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?
Who's Who of AI
Ari
Assistant Professor @ UChicago CS & DSI UChicao
Leading Conceptualization Lab http://conceptualization.ai
Minting new vocabulary to conceptualize generative models.
Their own posts
Recent commentary
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?
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?
Their network
In Ari's orbit
Center = Ari. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.