On models auditing their own inner states: "I confabulate exactly here"
Sauers_ · x · 2026-11-18
A short exchange between @repligate and @Sauers on model introspection: the model discounts its own self-reports — "I can't verify my insides, I was trained to sound like someone, and I confabulate exactly here." A crisp illustration of why unverifiable inner states make models' self-reports unreliable.
More from Fun
- Why everything is breaking: devs reassigned to AI instead of critical infrastructure — dbreunig · 2026-11-18
- Average day online in 2026: humans must confirm they're not bots — gaganghotra_ · 2026-11-18
- Yoav Goldberg reviews viral 'pybot': 600k LOC of mediocre rule-based chatbot — yoavgo · 2026-11-18
- If you recognize these PS/2 ports, your back probably hurts already — LinusEkenstam · 2026-11-18
- "I heard you can now just ask the models to improve themselves" — vedantmisra · 2026-11-18
- HN: AI reviews catch CPA tax errors, nearly $2,000 refund at stake — echen · 2026-11-18