原始内容
Marc Andreessen 🇺🇸 (@pmarca) 转发了 Przemek Chojecki | PC (@prz_chojecki) 的帖子：
19 Erdős problems I claimed with AI
Just to update the list for all the solution claims I've got with GPTs - 5.5 Pro solved the most, but each 5.2, 5.4, 5.6 solved something in the end.
3 New Claims with GPT-5.6 Sol Ultra in July: #421, #793, #415 (loose ends closed from previous attempt by GPT 5.4)
16 Previous Solution Claims: #258, #522, #603, #610, #750, #856, #858, #888, #896, #953, #956, #1092, #1133, #1148, #1151, #1190
5 almost/partial: #514 (finished with another person), #689 (finished with another person, bookkeeping is tricky), #741 (cleaning what's already proven from OpenAI/Deepmind), #906 (was completed by someone else first, a bit different), #1201 (partial, but close).
Most of the Solution Claims I've got with the help of GPT-5.5 Pro in the last two weeks of April 2026. What a marathon that was!
I'm writing solution claims not solutions, because not all of the solutions were verified - either formally or by other mathematicians. However each was run through an adversial LLM check at least twice and read by me.
Since May I wasn't as active pursuing these problems, but it makes sense to revisit some now with the release of GPT-5.6.
> **引用原帖 Przemek Chojecki | PC (@prz_chojecki):**
> 12 Erdos Problems I claimed with AI.
> They still need to be verified properly for the final score, though most look fine.
> Here's the full summary of my AI+math marathon:
> 12 Solution Claims: #522, #603, #610, #856, #888, #896, #953, #956, #1092, #1133, #1151, #1190
> 4 almost/partial: #514 (finished with another person), #741 (cleaning what's already proven from OpenAI/Deepmind), #906 (was completed by someone else first, a bit different), #1201 (partial, but close)
> On top of that I had 3 solutions verified before: #258, #858, #1148
> Also I've got MANY partial results in other problems, that I haven't posted. It was usually a small increment like optimizing a constant in an estimate that didn't really give any new math or insight into a full solution.
> One thing to add is I'm pretty humbled by the whole experience.
> Was it really me that solved them?
> Except for maybe 3-4 problems where my LLM guidance was very explicit with methods to be used, most of these didn't require much guidance besides crafting a good prompt with references, potential approaches, computations. Couple of them were proved basically in one prompt (prompt -> solution).
> A general remark is that except for maybe 2 problems, I've got solutions in max 4 prompts. GPT-5.5 Pro is way more efficient than GPT-5.4 Pro, but also if it doesn't hit quick, it makes little sense to push it.
> https://x.com/prz_chojecki/status/2050160288108892407