Why logits can simulate a jev-style classifier, explained

michellechen · x · 2026-09-19

The author explains the mechanism: while System One models' internals are unknown, their interface can be approximated with any LLM. Models emit logits for every vocabulary token; softmax turns them into probabilities and logprobs express them in log space — carrying more information than the finally selected tokens. With the right inputs and prompting, these scores yield a probability distribution that simulates a jev-style classification model.

Related event: Developers Turn LLM Logprobs into Probabilistic Decision Outputs(2 posts)→

Original post →

More from coding & agent

coding & agent channel →