Last week, OpenAI released a massive batch of math preprints. One shows that “ultraflat” ±1 polynomials exist once they’re long enough, but gives no explicit length bound. I burned 16 million tokens (about $15) to make that bound explicit.

Until a few months ago, I would have expected a task like this to take an expert or a team days or weeks. It took me about 30 minutes to produce a draft. The length bound is wildly impractical (a triple exponential), but now there’s a number.

The most accurate description of my role is “patron.” I have mixed feelings about that. But the leverage is thrilling, and I feel lucky that LLMs are good at the kinds of problems I enjoy. I’m realizing there’s a growing role for experts who can interpret the often alien output of LLMs: make it understandable, give it structure, and connect it to what we already know.

My feed shows a mix of reactions to OpenAI’s release: excitement, despair, snark, gallows humor. I don’t think knowledge workers, including STEM faculty, will become obsolete any time soon. But underneath those reactions are human problems. Machines already produce far more than people can read or check. We don’t yet know how to absorb that volume of work while continuing to pass on intuition and taste.

The reactions also reflect that a profession is tied not just to income but to identity. Some of the specialized expertise that took decades of training and work to develop is now available for a modest monthly subscription.

These challenges extend beyond mathematics to law, medicine, analytics, software development, journalism, and many other fields. They’re serious problems, but I also see people working on them. That keeps me optimistic.

Paper and code: github.com/jeff-kline/ultraflat-littlewood-rate