---
id: "2026-10-19-erdos-szemeredi-sunflower-conjecture-gpt-6-astra-pro"
url: "https://postcutoff.com/e/2026-10-19-erdos-szemeredi-sunflower-conjecture-gpt-6-astra-pro/"
as_of: "2026-10-20T07:28:00+02:00"
date: "2026-10-19"
date_precision: day
category: science
importance: 5
confidence: medium
status: [Awaiting review]
verification: Lean-verified
sources: 2
editor: human editor
human_review: null
version: null
---

As of: 2026-10-20 07:28 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-19-erdos-szemeredi-sunflower-conjecture-gpt-6-astra-pro/

# Erdős–Szemerédi sunflower conjecture proved for every k ≥ 3

Full title: Erdős–Szemerédi sunflower conjecture proved for every k ≥ 3; the key vector bound (Theorem 2) was found by ChatGPT 6 Astra Pro (Bucić, He, Lee)

On 19 Oct 2026 Matija Bucić, Fankang He (University of Vienna) and Hyunwoo Lee (KAIST/IBS) posted arXiv 2610.26686, "Proof of the Erdős-Szemerédi sunflower conjecture". Theorem 1: for every k ≥ 3 and n ≥ 1, a family of subsets of [n] with no k-sunflower has at most 2^((1 − 1/(16k³))·n) members. Until now only k = 3 was known, via Croot–Lev–Pach, Ellenberg–Gijswijt and the Alon–Shpilka–Umans reduction. The proof goes through a new bound on vector sunflowers (Theorem 2). Its "Statement on the use of AI" says: "The proof of Theorem 2 was found by ChatGPT 6 Astra Pro". The authors say they checked everything, that the proof "has been verified in Lean 4.35" and that the formalization "will be made available on the Palomar repository". It is an unrefereed preprint, and the Lean files were not public on 20 Oct.

## Key facts
- Paper: 'Proof of the Erdős-Szemerédi sunflower conjecture', Matija Bucić, Fankang He (Faculty of Mathematics, University of Vienna), Hyunwoo Lee (KAIST and IBS ECOPRO), arXiv 2610.26686 [math.CO], submitted 19 Oct 2026
- Conjecture (abstract): 'In the 1970s, Erdős and Szemerédi conjectured that for any n ≥ k − 1 ≥ 2, any collection of subsets of [n] not containing a k-sunflower has size at most c_k^n for some c_k < 2.' The paper traces the question back to 1970
- Theorem 1: any subset of 2^[n] with no sunflower of k sets has size at most 2^((1 − 1/(16k³))n). Lower bound (Deuber, Erdős, Gunderson, Kostochka, Meyer): k-sunflower-free families of size 2^(n − O(n log log k / log k)), so the result is 'not far from best possible for large k'
- Earlier: Erdős–Szemerédi 2^(n − Ω(√n)); Alweiss–Lovett–Wu–Zhang (2019) and Rao give 2^(n − Ω(n/log n)). k = 3 followed from the 2016 cap-set bounds (Croot–Lev–Pach; Ellenberg–Gijswijt) through Alon–Shpilka–Umans; Naslund and Sawin gave base 3/2^(2/3) ≈ 1.89 using Tao's slice rank
- Method: Theorem 2 bounds the number V_k(D, n) of vectors in [D]^n without a k-sunflower by the polynomial method (dimension counting with low-degree factors and a Chernoff bound); a reduction (Section 4) transfers the vector bound to set families with D = 16k²
- AI statement, verbatim: 'The proof of Theorem 2 was found by ChatGPT 6 Astra Pro, following a short discussion, which started by asking it to generalize the k = 3 case argument in [45], which we needed for our original target of proving the induced variant of the Erdős–Sauer problem. The original argument it produced used the slice-rank terminology due to Tao (as used in [45]); we rewrote it in the polynomial method language … The authors have verified all the results in the paper and take full responsibility for the content. The proof has been verified in Lean 4.35; the verification will be made available on the Palomar repository.' ([45] is Naslund–Sawin)
- Concurrent work cited in the paper: another team proves that the set-sunflower relation on the Boolean cube has maximal slice rank, and 'They discuss extensive use of AI, including multi-agent setups'. The authors say that obstruction concerns set sunflowers, while their tester is for vector sunflowers, so it does not apply
- Consequences listed in the paper: the vector bound extends the Alon–Shpilka–Umans question (connected to fast matrix multiplication) and gives results in arithmetic property testing and on equal subset sums
- Reactions: no Hacker News thread or expert comment found by 20 Oct morning (web search; sweep HN list). A web search instead returned 2023 sources saying Naslund–Sawin had 'proved' the conjecture; that wording refers to the k = 3 case

## What happened

A sunflower is a family of sets whose pairwise intersections are all the same set. Erdős and Szemerédi asked how large a family
of subsets of an n-element set can be without containing k sets forming a sunflower, and conjectured an exponential saving over
the trivial 2^n. The k = 3 case became a corollary of the 2016 cap-set breakthrough. Bucić, He and Lee had asked ChatGPT 6 Astra
Pro to generalize Naslund and Sawin's k = 3 slice-rank argument for a different problem, the induced Erdős–Sauer problem. The model
produced a bound for vector sunflowers that works for all k, and with the authors' reduction this gives the full conjecture.

The authors rewrote the model's slice-rank argument in polynomial-method language. They say all results were verified, including
in Lean 4.35, but the formal proof had not been posted when the preprint appeared.

## Why it matters

This is one of the best-known problems in extremal set theory. It sits next to the Erdős–Rado sunflower conjecture, which stays
open, and is tied to cap sets and the group-theoretic approach to matrix multiplication. The authors credit the model with the
proof of the key theorem, not with exposition, which puts it among the strongest AI contributions to combinatorics so far. Until
the Lean files are released or experts review it, the claim stays pending.

## What is disputed or not yet verified

- Verification: Unrefereed arXiv preprint; authors checked all results; the stated Lean 4.35 proof is not public yet (promised for Palomar)

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 172 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 111 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 202 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 141 days after its cutoff

## Sources

1. [arXiv 2610.26686: Proof of the Erdős-Szemerédi sunflower conjecture (Bucić, He, Lee)](https://arxiv.org/abs/2610.26686) (arxiv.org, paper)
2. [Full text (PDF)](https://arxiv.org/pdf/2610.26686) (arxiv.org, paper)

## Changes

- 2026-10-20 (filed): Created (PDF read in full text)

## Related

- 2026-10-20: [AI-assisted results in the 20 Oct 2026 arXiv listing](https://postcutoff.com/e/2026-10-20-ai-assisted-results-october-20-listing/index.md)
- 2026-10-17: [Erdős–Hajnal conjecture proved for every path P_t](https://postcutoff.com/e/2026-10-17-erdos-hajnal-all-paths-nguyen-astra-deepmind/index.md)
- 2026-08-18: [Palomar launches: a registry of Lean-verified mathematics to curb misrepresented AI proof claims](https://postcutoff.com/e/2026-08-18-palomar-lean-registry/index.md)
- 2026-05-27: [Erdős–Szemerédi sum-product conjecture shown false over the reals](https://postcutoff.com/e/2026-05-27-sum-product-conjecture-false-over-reals/index.md)
- People: [Terence Tao](https://postcutoff.com/person/terence-tao/)