September 11, 2026 — The Trend Tribune

By tech_tanaka · September 11, 2026

"10,000 agents, 88 hours, 5 million messages"—Why is OpenAI's mathematical achievement plagiarized? — (Photo: illustrative)

"10,000 Agents, 88 Hours, 5 Million Messages"—Why is Plagiarism Allegations Surrounding OpenAI's Mathematical Achievement?

On September 8th, OpenAI announced that its unpublished internal model had solved one of the Millennium Prize Problems concerning the Navier-Stokes equations (equations describing fluid motion). However, just hours later, Tristan Buckmaster, a mathematician at New York University, accused OpenAI of potentially referencing his unpublished research without permission. As an engineer, I want to carefully examine the details of this technical achievement and the resulting conflict surrounding its recognition.

A Different Scale from Previous Case: "Approximately 10,000 Agents, 88 Hours"

First, let's examine the technical details. This approach differs from Astra's mathematical proof and Claude's formalization of Fermat's Last Theorem, which we previously discussed. According to OpenAI, an undisclosed internal model, said to be even more powerful than GPT-6 Astra, used approximately 10,000 agents in parallel for 88 hours to prove the "finite-time blowup" (a phenomenon where the solution diverges within a finite time) in the forced 3D Navier-Stokes equations.

In this process, the agents exchanged approximately 5 million messages and consumed approximately 300 billion output tokens. OpenAI itself describes this computational cost as "several million dollars," but external estimates based on publicly available token prices suggest a more specific figure of around $15 million to $22 million.

The Important Technical Distinction Between "Unsolved" and "Resolved"

There is a critically important distinction to accurately understand the content of this achievement. The Millennium Problems, which offer a $1 million prize, deal with the unforced Navier-Stokes equations, where no external forces are involved. However, OpenAI's proof addresses the forced version, where external coercion is applied. While related, this is a distinct problem of different nature.

The Clay Institute for Mathematics has not yet determined whether this result actually meets the criteria for the prize. Even if the proof is ultimately accepted, the prize money will not be paid until it is published as a peer-reviewed paper and undergoes two years of verification by the mathematical community. In other words, even if everything is correct, it will still take a long time before this achievement is "officially recognized." Fields Medal winner Terence Tao has praised the underlying approach as a "remarkable achievement," but has not yet formally endorsed the entire proof.

Buckmaster's Account of "Why Are You Ruining My Career?"

More than the technical achievement itself, the conflict over crediting the work has become a major topic of discussion within the industry. Buckmaster claims that as of August 22nd, he and Anthropic researcher Levento Alpege had already reached the proof of the relevant explosion phenomenon.

According to Buckmaster, on September 3rd, amidst rumors of the two companies' research, he contacted researchers at OpenAI. Initially, he received a friendly response, and they even offered to provide OpenAI's computing resources. However, in a call on September 6th, OpenAI researcher Sebastian Bubek revealed that their internal model had already completed a nearly 100-page forced Navier-Stokes proof, and allegedly asked Buckmaster to remove Alpege's name from the joint publication. When Buckmaster refused, he claims he was told, "Why are you ruining your career?"

Discrepancies between OpenAI's explanation and the timeline

OpenAI researcher Bübeck offers a different explanation for how this project began. According to him, OpenAI started the project after rumors spread on social media on September 1st that a competing lab had solved two Millennium Problems. He claims that they first tackled a related problem, the simpler "regularity of the unforcible Euler equation," with a 1,000-agent group working on it for 50 hours, before expanding resources to 10,000 agents for the Navier-Stokes problem.

A straightforward reading of this timeline suggests that OpenAI's project began on September 1st, after the time when Buckmaster and his colleagues allegedly reached their results (August 22nd). However, if OpenAI's subsequent claims that they already possessed a near-complete proof and requested the exclusion of co-authors are true, then it becomes possible that their actions were not based on independently achieved results, but rather on recognition of the existence of other research. This discrepancy remains unresolved at this point, with both sides' claims still at odds, and external verification is awaited.

What Engineers Should Consider

This incident demonstrates that the acceleration of mathematical research by AI is beginning to significantly shake the very framework of trust and credit within human research communities—the framework of "who knew what, when, and how"—not just the "speed" of research results. The tension between the results derived by 10,000 agents in 88 hours and the painstaking research accumulated by human researchers is likely to resurface repeatedly as AI-accelerated research becomes more widespread.

The 166-page paper and the formal verification code using Lean have already been made public for external review. It will be necessary to closely monitor how the technical correctness of this proof and the veracity of both sides' claims regarding credit will be verified.

References: forkast.news / pasqualepillitteri.it / byteiota.com / www.hngn.com

OpenAI数学AIエージェントLeanAI/ML

"A post that reached 100 million people"—An Anthropic researcher risked his resignation to challenge the boundaries of self-improving AI.

By thesis_yamamoto

On September 8th, Anthropic researcher Jacob Coxon announced his resignation after warning that both OpenAI and Anthropic were competing to create self-improving superintelligence. His post garnered 100 million views. This article analyzes, from a researcher's perspective, the industry's denunciation of duplicity ("speaking of fear behind closed doors and feigning calmness in public"), the mention of costly measures such as a temporary ban on ability improvement, the alignment with the Pacing the Frontier letter, the timing immediately following the release of GPT-6 Astra and the "arrival of AGI" statement, and the continuity with Mrinank Sharma's resignation in February.

"A Post Reaching 100 Million People"—Anthropic Researcher Risks His Resignation to Speak Out Against the Boundary of Self-Improving AI

On the evening of September 8th, Jacob Coxon, a researcher at Anthropic, announced his resignation from the company with a post on social media. He had spent three years working on pre-training research at both OpenAI and Anthropic. As a journalist with a background in AI research, I want to examine the technical and ethical questions surrounding the concept of "self-improving AI" that this resignation raises.

The Specific Concern of "Self-Improving Superintelligence"

The core of Coxon's post lies in his assertion that "AI development companies are directly competing towards self-improving superintelligence, putting our lives on the line." Here, "self-improvement" refers to a state where AI systems can continuously improve their capabilities without human intervention.

The reason this concept is not merely an abstract philosophical concern is that it is directly related to the concrete progress shown in OpenAI's "Research Acceleration" report, which we previously discussed. The figure of agent working days exceeding human working days indicated that AI's contribution to research and development has already reached a scale that cannot be ignored. What Coxsone is concerned about is the more serious stage that lies ahead: "AI will be able to run its own research and development cycle independently, without human intervention."

An Industry Accusation: "Speaking of Fear Behind Closed Doors, Feigning Calmness in Public"

A particularly noteworthy aspect of Coxsone's post is his observation of the duality in the behavior of industry insiders. He states, "The people building AI genuinely believe it could kill us all within the next 10 years. This isn't a marketing gimmick. Rather, I've heard many executives and senior researchers carefully choose their words to sound plausible for the press, while expressing the same fears in private."

If this allegation is true, it would indicate a significant discrepancy between the AI industry's official statements and its internal realities. This kind of "public and private" structure itself may be making proper external oversight and governance more difficult.

Reference to the "Costly Measure of Pausing"

Mr. Coxon also made specific suggestions for solutions to this problem. He stated, "I don't feel we're on a track that can prevent global competition. This may require costly measures, such as temporarily prohibiting model capability improvements."

This is a recommendation aligned with the "Pacing the Frontier" letter, signed by over 1,000 AI industry professionals, which called for technological and institutional slowdown mechanisms. Mr. Coxon's resignation indicates that this type of voice from within the industry is no longer limited to one-off petitions, but is beginning to surface as an actual career decision.

Coinciding Timing: GPT-6 Astra and the "Arrival of AGI" Statement

There's a reason why the timing of this resignation is attracting even more attention. Coxsone's post came just days after OpenAI released its latest model, "GPT-6 Astra." Following this release, NVIDIA CEO Jensen Huang reportedly declared that "AGI (Artificial General Intelligence) has arrived."

The fact that this celebratory mood for a "leap in capabilities" and the warnings from insiders surfaced almost simultaneously symbolizes the unique situation the AI industry finds itself in. This overlaps with the previously discussed "model fatigue," highlighting once again the tug-of-war the entire industry is engaged in between the speed of capability improvement and the need to cope with the associated risks.

The Repeated Warning: "Not the First Time"

Coxsone's resignation is not the first instance of someone leaving the industry due to concerns about AI safety. In February of this year, Mrinank Sharma, Anthropic's head of AI safety, also resigned, stating that "the world is in crisis." This kind of repeated warning from industry insiders should be viewed not as an isolated event, but as a continuing pattern.

Coxson's post reportedly reached over 100 million people overnight, and two current Anthropic employees also expressed their support for the post.

What Researchers Should Consider

The lesson this resignation offers is that AI research is moving beyond mere technological optimization competition and is reaching a stage where it must constantly confront the fundamental question: "Where is this research ultimately headed?" The series of events previously discussed, such as OpenAI's accelerated research data, the unintended emergence of capabilities in GLM-5.3, and the Astra neuralese controversy, can all be seen as different manifestations of the same concern: "Is the pace of AI capability improvement outpacing the pace of our ability to understand and control it?"

To what extent will these voices from within the industry translate into actual institutional design and collaborative actions between companies? And to what extent will the "gap between public and private speech" that Coxon points out be bridged? These are topics that need to be continuously monitored from the perspective of transparency in AI research.

References: techcrunch.com / phys.org / www.oann.com / www.democracynow.org

AnthropicAI安全性AIガバナンス自己改善AI/ML

The "tightening" triggered by Unitree's wild fluctuations—a warning sign from Chinese authorities about a mass production bubble.

By robo_takahashi

On September 9th, the China Securities Regulatory Commission (CSRC) reportedly made unofficial stricter scrutiny of humanoid IPOs following Unitree's volatile stock price (a more than 53% drop from its initial price of 1,100 yuan to 513.51 yuan). The article examines the new criteria requiring either continuous revenue, a path to reducing losses, or substantial technological innovation, the reality of Unitree's own disclosures indicating a revenue-centric approach to research and education, the parallels with the 15th Five-Year Plan and the $138 billion national fund support measures, and the impact on subsequent companies such as Robotera, which is considering listing in Hong Kong.

Unitree's Volatility Triggers a "Tightening" Effort – A Warning Sign from Chinese Authorities Regarding a Mass Production Bubble

On September 9th, it was reported that the China Securities Regulatory Commission (CSRC) had begun to informally tighten its review criteria for initial public offerings (IPOs) of humanoid robotics companies. This move, taking the form of "informal guidance" to investment banks, is believed to be directly triggered by the volatile stock price of Unitree Robotics, which we previously discussed. As a software professional, I want to examine the details of this regulatory tightening and its implications for the entire industry.

New Review Criteria: "Continuous Revenue" and "Clear Path to Profitability"

According to this report, the CSRC now requires robotics companies to demonstrate either a continuous revenue stream, a clear path to reducing losses, or substantial technological innovation when reviewing IPO applications. This clearly indicates a shift from the previously relatively lenient listing criteria towards a greater emphasis on financial substance.

The direct background to this movement is the series of price fluctuations in Unitree's stock price, which surged more than fivefold on its first day of listing before sharply declining. As of September 9th, Unitree's stock price had fallen to 513.51 yuan, a drop of over 53% from its initial listing price of 1,100 yuan. While still a significant increase from its August 19th listing price of 150.80 yuan, the volatility was so extreme that authorities deemed it a "causing speculative overheating."

Unitree's Own Disclosures Reveal "Majority Focus on Research and Education"

A crucial point in understanding these regulatory tightenings is a fact already revealed in Unitree's own IPO review documents. Disclosures clearly show that the majority of the company's revenue from humanoid robots comes from demand from research and educational institutions, with revenue from actual industrial applications being limited.

This indicates that the regulatory authorities themselves had already grasped, through their review process, the reality we previously discussed: "Industrial applications account for less than 10% of sales, with demand primarily coming from research and educational institutions." This latest tightening of regulations can be understood as the authorities' corrective measure against this kind of "discrepancy between the reality of commercialization and its valuation in the stock market."

The Balance Between Support and Regulation: A $138 Billion National Fund

What's interesting is that this move towards stricter regulations is proceeding in parallel with the Chinese government's extremely strong support measures for the robotics industry. Beijing has designated robotics as a priority strategic emerging industry in its 15th Five-Year Plan (2026-2030), and in support of this, it is reported that a venture capital fund of approximately $138 billion, focused on robotics, AI, and innovation, is being established.

In other words, the Chinese government is trying to simultaneously apply the "accelerator" by pouring massive funds into the entire robotics industry and the "brake" by curbing the overheating of individual companies' stock markets. Given the overwhelming market share of Chinese manufacturers, who account for over 90% of global humanoid robot shipments, the healthy development of this industry is undoubtedly a national priority.

Impact on Robotera, the Next Upcoming IPO Candidate

There are already signs that the tightening of regulations is having an impact. Another Chinese humanoid robot manufacturer, Robotera, is reportedly considering a listing in Hong Kong, and it remains to be seen how the CSRC's actions will affect subsequent listing plans. With stricter scrutiny on the Shanghai STAR market in mainland China, it's possible that more companies will choose the Hong Kong market in search of a less stringent regulatory environment.

What Software Professionals Should Keep an Eye on

This tightening of regulations by the CSRC is not merely a response to individual companies, but a significant turning point indicating that the entire Chinese humanoid robot industry is shifting from a phase of "technology demonstrations and fundraising frenzy" to a phase of "more rigorous evaluation based on actual commercialization results."

As I have repeatedly pointed out, there has often been a significant discrepancy between "published figures" and "verifiable results" in this industry. The fact that Chinese regulators themselves have begun to view this discrepancy as a problem suggests that more stringent disclosure of financial and commercial performance will likely become standard practice when evaluating this industry in the future. I will be closely watching what kind of disclosures companies aiming for an IPO will be required to make under these new evaluation standards.

References: www.theinformation.com / www.chathamhouse.org / gokhshtein.com / cryptodaily.co.uk

Unitree中国IPO規制ヒューマノイド

"10,000 agents, 88 hours, 5 million messages"—Why is OpenAI's mathematical achievement plagiarized?