2026-02-26 15:02 UTC
@abucci @ngaylinn It's a very important distinction. Consider asking an LLM what my birthday is when that information is not in the training data.
A next token predictor will represent it's uncertainty by assigning all dates equal probability. If you then sample from this, you get hallucinations.
A model trained by RL will instead pick a very definite answer which (ideally) is "I don't know." Uncertainty in the semantics, but not in the distribution on actions.
Replies (1)
-
@pbloem@sigmoid.social 2026-02-26 15:05
@abucci @ngaylinn That doesn't make it an agent, but it allows for a clear distinction between a statistical prediction model, and one trained towards a specific behavior. Hence, more "agentic". If you don't want to start the discussion before we've resolved the philosophical issue of what an agent is, I don't know what to tell you. I don't think we can afford to wait.