First 48 hours of GPT-6.2 Astra
A high-school student posts 25 claimed results, half Lean-checked
Awaiting review
The takeaway
Within two days of GPT-6.2 Astra’s release (17 Nov 2026), people outside academic maths began posting results made with it.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Lean-verified
- Importance
- 3 of 5
- Last verified
- 19 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Repo github.com/1337isnot1337/solved-conjectures (created 18 Nov 2026), ‘Curated by Cinna Davis and GPT-6.2 Astra’; 25 results listed on 19 Nov: 12 ‘Formalized stated result’ in Lean, 12 partial formalizations, 1 not formalized; README: ‘These notes are unrefereed. Computational checks and Lean verification are distinct from peer review and priority claims.’
- Results include Erdős Problem 607 (line-size sets, partial Lean), Pham’s Bulgarian-solitaire denominator degree (2022) and a generalized Bulgarian solitaire limit shape (2020), Batista’s Egyptian fractions, Haralambis’s three-difference conjecture, Carraher et al.’s distance-graph conjecture 38 (2014; their conjecture 39 refuted), Natalchenko–Sagdeev nonmixed-gap packing (2024), Spivey’s three-player War, several Loneliness Spectrum cases, and Graffiti.pc Conjecture 318 (contributed by @JesterInABarrel, ‘in 12m 24s’); partial results on Frankl’s union-closed sets conjecture, the Lonely Runner Conjecture and Hadamard order 2396
- Davis on X (19 Nov): ‘Just solved 2 more conjectures with 6.2 Astra, Lean-verified … it’s a turkey shoot. Total solved: 25’; earlier: two conjectures ‘in 30 minutes’
- Rohan Arun (18 Nov): ‘we discovered an improvement on e-means using 6.2 Astra within 10 minutes’, repairing two counterexamples published the day before (seed-loss check; pseudo-count floor with threshold τ* = 0.1927; failures 49% → 0% over 20,000 runs); paper on Zenodo, code on GitHub
- Other GPT-6.2 Astra maths in the first days: Xi Chen’s Hexagon claim on Bourgain’s cubic NLS growth question (2026-11-18 entry); Epoch’s FrontierMath Erdős score 7 of 68 unsolved problems at launch; on 18–19 Nov also Jeffrey Shallit’s additive-square-free word on six integers, ‘almost entirely written by GPT 6.2 Astra’ with a Lean check (own entry), Chen, Mao and Zhang’s faster bounded integer programming obtained ‘with almost no human input’ (arXiv 2611.24386), Sunay Venkat’s two Erdős-related deposits and about a dozen irrationality deposits by twinheterodox with Codex GPT-6.2-Astra max (see the 2026-11-19 listing)
- Context: VentureBeat headlined power users’ reviews ‘essentially unusable’ and Theo Browne’s ‘Why I hate GPT-6.2 Astra’ had 1.9M views in the same 48 hours
- Repo state on 19 Nov evening (README, last push 20:09 UTC): still 25 results, now 15 fully Lean-formalized (✅), 9 partly (🟡) and 1 not (⬜); Erdős Problem 607 moved to fully formalized. Results on the list besides those above include random lonely runners on a hyperplane, Kravitz’s Loneliness Spectrum Conjecture for five to eight speeds (partial Lean), equality cases for eight to fourteen lonely runners, Liu–Zhu packing, two-difference Motzkin density, rectangle partition asymptotics, two CUT-game results and a counterexample to Ethan Li’s coefficient-sign prediction for chromatic symmetric functions (a 31-vertex tree; it ‘does not refute’ the Dahlberg–She–van Willigenburg conjecture). ‘Partial results on open problems’ now lists Frankl’s union-closed sets, the Lonely Runner Conjecture, Erdős–Gyárfás power-of-two cycles and Hadamard order 2396
- First outside contribution: pull request #35 by aadrianvv (19 Nov) completing the Lean formalization of the Li counterexample; about 40 issues opened by Davis track individual Lean targets. Repo: 7 stars, 1 fork
Show 2 more
- Reach: Davis’s ‘Total solved: 25’ post had about 2.7k views and the 8-speed Loneliness Spectrum post (19 Nov, 13:24 UTC) about 1.2k by the evening; Arun’s e-means post about 14.6k (fxtwitter)
- Reactions: no comment found on the Erdős Problems page for #607 (0 comments, 0 proof claims on 19 Nov; #607 itself was settled by Szemerédi–Trotter, so the repo’s result is a refinement), on HN, Mathstodon or blogs
What happened
GPT-6.2 Astra was released on 17 November. By the next day independent users were posting results produced with it. The largest batch is a GitHub repo by Cinna Davis, a high-school student who earlier contributed to Rohan Arun’s integer-multiplication project. It grew from two results on 18 November to 25 on 19 November, each with a written proof and a Lean status, and with one result contributed by another user. The same evening Arun posted a repaired version of the e-means clustering algorithm, answering two counterexamples published a day earlier.
Why it matters
It shows how fast a new frontier model turns into a stream of formal-looking results from outside academia. The problems are mostly small and specialised, and none of the claims has been refereed; the Lean column is the useful signal, and only about half the results are fully formalized as stated.
What is disputed or not yet verified
| Verification | Unrefereed; 12 results Lean-formalized as stated, others partial or computational only |
|---|
Sources
9 sources from 4 sites. Numbers match the chips in the text.
9 sources: 3 primary, 6 reactions
Primary
- Cinna Davis: solved-conjectures repo (GitHub)github.com, code
- Rohan Arun: Stabilizing Exponential Moving Average k-Means (Zenodo)doi.org, paper
- Rohan Arun: emeans-plus code (GitHub)github.com, code
Reactions
- Cinna Davis on X: ‘Total solved: 25’ (19 Nov)x.com, discussion
- Cinna Davis on X: first two conjectures ‘in 30 minutes’x.com, discussion
- Rohan Arun on X: e-means repaired with GPT-6.2 Astrax.com, discussion
- @JesterInABarrel: Graffiti.pc Conjecture 318 in 12m 24sx.com, discussion
- Cinna Davis on X: Kravitz’s Loneliness Spectrum Conjecture for 8 speeds (19 Nov)x.com, discussion
- Erdős Problems #607erdosproblems.com, discussion
Changes
- Repo re-checked on 19 Nov evening (new results list, Lean counts, first outside PR, post reach); reaction search (science agent)
- Filed (owner’s request) from the repo README, the authors’ X posts and the Zenodo record