A researcher says GPT-5.6 Sol helped crack 6 Erdős problems in 5 days
rohanpaul_ai · x · 2026-07-23
One researcher says GPT-5.6 Sol helped solve 6 open Erdős problems in 5 days
A quoted post claims a single researcher solved 6 open Erdős problems in 5 days using GPT-5.6 Sol.
He reportedly tried 13 problems in total, for a success rate of about 46%. The key takeaway is not just model output, but the prompting style:
- The prompt was written like a contract, not a question.
- It restated the problem and defined what a valid proof must establish.
- It listed weaker results that would not count, to block near-misses.
- It also instructed the model to explore multiple avenues in parallel, keep incompatible hypotheses alive, and actively search for counterexamples to its own lemmas.
The post frames this as a lesson in how to steer a model toward proof search rather than a simple answer.
Related event: GPT-5.6 Sol Reportedly Solves Six Erdős Problems in Five Days(2 posts)→
More from AGI Musings
- Mat Dryhurst argues the online creator market now pits everyone against everyone — matdryhurst · 2026-07-23
- A math joke turns LLM criticism into a threat to dump unreadable theory on arXiv — airkatakana · 2026-07-23
- X blocking is the gentler option, but the post demands useful AGI discourse — suchenzang · 2026-07-23
- AI may change mathematics the way AlphaGo changed professional Go — Kangwook_Lee · 2026-07-23
- Open-source AI is improving faster while costs fall, the post argues — bindureddy · 2026-07-23
- OpenAI’s silence on GPT-OSS is making the world more dangerous, reply says — _aidan_clark_ · 2026-07-23