Post #1490139
2026-02-26 14:55 UTC
@pbloem@sigmoid.social @ngaylinn@tech.lgbt Switching from more than one possible output to single output does not make a fundamental difference. Tacking a sampling algorithm, no matter how complicated, to the end of a selection process does not confer a system agency in any meaningful sense of that word. That's a small-world operation, and meaningful agency manifests in a large world.
Leaving aside the intellectual vacuity of behaviorism, of course, as well as impossibility of judging "did well" or "did poorly" in a coding environment.
Replies (1)
-
@pbloem@sigmoid.social 2026-02-26 15:02
@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.