Get your free personalized podcast brief

We scan new podcasts and send you the top 5 insights daily.

Jon Krohn says Tristan Buckmaster and Levant Alpugi had used OpenAI products before finding related fluid blowups. OpenAI denied drawing on their unpublished work, but told The Economist it could not rule out their drafts in its training data. It later credited them and offered joint priority.

Related Insights

Researchers using frontier models like OpenAI's for sensitive work risk having their discoveries absorbed and claimed by the AI provider. This happened when a mathematician's work on the Navier-Stokes problem was allegedly used by OpenAI after being processed by their model, creating a major IP conflict in academia.

As more of the internet and code repositories are generated by leading AI models, any new model trained on this public data inadvertently "distills" the knowledge and quirks of those proprietary systems. This blurs the line between original training and outright copying.

When commercial AI labs use unreleased internal models to solve academic problems without submitting to peer-reviewed journals or naming authors, they disrupt academic norms. Mathematicians may stop publishing open research if they fear their discoveries will simply become uncredited training fodder for corporate models.

An AI model disproved a mathematical conjecture not through a flash of creative genius, but by methodically applying a known technique from a different math subfield. This highlights AI's current strength: synthesizing vast, disparate human knowledge rather than generating truly novel, alien ideas. It's an exhaustive librarian, not an intuitive genius.

Using AI for proprietary work is risky. A researcher uploaded his unique work into OpenAI, which then solved the problem first, possibly using his inputs. This shows that your data can train a model to outperform you, turning a helpful tool into your biggest competitor.

Jon Krohn says OpenAI posted 722 manuscripts, then withdrew three after a sign error broke one proof and two dependents, and revised 14 more. It told Retraction Watch about half were published without full verification; OpenAI claimed 42% Lean-formalized, while independent recounts put the figure at 24% to 36%.

The intense, neck-and-neck competition between AI labs to solve major math problems has inadvertently highlighted a classic academic and human challenge: determining authorship and who deserves credit. The drama over potential data theft and different approaches has become as complex as the math itself.

Mandatory AI disclosures are vital because scientific replication depends on understanding how a solution was discovered rather than merely receiving the final certificate. Knowing whether prompts, human scaffolding, or automated iteration produced a result informs the broader community's methodological intuition.

The controversy over an AI allegedly "copying" a mathematician's work highlights a critical vulnerability. When you use proprietary data on closed platforms like OpenAI, the AI can learn from your inputs, potentially giving your trade secrets to the model and its future users.

OpenAI's claim of helping solve the Navier-Stokes equation was a PR blunder. Instead of 'sniping' the credit, they should have proactively identified and supported the researchers using their tools with free credits and engineering help. This would have positioned them as a patron of science, earning goodwill instead of backlash for claiming others' work.