AI is solving problems by the hundreds. Who chooses where to dig next?

OpenAI has released 722 mathematical manuscripts across 372 families of results. AI enthusiasts asked GPT-6 Pro and Fable 5.1 to rank mathematical discoveries from the past three years. By their count, 81% of the top-ranked results came from OpenAI’s release that same day, October 6. According to the models, one company had practically rewritten the recent history of mathematics in a single release.

That is how a hallucination of scale takes hold: suddenly, the model is a superhuman scientist churning out great discoveries. The results themselves may be genuine and enormously important. The fantasy lies in the leap from hundreds of solved problems to an independent scientific genius. A generated leaderboard can blend novelty, prestige and scientific significance into one dazzling ranking.

“81% of them have been released today. wtf” — X user, on the model-generated ranking

Today’s agents can rapidly work through “generative lakes” that humans have already found: open problems, technical gaps and research directions with established questions, language, literature and criteria for success. Within those lakes, an agent can devise a new proof, an unexpected approach or a powerful generalization. The territory, though, has already been mapped. The machine has a starting point and a way to tell whether it is making progress.

The productivity gain is real and potentially enormous. Years of work could become much shorter; neglected lines of inquiry could finally get the attention they deserve. But a powerful tool for doing research and the judgment to set its direction remain very different capabilities, however similar they look from the outside. Research taste means choosing a fruitful question, developing the language to ask it and recognizing a promising research program before anyone has drawn up a list of problems.

OpenAI also has a more practical problem. Since the Navier–Stokes episode and the ensuing dispute over authorship and academic norms , it has spent roughly a month working out how to turn machine-generated results into conventional scientific publications. We have already written about the need to trace an AI discovery’s origins. One bottleneck is now glaringly obvious: someone willing to serve as corresponding author. This goes well beyond an academic rubber stamp. Someone must understand the result, stand behind it, answer reviewers and remain accountable after publication. This batch arrived as a GitHub repository, rather than a collection of journal papers with named human authors. The proofs and Lean formalizations are serious mathematical work. Getting them into the scientific literature still requires human work and responsibility.

The accumulated backlog could last a long time, and clearing it quickly will open new directions in its own right. The harder question comes when the most attractive lakes have been drained. Can the machine find the next fruitful research program on its own—and understand why it is worth pursuing? The cutting power is already impressive. Who decides where to put the tool to work next?