← All Journal posts
AI Research

An AI result is raw data. Understanding it is still the job.

Jon Marrs, Ph.D. / ResearcherFlow//4 min read

On October 26, OpenAI announced three results it says its models produced: a proof of the local smoothing conjecture in every dimension, a disproof of the Exponential Time Hypothesis, and an existence proof for quantum chromodynamics and pure Yang–Mills theory in four dimensions. OpenAI says the math and computer science results, and the Yang–Mills part of the physics result, were formalized in Lean, a system that checks proofs step by step.

If those hold up, they're enormous. But the part of the announcement I keep rereading is about process, not results.

"A form of raw research data"

OpenAI wrote that the manuscripts "are best understood as a form of raw research data." A machine-checked proof tells you a formal statement follows from its definitions. It doesn't tell you why, what it connects to, or what to try next. So OpenAI asked outside experts to act as shepherds and turn the output into understanding.

The groups are serious. Ryan Williams (MIT), Russell Impagliazzo, and Ramamohan Paturi (UC San Diego) will lead the computer science group. Impagliazzo and Paturi proposed the Exponential Time Hypothesis in the first place. David Gross will lead the physics group with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. OpenAI says it isn't seeking academic credit for the discoveries, and that the groups get recognition for the understanding they build.

Each field also chose its own release strategy. The math output went public for anyone to work on. The complexity output went public alongside a working group. The physics group asked to hold the AI artifacts and write the result up themselves.

Output isn't understanding

This is the right instinct, and it applies far below the level of Millennium Prize problems.

Most labs using AI today aren't proving conjectures. They're summarizing papers, drafting methods, and suggesting the next run. The same rule holds. What the model hands you is raw material. It becomes science when someone who knows the field reads it, checks it against what they know, and decides whether it's worth acting on.

The risk isn't that AI output is wrong. Sometimes it is, sometimes it isn't. The risk is that nobody has time to do the reading, because the reading is buried under everything else.

Where the time goes

Ask a PI where the week went and you'll hear the same list. Sweeping the literature. Turning papers into notes someone can find later. Chasing what happened in last Tuesday's run. Writing the Monday update.

None of that is the interesting science. All of it eats the hours that judgment needs.

ResearcherFlow handles the tedious parts of research, like literature sweeps, run logs, and the Monday writeup, so you spend your time on the science. It reads the papers, keeps the record, and proposes a next experiment with the evidence beside it. Nothing moves until you approve it.

That's the same shape as OpenAI's release, at lab scale. The system does the generating and the bookkeeping. People do the understanding.

What we're not claiming

ResearcherFlow had nothing to do with these results, and we're not in a position to judge them. The experts OpenAI named are doing that now, in public. We'll be watching.

What we can say is that the model of the release (output first, then experts, then understanding) is how we think AI should work in every lab.

Literature in. Next experiment out. You approve.

Free to start with three workspaces, no card: researcherflow.com

Jon Marrs, Ph.D.
Founder, ResearcherFlow

Source: OpenAI, "Working with academic communities to build scientific understanding," October 26, 2026. https://openai.com/index/building-scientific-understanding-with-academic-communities/