An AI Just Cracked a 30-Year-Old Statistics Puzzle in 90 Minutes

Here is a small story that says a lot about where AI is right now. A statistics professor at the University of Pennsylvania, Edgar Dobriban, handed a hard, unsolved problem to OpenAI’s newest model, GPT-5.6 Sol Pro. About 90 minutes later, he had an answer that had escaped human experts for years. The model that came before it, GPT-5.5, had chewed on the same problem for more than 20 hours and come up empty.

The puzzle sounds technical, but the idea behind it is simple. When scientists run thousands of tests at once (say, scanning the whole human genome to find genes linked to a disease), a lot of “hits” turn out to be false alarms just by chance. In 1995, two statisticians built a method to keep those false alarms under control. It is called the Benjamini-Hochberg procedure, and it is everywhere: the original paper has been cited more than 130,000 times. For years, people assumed the method also holds up when the data points are linked to each other (which real-world data usually is), but nobody had ever proven it. Dobriban used the AI to build a clean counterexample showing the method can quietly miss its own target. Simulations backed it up, and he published the code.

What’s behind it: Worth keeping expectations grounded here. The gap the AI found is small (a false-alarm rate of 0.104 where 0.1 was promised), so this matters more for theory than for your daily lab work, and it does not mean the method is broken. And the model did not invent a brand-new idea out of thin air. It combined known techniques in an unusual way that humans had not stitched together. That keeps a big question open: can models trained on human knowledge reason their way to genuinely new insights, or are they very fast at recombining what already exists? For now, “very fast recombination” is already turning out to be useful.

What this means for you: If you are just curious about AI, this is a good example of what the current tools are actually good at: taking a well-defined, gnarly problem and racing through combinations no single person would sit down and try. If you work in research or data, the practical takeaway is that these models are becoming real thinking partners for hard problems, not just writing assistants. The honest caveat, as one Berkeley statistician put it, is bittersweet: a milestone result used to mean a human colleague to celebrate. Now it can mean a chat log. That shift is worth sitting with.

Sources

• Edgar Dobriban on X
• Preprint (PDF)
• Code on GitHub
• The Decoder coverage