Video summary

Czy OpenAI okradło matematyka?

Main summary

Key takeaways

News and Commentary

Overview

The speaker reacts to OpenAI’s September 8 announcement that an AI system has formally “solved” the Millennium Problem involving the (Navier–)Stokes equations. They frame it as a major breakthrough—especially because only a few of the seven Clay Millennium Problems have been solved so far, and this would be the first time a solution is claimed without a human directly proving it.

At the same time, they argue that the community must wait for the proof to be accepted and thoroughly checked for loopholes, hidden assumptions, and correctness.

Claims about the nature of the breakthrough

The Clay Millennium Problems framework

  • The Millennium Problems were created by the Clay Mathematics Institute (seven problems announced in 2000).
  • Each problem carries a $1M reward after formal publication in a peer-reviewed journal.
  • There is also a two-year window during which refutation is expected/allowed if issues are found.

Focus on the Navier–Stokes “blow-up” question

The speaker highlights the long-unsolved challenge of whether Navier–Stokes/Stokes-type equations can produce non-physical “blow-ups,” such as:

  • Infinite or unbounded velocities arising from real initial conditions.

This is tied to mathematical rigor in fluid dynamics.

Compute scale used for the AI effort

They claim OpenAI’s approach relied on extremely large internal computation, described as:

  • many “agents” (on the order of ~1,000)
  • exchanging millions of messages
  • producing hundreds of billions of tokens
  • with Navier–Stokes alone requiring very large compute resources

Economics as a red flag

The speaker argues that the $1M prize is tiny compared to the described compute costs, implying:

  • huge resources may already be determining “wins,” not scientific process
  • the economics themselves suggest potential distortion of incentives and fairness

Main allegations: priority and intellectual-property concerns (“Did OpenAI steal it?”)

The video’s central point is not only celebrating the math result, but raising concerns about:

  • intellectual property
  • scientific priority
  • transparency about how the solution was produced

Alleged timeline and race dynamics

The speaker alleges OpenAI may have started pursuing this direction after hearing rumors that researchers—possibly including people associated with Anthropic or competitors—were close to solving parts of the Navier–Stokes problem.

They identify several individuals as relevant to parallel work, including:

  • Tristan (also referred to later as Levent)
  • an Anthropic employee (named Levent in subtitles)
  • Sebastian Bubek (described as a mathematics department head at OpenAI)

Narrower related work (Euler subproblem)

They suggest that these researchers may have been working on a related but narrower version—such as:

  • an Euler equations special case

They argue that when OpenAI entered with massive compute, it could have affected authorship/credit.

Fast confirmation via formal proof checking

The speaker describes an alleged process where OpenAI’s verification/confirmation was fast using:

  • Lean, a logic/programming tool commonly used to check formal proofs

They also suggest that the proof-checking and/or coding assistance may have involved OpenAI tools (for example, a codex-like system), raising the possibility that:

  • competitors’ work could have influenced the system’s output
  • even if only indirectly (e.g., through training patterns or tooling)

Evidence and uncertainty the speaker highlights

Lack of direct proof of wrongdoing

The speaker repeatedly stresses that proving misconduct is unclear, calling it:

  • “one person’s word against another”

However, they argue that several public statements look “very bad” and merit scrutiny.

Statements about data use and model improvement

They reference OpenAI’s claim that it “can’t be ruled out” that user data from OpenAI products helped improve a de-identified model.

The speaker interprets this as potentially allowing:

  • competitor-related information (or similar patterns)
  • to feed into training or future solving

Limited transparency about the agent prompts

The speaker emphasizes that OpenAI reportedly has not publicly released:

  • the prompts used to launch the agent swarm

They argue this makes independent verification difficult, especially for assessing:

  • whether the solution relied on others’ unique insights

Early technical doubts by other mathematicians

They also note that other mathematicians reportedly raised technical doubts early, including concerns about:

  • potentially misleading or problematic assumptions in the proof

Even if the formal proof checks out, they imply correctness scrutiny may still remain.

Broader argument: AI shifts the fairness of scientific discovery

The speaker argues this incident reflects systemic problems beyond a single case.

Priority distortion

  • Companies with far more computational and financial resources can “buy” speed.
  • This may lock out researchers who don’t have comparable budgets.

Transparency vs. proprietary control

They argue private companies may hide:

  • prompts
  • training-policy connections relevant to proofs
  • agent directives

This limits the community’s ability to adjudicate credit fairly.

Lack of accountability

They claim AI providers can potentially:

  • charge for results
  • while not being fully responsible if outputs are wrong or problematic

Plagiarism/IP concerns at a larger scale

They compare the situation conceptually to earlier plagiarism/IP cases (via an anecdote about copied software for scientific measurement automation), arguing:

  • power dynamics here are worse
  • the scale and speed are unprecedented

Call to action / desired changes

The speaker’s conclusion is that the community and regulators should address:

  • How priority should be established when AI tools are involved
  • What transparency is required, such as:
    • prompts
    • agent workflows
    • training-policy details relevant to proofs
  • How to handle intellectual property and consent in training
  • How to ensure scientific integrity when the “winner” may be decided by compute rather than open scholarly process

Overall takeaway

The video is both:

  • a reaction to the astonishing claim that an AI system resolved a Millennium Problem, and
  • a warning that—regardless of whether OpenAI’s proof is ultimately correct—the way such proofs are generated may already be creating conflicts over:
    • ownership
    • credit
    • and the future structure of scientific discovery

Presenters / Contributors

Presenter / narrator

  • Unnamed in subtitles: “the speaker/author of the video” (speaks throughout)

Mentioned individuals

  • Grigory Perelman
  • Tristan (math professor in New York; also referred to in subtitles as Levent)
  • Levent (Anthropic employee; as named in subtitles)
  • Sebastian Bubek (described as a mathematics department head at OpenAI)
  • Stan Polasek (mentioned as noticing potential issues early)
  • Terence Tao (mentioned as receiving a Fields Medal; discussed in relation to early correspondence)
  • “Tomek Rożek” (mentioned as potentially covering similar info first, in a “views/priority” context)
  • Andrzej Dragan (mentioned as someone who could explain Lean / verification in more detail)

Original video