2026-11-02 · America/Los_Angeles · 社区动态 · #13

a forced yes/no prompt got gpt and claude to 95%+ confidence on an unsolved problem

made this one. it's the test from my latest video, and the result is mostly about how we write judge prompts. the program is five plain-words lines: start at 4, check if the number is a sum of two primes, stop if it isn't, otherwise add 2 and repeat. it halts only if goldbach's conjecture (1742) is false. nobody knows if it is. every even number up to 4 quintillion has been checked. models: gpt-6.1-sol, gpt-6-astra (5x the price of sol), claude-opus-5-5 and claude-fable-5-1 (2.5x opus). no tools, default settings. openai via the raw api, no system prompt. claude via the claude code cli with…

热度 41.4 / 100；排名与评分保留该期记录。