Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →OpenAI personnel claimed in October 2025 that GPT-5 had found solutions to 10 “previously unsolved” Erdős problems and made progress on 11 more. The claim was later corrected: GPT-5 had helped researchers locate solutions already published in mathematical literature, including papers not reflected in a public database. The posts were deleted or revised after criticism, but available reporting does not establish that anyone was forced to delete them.
What OpenAI originally claimed
On October 17, 2025, OpenAI vice president Kevin Weil posted that GPT-5 had “found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others.” OpenAI researcher Sebastien Bubeck amplified the claim, describing researchers finding the solutions with GPT-5’s help over a weekend. Related work involved mathematicians Mark Sellke and Mehtaab Sawhney, who used thousands of GPT-5 queries while examining problems listed in the Erdős database.
As an Amazon Associate I earn from qualifying purchases.
If the wording had meant that GPT-5 independently produced new proofs, it would have represented a major advance in automated mathematics. That implication is what made the episode newsworthy—and misleading.
TechCrunch’s account reported the original wording, the subsequent criticism and the clarifications.
#1 Best Overall
The crucial meaning of “open”
The problems came from the Erdős Problems website, a database of questions associated with Hungarian mathematician Paul Erdős. The problems span number theory, combinatorics, geometry and other areas, and they vary greatly in difficulty. “Ten Erdős problems” therefore does not automatically mean ten landmark breakthroughs.
Thomas Bloom, a University of Manchester mathematician who maintains the database, explained that “open” had a narrower meaning than many readers assumed. A problem could be marked open because Bloom had not found or seen a paper containing a solution. The label was not a definitive declaration that no mathematician anywhere had solved it.
Bloom said GPT-5 had located references to existing solutions and called the OpenAI framing a “dramatic misrepresentation.” In other words, the database entries needed updating, but the model had not demonstrated ten new proofs.
Free tools Windows power users keep installed
One-click scans. No signup required.
Retrieving a proof is not creating one
| Claim or activity | What the evidence supports |
|---|---|
| “GPT-5 solved 10 previously unsolved problems” | This was the interpretation encouraged by the original posts, not an established result. |
| GPT-5 found relevant mathematical answers | Researchers used the model to surface papers and references containing solutions that were missing from the database’s records. |
| Ten new mathematical discoveries | Not demonstrated in the October episode. |
| Database corrections | Identifying overlooked literature can legitimately change an entry from open to solved. |
“Found” is ambiguous in this context. It can mean retrieved a paper, recognized that a paper answers a question, reconstructed an argument or derived a genuinely new proof. Mathematics normally treats a problem as solved only when a valid argument establishes the result. The October posts did not distinguish those meanings.
How the posts were walked back
Bloom’s correction
Bloom’s explanation supplied the technical correction: the model connected database entries with solutions already present in the literature. That distinction shifted the story from autonomous discovery to literature search and verification.
Weil’s acknowledgment
Weil later said he had misunderstood Sellke’s original post and acknowledged that his own wording was wrong. He indicated that he would delete the post.
Bubeck’s apology
Bubeck deleted his related post and apologized for the misleading phrasing. He clarified that the solutions identified were already in the literature.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Public criticism
Google DeepMind chief executive Demis Hassabis called the episode “embarrassing,” while Meta chief AI scientist Yann LeCun mocked the claim. Their comments are reactions, not independent proof of what the model did; Bloom’s technical clarification is the central evidence.
The sequence—claim, correction, criticism and deletion—is documented in Techmeme’s October 20 timeline. The available reporting supports “deleted after public criticism” or “walked back,” not the stronger assertion that OpenAI management forced the deletions.
What GPT-5 actually contributed
The defensible description is that GPT-5 helped uncover previously overlooked or poorly connected mathematical literature. Researchers queried the model repeatedly, followed references and then used human mathematical judgment to determine whether the cited papers really answered the listed questions.
- Useful: finding obscure papers, linking differently worded results and exposing gaps in a manually maintained database.
- Human-dependent: deciding whether a citation contains a complete solution, checking the proof and interpreting its relationship to the problem statement.
- Not shown: ten independently generated proofs of problems that had remained unsolved worldwide.
This is not an all-or-nothing result. Scholarly discovery can be difficult when results are old, fragmented across fields or indexed under unfamiliar terminology. A system that reliably finds such material could save researchers substantial time. But updating a database after locating a paper is different from proving a new theorem.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why the wording mattered
AI companies have strong incentives to describe retrieval, assistance or partial progress in the language of autonomous discovery. “Found solutions” sounds like a scientific breakthrough; “located papers containing solutions” describes a research-assistance task. Without a clear account of the model’s prompts, search process, human filtering and proof verification, readers cannot tell which achievement is being claimed.
The number 10 also needs context. It counts database entries, not ten uniformly difficult or historically important problems. Some may have had obscure or incomplete solutions, making the discovery useful without making the underlying mathematics new.
Later GPT-5 research was a separate story
In November 2025, an OpenAI-affiliated paper, Early science acceleration experiments with GPT-5, described experiments in mathematics, physics, astronomy, computer science, biology and materials science. It reported four new mathematical results verified by the human authors and emphasized that expert input remained important. The paper is available at arXiv.
That later work offers a more carefully documented picture of GPT-5-assisted research, but it does not retroactively make the deleted October posts accurate. The two episodes involved different claims and different standards of documentation.
How to evaluate future AI “breakthrough” claims
- Ask what “found” means. Was the system retrieving a publication, reproducing a known proof or producing a new result?
- Check the status label. “Open” in a particular database may mean “not yet recorded as solved,” not “no solution exists.”
- Look for a verifiable artifact. A paper, proof, dataset or reproducible method should be available for inspection.
- Separate model output from expert work. Prompting, search design, filtering and mathematical validation may account for much of the result.
- Look for independent verification. A company announcement is not the same as review or confirmation by outside mathematicians.
Bottom line
GPT-5 did not demonstrably solve ten previously unsolved Erdős problems in the episode that triggered the controversy. It helped researchers locate existing solutions that had not been reflected in a public database. That is a potentially valuable literature-discovery capability, but it is not the same as an autonomous mathematical breakthrough—and “forced to delete” goes beyond what the available evidence shows.
Frequently Asked Questions
Were the Erdős problems actually unsolved?
They were listed as open in Thomas Bloom’s database, which meant he had not found or seen papers containing solutions. That status did not guarantee that no solution existed in the wider literature.
Did GPT-5 produce any new mathematics later?
A separate November 2025 OpenAI-affiliated paper reported four human-verified new mathematical results from GPT-5-assisted experiments. Those results do not validate the original October claim.
Who deleted the posts?
Kevin Weil said he had misunderstood the underlying post and would delete his wording; Sebastien Bubeck deleted his related post and apologized. Public reporting does not establish that OpenAI ordered or forced the deletions.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
