

There's one very real way that evolution is happening in LLMs: through the selection of attributes that are heritable and selected for in synthetic data. Have we isolated any of these yet?

LLMs don't perfectly remember the documents in their training set. But has anyone characterized how their reproductions differ from the original?

For a given LLM there must be certain data drawn from distribution D, such that if you finetune on them the LLM performs worse on D. It misgeneralizes, due to its priors, as we all do. Are there any interesting cases of this that don't feel totally adversarial and artificial?

Does the ICL-finetuning gap close with scale?