NYU Professor Accuses OpenAI of Scooping Navier-Stokes Math Proof

Fireship

Corporate AI labs using massive compute swarms to front-run academic rumors threaten to destroy open scientific collaboration by forcing researchers to conceal early discoveries.

AI-assisted summary of OpenAI's biggest math breakthrough is getting ugly... — original source. Check the original for full context.

Source published: 2026-09-11

Key ideas

NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge made a breakthrough by breaking the Euler equations using AI tools Claude Code and Codex. (0:00)
OpenAI reportedly deployed 10,000 AI agents and $20 million in compute to solve the Millennium Prize Navier-Stokes problem, matching the novel technique developed by Buckmaster and Alpöge. (2:41)
A bitter attribution dispute erupted over allegations that OpenAI leveraged Buckmaster's private Codex sessions and demanded the exclusion of Alpöge due to his Anthropic affiliation. (2:41)
Fields Medalist Terence Tao warned that corporate AI swarms front-running academic rumors could force researchers into secrecy and dismantle open science. (4:25)

Detailed notes

## Background This video was published on YouTube by the channel Fireship on September 11, 2026. According to the page description, it covers OpenAI's claim of cracking a 90-year-old math problem and the ensuing dispute with an NYU professor who was unhappy about the situation.

## TL;DR The video details the controversy surrounding OpenAI's claim that its agent swarm solved the historic Navier-Stokes Millennium Prize Problem. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had made a breakthrough on the related Euler equations using AI coding tools, only for OpenAI to allegedly deploy 10,000 agents and $20 million in compute using the exact same novel approach after hearing rumors of their work. Following an escalating confrontation between Buckmaster and OpenAI's Sébastien Bubeck over credit and private user data, both parties released competing papers and statements. The fallout culminated in Fields Medalist Terence Tao warning that corporate AI swarms front-running scientific research could destroy centuries of academic openness.

## Atmosphere & style The presentation carries Fireship's trademark satirical, fast-paced tech commentary, weaving deadpan internet humor and absurd pop-culture references through an otherwise serious academic conflict. Beneath the comedic editing and breezy delivery lies a tense, cynical narrative arc about massive corporate compute muscle eclipsing traditional human scholarship. It feels entertaining yet unsettling, capturing the chaotic anxiety of researchers who suddenly find themselves outpaced and pressured by aggressive AI labs.

## Bulletin - [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge made a breakthrough by breaking the Euler equations using AI tools Claude Code and Codex. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) OpenAI reportedly deployed 10,000 AI agents and $20 million in compute to solve the Millennium Prize Navier-Stokes problem, matching the novel technique developed by Buckmaster and Alpöge. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) A bitter attribution dispute erupted over allegations that OpenAI leveraged Buckmaster's private Codex sessions and demanded the exclusion of Alpöge due to his Anthropic affiliation. - [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) Fields Medalist Terence Tao warned that corporate AI swarms front-running academic rumors could force researchers into secrecy and dismantle open science.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) The Navier-Stokes equations, foundational to fluid dynamics for nearly two centuries, have long lacked a formal proof regarding whether singularities or mathematical breakdowns can occur. - [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) The Clay Mathematics Institute established a $1 million Millennium Prize bounty 26 years ago to determine whether Navier-Stokes equations always remain smooth or can fail. - [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) Tristan Buckmaster and Levent Alpöge achieved a critical breakthrough on August 15th by breaking the Euler equations, a simplified inviscid version of Navier-Stokes. - [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) The researchers experienced a massive productivity boost by incorporating AI tooling, specifically Claude Code and OpenAI's Codex, into their pure math workflow. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) OpenAI reportedly mobilized 10,000 agents and around $20 million in compute power to break Navier-Stokes using the identical long-shot method Buckmaster and Alpöge spent a year developing. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) OpenAI researcher Sébastien Bubeck acknowledged that their mathematical effort began only days prior, triggered by circulating rumors of Buckmaster's initial breakthrough. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) During confrontations with Buckmaster, OpenAI reportedly evaded whether Buckmaster's private research sessions in Codex had been incorporated into training datasets. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) Buckmaster claimed OpenAI gave him an ultimatum to either take second billing or join as an author on the condition that Alpöge was removed due to his job at Anthropic. - [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) OpenAI and Sam Altman strongly rejected Buckmaster's claims, asserting that no user data was accessed, researchers acted with integrity, and no demands were made to exclude Alpöge. - [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) Both parties released competing papers on the same Tuesday, with Buckmaster and Alpöge including an exhaustive four-page statement detailing the conflict. - [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) Terence Tao emphasized that if corporate AI labs aggressively scoop academic discoveries based on early rumors, the open-science culture underpinning centuries of progress will collapse.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) The Navier-Stokes Millennium Problem and Fluid Dynamics

The episode opens with the narrator remarking that while he was busy losing all his money on "the future of finance," OpenAI announced it had allegedly conquered one of the most stubborn challenges in mathematics. The target is the Navier-Stokes equations, first conceived in the 1850s—an era when computing was virtually nonexistent. Originally sketched out by French engineer Claude-Louis Navier and subsequently refined a couple of decades later by Irish mathematician George Gabriel Stokes, the equations function essentially as Newton’s second law ($F = ma$) applied to fluid dynamics. By taking a fluid's existing state—its velocity and surrounding pressure—the equations forecast how that liquid or gas will behave a moment later. For nearly two centuries, engineers and scientists have depended on Navier-Stokes to model hurricane trajectories, design aerodynamic airplane wings, and simulate blood circulation inside our "meatbag bodies."

Despite their immense real-world utility, mathematicians refuse to accept equations on blind faith without rigorous proof. The fundamental unresolved question is deceivingly simple: can the Navier-Stokes equations ever break or produce a singularity? The problem proved so notoriously intractable that 26 years ago, the Clay Mathematics Institute designated it one of the seven Millennium Prize Problems, attaching a $1 million bounty for anyone who could either prove the equations remain smooth indefinitely or demonstrate a single scenario where they break down.

Enter Tristan Buckmaster, an NYU mathematics professor who received the 2019 Clay Research Award for his breakthrough contributions to these exact equations. Last year, Buckmaster joined forces with Levent Alpöge, a mathematician employed at Anthropic who pursues century-old fluid dynamics challenges in his spare time. Progress was initially slow and steady until mid-August, when the team leaned heavily into AI tooling—leveraging Claude Code and Codex—and saw their productivity skyrocket. Their unconventional approach proved effective on August 15th, when they successfully caused the Euler equations (a simplified, inviscid formulation of Navier-Stokes) to fall apart. While not yet the full million-dollar prize, it stood as the closest anyone had ever come to breaking the real equations.

```mermaid flowchart TD accTitle: Navier-Stokes Dispute Flow accDescr: Diagram tracing the research progress of Buckmaster and Alpöge to OpenAI's agent swarm and the resulting dispute.

subgraph Research["Independent Research (NYU & Anthropic)"] A["Buckmaster & Alpöge collaborate"] --> B["Use AI tools: Claude Code & Codex"] B --> C["August 15: Euler equations break"] end

subgraph OpenAI_Action["OpenAI Response"] D["Hear research rumors"] --> E["Deploy 10,000 agents + $20M compute"] E --> F["Claim Navier-Stokes breakthrough using identical novel technique"] end

subgraph Escalation["The Confrontation"] C -.-> D F --> G["Sept 3 / Sunday Call: Bubeck vs. Buckmaster"] G --> H["Dispute over training data & paper attribution"] H --> I["Tuesday: Dual paper releases & public fallout"] end ```

[02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) OpenAI's Agent Swarm and the Sunday Confrontation Call

Shortly after the Euler breakthrough, OpenAI displayed a sudden and intense appetite for pure mathematics. In late August, while training a new model, OpenAI researchers fed it every unsolved Millennium Prize Problem simultaneously. Backed by 10,000 agents and roughly $20 million in compute power, OpenAI claimed to have successfully broken Navier-Stokes. Strikingly, OpenAI's model had arrived at the solution via the exact novel, long-shot method that Buckmaster and Alpöge had spent a year developing and running through Codex.

Alarmed by the coincidence, Buckmaster emailed OpenAI on September 3rd to demand an explanation. The following Sunday, he joined a phone call with Sébastien Bubeck, the OpenAI researcher spearheading the project. The narrative surrounding the project shifted rapidly during the discussion: - Initially, Buckmaster was told the system solved the problem with minimal prompting based purely on the raw problem statement. - Bubeck eventually admitted the broader team had actively worked on it, the prompt itself had been generated by Codex, and work had only commenced days earlier after hearing circulating rumors about Buckmaster’s breakthrough. - When Buckmaster asked point-blank whether OpenAI’s model had been trained on his private Codex sessions, he was told the system did not query real-time user data; pressed on whether those sessions were incorporated into training sets, Bubeck reportedly provided no answer.

Buckmaster claims OpenAI then presented him with an ultimatum containing two choices: 1. Buckmaster and Alpöge could publish their preliminary Euler results first, after which OpenAI would release its full Navier-Stokes proof, merely crediting Buckmaster as "the person who got closest to solving it." 2. Buckmaster could serve as author on the Navier-Stokes paper himself, on the strict condition that Alpöge be completely removed from the paper due to his employment at rival AI lab Anthropic.

Refusing both terms, Buckmaster threatened to go public. Bubeck allegedly warned him, asking why he would choose to "ruin his career," adding that "if he didn't want [Bubeck] to be nice, then he didn't have to be nice." OpenAI and CEO Sam Altman have pushed back vigorously against these allegations, claiming they never asked for Alpöge to be scrubbed, maintaining that their researchers acted with complete integrity, and insisting Buckmaster was the only party issuing threats.

[04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) The Aftermath, Terence Tao's Warning, and Front-Running Science

On Tuesday morning, Buckmaster and Alpöge published their papers accompanied by an exhaustive four-page statement recounting the entire dispute. By Tuesday afternoon, OpenAI published its own solution along with a blog post adamantly stating that no user data had been accessed and arguing that the two proofs were fundamentally distinct. The host jokes that by nightfall, Madame Zeroni from *Holes* joined the fray, demanding that Sam Altman and Sébastien Bubeck carry her up the mountain under threat of eternal curse.

Rumors have already begun circulating that OpenAI is closing in on verifying yet another Millennium Prize Problem, underscoring that transformative mathematical breakthroughs now seemingly require "a little elbow grease, 10,000 agents, and $20 million worth of compute."

Field Medalist Terence Tao offered what the host describes as the most level-headed response to the chaos: - Tao congratulated Buckmaster and Alpöge on their groundbreaking mathematical achievement. - He issued a sobering warning to the scientific community: if the mere rumor of ongoing academic research can unleash a swarm of corporate AI agents designed to front-run and scoop human discoveries, researchers will inevitably stop sharing early findings openly. - Such secrecy would dismantle centuries of open-science tradition and collaborative academic progress.

The video concludes with a lighthearted pitch for sponsor Railway, framing reliable infrastructure as a company that keeps its promises, before signing off.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) the future of finance — Narrator

> [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) meatbag bodies — Narrator

> [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) can the Navier-Stokes equations ever break or produce a singularity? — Narrator

> [00:00](https://www.youtube.com/watch?v=aspmNhKAFMc&t=0s) it stood as the closest anyone had ever come to breaking the real equations — Narrator

> [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) the person who got closest to solving it — Sébastien Bubeck

> [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) ruin his career — Sébastien Bubeck

> [02:41](https://www.youtube.com/watch?v=aspmNhKAFMc&t=161s) if he didn't want [Bubeck] to be nice, then he didn't have to be nice — Sébastien Bubeck

> [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) a little elbow grease, 10,000 agents, and $20 million worth of compute — Narrator

> [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) if the mere rumor of ongoing academic research can unleash a swarm of corporate AI agents designed to front-run and scoop human discoveries, researchers will inevitably stop sharing early findings openly — Terence Tao

> [04:25](https://www.youtube.com/watch?v=aspmNhKAFMc&t=265s) Such secrecy would dismantle centuries of open-science tradition and collaborative academic progress. — Narrator

## Mentioned OpenAI, Fireship, Claude-Louis Navier, George Gabriel Stokes, Tristan Buckmaster, Levent Alpöge, Anthropic, Claude Code, Codex, Clay Mathematics Institute, Sébastien Bubeck, Sam Altman, Terence Tao, Madame Zeroni, Holes, Railway, New York University (NYU)

## Video description Railway is the smoothest way to deploy software: https://railway.com/?referralCode=fireship

OpenAI claims it cracked a 90 year old math problem, but one NYU professor isn't happy about it. Let's dive in.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=aspmNhKAFMc)

中文简报

纽约大学教授指责OpenAI抢发纳维-斯托克斯数学证明成果企业AI实验室利用海量算力集群抢跑学术传闻，迫使研究人员隐瞒早期发现，恐将破坏开放的科学合作模式。该视频详细阐述了围绕OpenAI声称其智能体集群解决了历史性的纳维-斯托克斯千禧年大奖难题所引发的争议。纽约大学数学家Tristan Buckmaster与Anthropic研究员Levent Alpöge利用AI编程工具在相关的欧拉方程上取得了突破，但在听到有关其工作的传闻后，OpenAI据称动用了10,000个智能体和2000万美元的算力，采用了完全相同的全新方法。随着Buckmaster与OpenAI的Sébastien Bubeck围绕成果归属和私有用户数据的冲突升级，双方均发布了竞争性的论文和声明。这一风波最终引来菲尔兹奖得主陶哲轩（Terence Tao）的警告，他指出企业AI集群抢跑科学研究的做法可能会摧毁数百年来的学术开放传统。纽约大学数学家Tristan Buckmaster与Anthropic研究员Levent Alpöge利用AI工具Claude Code和Codex打破了欧拉方程，取得了突破性进展。 (0:00) 据报道，OpenAI动用了10,000个AI智能体和2000万美元算力来解决千禧年大奖纳维-斯托克斯难题，其采用的方法与Buckmaster和Alpöge开发的新颖技术一致。 (2:41) 针对OpenAI利用Buckmaster私有Codex会话数据以及因Alpöge隶属Anthropic而要求将其排除在外的指控，双方爆发了激烈的成果归属争议。 (2:41) 菲尔兹奖得主陶哲轩警告称，企业AI集群根据学术传闻抢跑成果的做法，可能会迫使研究人员走向保密，并瓦解开放科学体系。 (4:25)

Reading summary

A dispute has emerged between academic mathematicians and OpenAI over breakthrough proofs for fluid dynamics equations, spotlighting allegations of data scraping, research scooping, and aggressive corporate tactics. While researchers Tristan Buckmaster and Levent Alpöge accuse OpenAI of leveraging rumors and potentially private session data to replicate their work, OpenAI denies accessing user data and maintains its proof was developed independently. Academic Progress on Millennium Prize Equations (0:00) First formulated in the 1850s as an application of Newton’s second law to fluid dynamics, the Navier-Stokes equations describe the movement of liquids and gases. Despite widespread practical use in modeling weather, aerodynamics, and circulation, mathematicians have spent decades trying to establish whether the equations remain smooth indefinitely or break down into singularities. The question remains one of the Clay Mathematics Institute's Millennium Prize Problems, carrying a $1 million reward established 26 years ago. NYU mathematics professor Tristan Buckmaster, a 2019 Clay Research Award recipient, partnered last year with Anthropic mathematician Levent Alpöge to tackle these dynamics. Incorporating AI coding tools including Claude Code and Codex, the pair achieved a milestone on August 15 by proving that the inviscid Euler equations—a simplified formulation of Navier-Stokes—break down, representing the closest published step toward resolving the broader Millennium problem. OpenAI's Compute Swarm and Replicated Methodology (2:41) Shortly after Buckmaster and Alpöge made their Euler breakthrough, OpenAI intensified its focus on pure mathematics. In late August, OpenAI researchers fed all unsolved Millennium Prize Problems simultaneously into a newly trained model, backed by an estimated 10,000 agents and $20 million in compute power. OpenAI soon claimed to have fully broken the Navier-Stokes equations. Crucially, the model solved the problem using the exact novel, unconventional method that Buckmaster and Alpöge had spent a year developing and running through OpenAI's Codex tool, prompting suspicions over how the company converged on the same approach. Conflicting Attribution Accounts and Data Disputes (2:41) Following an inquiry email from Buckmaster on September 3, a phone call took place between Buckmaster and OpenAI project lead Sébastien Bubeck. According to Buckmaster, Bubeck acknowledged that OpenAI launched its effort only after hearing rumors of the academic breakthrough and that Codex generated the prompts. When asked whether Buckmaster's private Codex sessions were incorporated into model training sets, Bubeck reportedly gave no answer. Buckmaster further alleged that OpenAI gave him an ultimatum: accept secondary credit on the Euler proof while OpenAI published Navier-Stokes, or receive authorship on the Navier-Stokes paper on the condition that Alpöge be removed due to his employment at Anthropic. OpenAI and CEO Sam Altman vigorously rejected these claims, stating that their researchers acted with integrity, never requested Alpöge’s exclusion, never accessed private user data, and that their final proof is fundamentally distinct. The Threat of AI Front-Running Open Science (4:25) Following the dispute, both camps published their respective papers on the same Tuesday, with Buckmaster and Alpöge attaching a four-page statement detailing the conflict. Fields Medalist Terence Tao commended the mathematicians for their work but issued a stark warning regarding the incident's implications for academia. Tao cautioned that if circulating rumors of early academic breakthroughs trigger corporate AI swarms capable of front-running human discoveries with massive compute, researchers will stop sharing early results. Such defensive secrecy, Tao warned, threatens to undermine centuries of open collaboration in scientific research. The controversy highlights mounting tensions at the intersection of academic research and corporate AI capabilities. As labs deploy tens of thousands of automated agents backed by tens of millions of dollars in compute, the scientific community faces unresolved questions over data provenance, intellectual attribution, and the survival of open scientific exchange.

Source quotes

[0:00] Narrator: the future of finance [0:00] Narrator: meatbag bodies [0:00] Narrator: can the Navier-Stokes equations ever break or produce a singularity? [0:00] Narrator: it stood as the closest anyone had ever come to breaking the real equations [2:41] Sébastien Bubeck: the person who got closest to solving it [2:41] Sébastien Bubeck: ruin his career [2:41] Sébastien Bubeck: if he didn't want [Bubeck] to be nice, then he didn't have to be nice [4:25] Narrator: a little elbow grease, 10,000 agents, and $20 million worth of compute [4:25] Terence Tao: if the mere rumor of ongoing academic research can unleash a swarm of corporate AI agents designed to front-run and scoop human discoveries, researchers will inevitably stop sharing early findings openly [4:25] Narrator: Such secrecy would dismantle centuries of open-science tradition and collaborative academic progress.

Original transcript

[0:00] Earlier this week, while I was busy [0:01] losing all my money in what I thought [0:03] would be the future of finance, OpenAI [0:05] claimed to have solved another one of [0:06] mathematics' hardest problems. This [0:08] time, it was one invented back in the [0:10] 1850s, when the only program on Earth [0:12] was a degenerate gambler who I know [0:14] would have loved Horse Tender. The [0:16] problem is called Navier-Stokes, and [0:18] it's a set of equations we've grown to [0:20] rely on for determining the movement of [0:22] fluids and gases. But, unlike those [0:24] other videos you've watched about the [0:25] movement of fluids and gases, this one's [0:27] plot is way more interesting. [0:29] >> Because this is just like crazy drama [0:31] for us. [0:31] >> In this video, we'll find out what makes [0:33] Navier-Stokes so hard, look into whether [0:35] OpenAI actually solved it, and learn why [0:37] a math professor at NYU, a mathematician [0:40] at Anthropic, and Sam Altman are all [0:42] telling very different stories about [0:43] what happened on a phone call last [0:45] Sunday. It is September 11th, 2026, and [0:47] you're watching The Code Report. Back in [0:49] the 1800s, a French engineer named [0:52] Claude-Louis Navier wrote down some [0:54] equations describing how fluids move. [0:56] Then, a couple decades later, an Irish [0:58] mathematician named George Stokes fixed [1:00] the engineer's work, and eventually they [1:02] got branded as the Navier-Stokes [1:04] equations. They're basically force [1:06] equals mass times acceleration, but for [1:08] water. The idea is you give the [1:09] equations the current state of a fluid, [1:11] like how fast it's moving and how much [1:13] pressure it's under, and what you'll get [1:15] back is how that fluid will move a [1:16] moment later. And for almost 200 years [1:18] now, engineers have trusted them for [1:20] some pretty crucial stuff, like [1:22] forecasting hurricanes, designing [1:23] airplane wings, and simulating how blood [1:26] pumps through their meatbag bodies. But, [1:28] of course, mathematicians don't trust [1:29] anything until it's proven. And there's [1:31] one question about these equations that [1:33] nobody could answer, which in very [1:34] simple terms is, can they ever break? [1:37] That question has been so difficult that [1:39] 26 years ago, the Clay Mathematics [1:41] Institute made it one of their seven [1:42] Millennium Prize Problems and put a [1:44] million-dollar bounty on its head. To [1:46] claim the money, you either need to [1:47] prove the equations never break, or show [1:49] one case where they do break. As you can [1:51] imagine, there aren't a lot of people [1:53] who care about that problem, let alone [1:55] have any chance of solving it. But, [1:56] someone who does is Tristan Buckmaster, [1:59] a math professor at NYU who won the Clay [2:01] Research Award back in 2019 for some [2:03] work he did on these exact equations. [2:06] Last year, he teamed up with Levent [2:07] Alpige, a mathematician who happened to [2:10] work at Anthropic and in his spare time [2:12] apparently enjoys working on century-old [2:14] fluid dynamics problems. Together, the [2:16] team was making slow but steady progress [2:18] until mid-August of this year when [2:19] Levent realized what the company he [2:21] works for actually does. At that point, [2:23] they started leaning more heavily on Cod [2:25] code and Codex for help and their [2:26] progress took off. The approach they [2:28] ended up taking was novel but also a bit [2:30] of a long shot and on August 15th, it [2:33] paid off when they got a simpler version [2:35] of Navier-Stokes called the Euler [2:37] equations to fall apart. It wasn't the [2:38] million-dollar problem, but it was the [2:40] closest anyone had ever gotten. Then, as [2:42] if they had just gotten vaccinated, [2:44] OpenAI suddenly got very interested in [2:46] math. In late August, as they were [2:48] training a new model, they decided to [2:49] give it every unsolved millennium [2:51] problem at once to see what would [2:52] happen. The 10,000 agents and 20 million [2:55] dollars worth of compute later, they [2:57] claim to have broken Navier-Stokes. And [2:59] what do you know, the process they used [3:01] was the exact novel process that [3:03] Buckmaster and Alpige had spent the last [3:05] year working on and throwing into Codex, [3:07] which brings us to the phone call. On [3:09] September 3rd, Buckmaster emailed OpenAI [3:12] to ask what was going on. That Sunday, [3:14] he got on a call with Sebastian Boubek, [3:16] the OpenAI researcher running the [3:18] project. Buckmaster was told all OpenAI [3:20] did was give the model the problem [3:21] statement with very little input. Then, [3:23] as the call went on, it became clear [3:25] that the whole team worked on it. The [3:27] prompt was written by Codex and they had [3:28] only started it a few days earlier when [3:30] they heard rumors about his work. [3:32] Buckmaster then questioned whether the [3:33] model had been trained on his Codex [3:35] sessions and was told only that it [3:37] didn't look up user data, but when he [3:39] pressed on training specifically, [3:40] according to him, he got no answer. From [3:42] there, he claims OpenAI laid out two [3:44] options. The first was that Buckmaster [3:46] and Alpige could publish their Euler [3:48] results first, then OpenAI could publish [3:50] its Navier-Stokes proof and credit [3:52] Buckmaster as the person who got closest [3:54] to solving it. The second was that [3:56] Buckmaster could write the Navier Stokes [3:58] paper himself, but leaving off Alpagay [4:00] because he works at Anthropic. And so, [4:01] Buckmaster did what any gigachat would [4:03] do, and he refused both options and [4:06] threatened to go public. Then, according [4:07] to his statement, Boubacar asked [4:09] Buckmaster why he'd ruin his career and [4:11] told him that if he didn't want Boubacar [4:12] to be nice, then he didn't have to be [4:14] nice. But, of course, OpenAI has a [4:16] different story, claiming that they [4:17] never asked for Alpagay to be removed [4:20] from his work. And Sam says he believes [4:21] everyone on their end acted with [4:22] integrity, and Buckmaster was the only [4:24] one making threats. So, on Tuesday [4:26] morning, Buckmaster and Alpagay posted [4:28] their papers along with a four-page [4:30] statement describing how it all went [4:31] down. That afternoon, OpenAI posted [4:34] their own solution along with a blog [4:36] post insisting that no user data was [4:37] accessed, and that the two proofs are [4:39] significantly different. And that night, [4:41] Madam Zeroni posted her own statement [4:43] saying that Sam and Boubacar must carry [4:45] her up the mountain or OpenAI will be [4:47] cursed for all of eternity. [4:48] >> You and your family will be cursed for [4:51] always and eternity. [4:53] >> And just yesterday, rumors started [4:55] circling that OpenAI is very close to [4:57] verifying another millennium problem, [4:59] approving that all it takes to change [5:00] the world is a little elbow grease, [5:02] 10,000 agents, and $20 million worth of [5:04] compute. Of course, Terence Tao has the [5:07] most reasonable take on all this, [5:08] congratulating Buckmaster and Alpagay on [5:10] their work, and warning that if the mere [5:12] rumor of your research can trigger a [5:14] swarm of agents racing to front-run it, [5:16] mathematicians will simply stop sharing [5:18] ideas, which will undo centuries of open [5:20] science. The lesson here might be to be [5:22] careful which companies you trust, which [5:24] is why I host all of my projects on [5:25] Railway, the sponsor of today's video. [5:28] They once again refused to waste your [5:29] time with a full ad, which makes sense [5:31] for a company that lets you deploy [5:33] anything with just a few clicks. So, if [5:35] you want to say thanks, you can try [5:36] deploying your next project for free [5:38] with the link below. This has been The [5:39] Code Report. Thanks for watching, and I [5:41] will see you in the next one.