Mathematician Accuses OpenAI of Stealing 20 Years of Research on Non-Sofic Groups

OpenAI is igniting the fury of the entire mathematics community!

Recently, NYU mathematician Tristan Buckmaster clashed fiercely with OpenAI over the attribution of the "NS equations."

A statement even revealed that OpenAI called mathematicians to pressure them and forcibly take credit for their work.

Unexpectedly, just as one wave subsides, another arises.

Today, Andreas Thom, a leading mathematician at TU Dresden and an expert in group theory, shared email "evidence" on social media.

He accused OpenAI's new Astra model of plagiarizing his 20-year research in the field of non-sofic groups.

OpenAI firmly denied this.

Within just a few hours, this news shot to the trending list.

Everyone was stunned again, with one netizen suggesting that mathematicians should return completely to the era of blackboards, pens, and paper.

The origin of this matter dates back to OpenAI's high-profile announcement in early August—

All ten "Fields Medal-level" problems have been successfully solved by Astra.

One of them includes the Gromov soficity conjecture, which had puzzled the field of group theory for 27 years.

Astra found evidence for the existence of a non-sofic group.

As soon as the news broke, the entire AI community cheered, saying, "AGI has opened the door to pure mathematical discovery."

However, top experts within the industry are becoming increasingly suspicious.

As a preeminent scholar with over two decades of experience in the field, Andreas Thom and his long-time collaborator Gábor Kun keenly observed:

The most critical and ingenious deduction step in Astra's proof process closely aligns with the technical pathway they are currently advancing.

More strangely, in Thom’s own words—

My approach and Kun's method were not the mainstream direction for tackling non-flexible groups; at the time, the academic community favored other paths, such as quantum game theory.

How did OpenAI’s models precisely identify and master our highly obscure and detailed technique?

The answer was revealed quickly, and the truth was utterly chilling.

Months before OpenAI officially released its results, Thom and his colleagues extensively used ChatGPT to explore the "extension matching problem," typing into the chat window numerous unpublished core derivations and the latest proof details related to this conjecture.

Twenty years of dedication spent in obscurity, months of iterative dialogue with AI—suddenly rebranded as OpenAI’s groundbreaking autonomous achievement.

Anyone else would never let this slide.

After detecting the anomaly, Andreas Thom chose not to act immediately.

He maintained the restraint expected of a scholar and sent a formal email to OpenAI’s two core researchers, Mark Sellke and Sébastien Bubeck.

In the letter, Thom raised two extremely precise and logically rigorous inquiries:

Have our conversation records discussing unpublished results in ChatGPT over the past several months been included in the model's training data?

2. During the model's problem-solving and reasoning process, can these records be directly retrieved and accessed?

In response, researcher Mark Sellke sent a very brief email containing only one cold, impersonal sentence:

Regarding your conversation with ChatGPT: That never happened.

No data audit explanation, no privacy mechanism details, not even any formatted pleasantries—just a blunt, arrogant, and unequivocal rejection.

Thom also disclosed a date—June 29—when he disabled the model training option.

However, he believes that turning off the switch can only apply going forward; it cannot address what occurred in earlier conversations or determine whether derivative data has already been included in the training process.

Users can see a setting but cannot view the backend data flow.

Thom later recalled that OpenAI’s response was not only rude but also steeped in deceitful sophistry:

He clearly distinguished between "entering the training set" and "online invocation," but Sellke attempted to dismiss all substantive issues with a vague statement of "it never happened."

However, the cost of arrogance came swiftly.

Just weeks later, OpenAI was backed into a corner in the dispute over the Millennium Prize problem known as the Navier-Stokes equations.

Faced with increasing pressure from Tristan Buckmaster and Levent Alpöge, OpenAI’s official public relations stance suddenly reversed:

Although insisting that no private data from specific accounts was accessed, it acknowledged—

We cannot rule out using de-identified data from customer products to improve models.

The boomerang struck precisely back at its own throat.

Previously, when confronted by Thom, you firmly denied any such thing, but now you’ve admitted that de-identified data was used to improve the model.

So, as long as you erase the mathematician’s name, humanity’s unpublished, groundbreaking insights can be openly “de-identified and sanitized” on Silicon Valley giants’ servers and magically transformed into AI’s autonomous intelligence?!

In his latest long-form article, Andreas Thom mercilessly dismantles this wordplay:

Anonymization may remove a scholar's name, but it cannot erase the intellectual core of a mathematical idea.

From Tristan Buckmaster to Andreas Thom, who has now come forward.

The first could be called a “coincidence,” the second explained away as a “misunderstanding,” but the successive, collective defections and accusations by top scholars have completely shattered OpenAI’s carefully crafted “AGI scientific myth.”

As Valerio Capraro emphasized in his lengthy article, this is no longer a few isolated academic disputes:

If these allegations are proven true, we are witnessing one of the largest intellectual theft scandals in the history of science.

So, who will dare to open this dialog box leading to the abyss next time?

Today, OpenAI owes the global mathematical community a genuine explanation.

Reference materials:

https://x.com/IntCyberDigest/status/2097825474961395899

https://x.com/ValerioCapraro/status/2097791836269977996?s=20

https://mathstodon.xyz/@andreasthom

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by Peach.