Forced Yes/No Prompts Make Models Overconfident on Goldbach

A video testing top models on a Goldbach-conjecture halting problem showed they hedge when allowed to say 'I don't know,' but claim 95%+ confidence when forced into a yes/no answer.

2026-11-02 ~ 2026-11-02 · 2 related posts

• GPT and Claude bet 95%+ on an unsolved 1742 math problem — until 'I don't know' is allowed — Nir777 · 2026-11-02
• Forced yes/no prompts push GPT and Claude to 95%+ confidence on Goldbach's unproved halt test — Nir777 · 2026-11-02