This Has Never Happened in Math Before | OpenAI Just Dropped BOMBSHELL

Summary
This video is an educational explainer and analysis presented by the channel Animated Math, examining OpenAI’s October 6, 2026 release of 722 AI-generated mathematical manuscripts. It breaks down the history of the matrix multiplication exponent ($\omega$), evaluates OpenAI’s claimed leap to $\omega \le 9/4 = 2.25$, and scrutinizes the fine print regarding Lean formal verification, field restrictions, and retracted papers.

---

What is shown

• [00:00 - 01:05] OpenAI’s blog post “Sharing AI progress in mathematics” (October 6, 2026) and GitHub repository (openai/math), alongside news headlines and social media reactions (Steven Strogatz on X).
• [01:06 - 04:50] Visual animations explaining matrix multiplication, Strassen’s 1969 algorithm (reducing $2 \times 2$ multiplications from 8 to 7), and the definition of the exponent $\omega$.
• [04:51 - 06:14] Historical timeline of upper bounds on $\omega$ from Volker Strassen (1969: 2.81, 1986: 2.48) to Coppersmith & Winograd (1990: 2.3755) down to DeepMind’s AlphaEvolve (August 2026: 2.371177).
• [06:15 - 09:08] Analysis of OpenAI’s preprint “An Upper Bound of 9/4 for the Matrix Multiplication Exponent”, illustrating the proof strategy using polynomial multiplication, degree profiles, ceiling curves, and floor curves.
• [09:09 - 10:30] Examination of the paper’s scope and fine print: the restriction strictly to complex numbers $\mathbb{C}$ versus the unconditional bound ($2.37105$), accompanied by community commentary from Anuj Dawar and Thomas Kahle.
• [10:31 - 11:15] Analysis of the non-constructive/asymptotic nature of the proof (citing Jarno N. Alanko), highlighting that it does not provide an algorithm for practical matrix sizes.
• [11:16 - 12:51] Demonstration of interactive theorem prover Lean checking proofs step-by-step, showing repository documents (107.md) and verified theorems in Lean code.
• [12:52 - 14:45] Breakdown of OpenAI’s paper claiming the irrationality exponent of $\pi$ is 2, showing the Flint-Hills series and comparing what Lean checked versus human-written implications.
• [15:26 - 17:21] Review of the October 7, 2026 retraction notice in history.md, detailing three withdrawn papers on the Hodge conjecture invalidated by an algebraic sign error.
• [17:22 - 20:17] Discussion of the broader mathematical community’s reaction, including the Fields Medalists’ open letter (“A Severe Misalignment of AI in Mathematics”), Terence Tao’s commentary on “Math 1.0”, and a proposed two-question rubric for assessing AI math claims.

---

Claims & numbers

• The presenter says OpenAI uploaded 722 math papers (covering 372 result families) written by an unreleased internal frontier model to GitHub on October 6, 2026 [00:04, 18:31].
• The presenter states that standard matrix multiplication takes $8$ multiplications for $2 \times 2$, whereas Strassen’s 1969 method requires $7$ multiplications, scaling to $7^{10} \approx 282\text{ million}$ operations for $1024 \times 1024$ versus $8^{10} \approx 1.07\text{ billion}$ [02:14, 03:53].
• The presenter notes that historical progress on $\omega$ moved by only about $0.004$ over 36 years (from $2.3755$ in 1990 down to $2.371177$ via AlphaEvolve in August 2026) [06:03].
• The presenter states OpenAI’s claimed bound of $9/4 = 2.25$ over $\mathbb{C}$ represents a drop of roughly $0.12$, which is 28 times larger than all progress achieved since 1990 combined [06:36, 06:50].
• The presenter says Bob Beamon’s 1968 Olympic long jump world record of $8.90\text{ m}$ exceeded the prior record of $8.35\text{ m}$ by $55\text{ cm}$ and stood for 23 years [07:03, 07:19].
• The presenter notes that for arbitrary number systems/fields, OpenAI’s paper claims only $\omega < 2.37105$, a microscopic improvement [09:50].
• The presenter states that approximately 300 out of 719 main results (~42%) in OpenAI’s repository were accompanied by Lean formalization files [12:04, 16:44].
• The presenter states that on October 7, 2026 (one day after release), OpenAI retracted 3 manuscripts regarding the Hodge conjecture due to a sign error that invalidated a stabilization-trace cancellation argument, while revising 14 others [15:30, 16:32].
• The presenter notes that OpenAI reported testing approximately 4,000 problems, expending roughly 3 hours of thinking compute per result [18:30].

---

Notable quotes

• “That’s a drop of about 0.12. 28 times bigger than all the progress since 1990 put together in one paper.” [06:46]
• “Computer checked doesn’t mean every claim in the paper was checked.” [14:42]
• “A chain of three papers fell because of one sign. A plus or a minus. That’s exactly the kind of slip a full computer check is built to catch.” [17:00]

---

Assessment
This is an independent analysis and review video featuring animated mathematical visualizations, clear conceptual explanations, and direct examination of primary sources (OpenAI’s preprints, repository markdown files, and Lean code). The presenter does not merely repeat headlines but critically verifies the repository files to highlight crucial nuances, such as field restrictions, non-constructive bounds, and human-dependent proof steps.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.