AI Closes 30-Year Math Gap; Prompt Engineering Becomes Theory Solver

This matters more than it sounds. Convex optimization is the backbone of modern machine learning, quantitative finance, and systems design. When researchers have been stuck on a problem for 30 years, it's not because they're lazy. It's because the problem is genuinely hard. Yet GPT-5.6 saw the structure in a way human researchers hadn't quite framed it yet.

What this really signals is a phase shift in how AI tackles theoretical computer science. For years, we've viewed LLMs as chat interfaces or code generators—useful, but ultimately operational. This result suggests they're becoming legitimate research tools for domains that have traditionally required deep mathematical intuition. The mechanism? Prompt engineering. Not architectural innovation. Not massive scale. Just asking the right question.

For founders, this has immediate implications. If you're building infrastructure that depends on optimization—scheduling systems, resource allocation, portfolio optimization—you now have a new tool in your toolkit. Rather than licensing proprietary solvers or waiting for academic breakthroughs, you might feed hard optimization problems directly to frontier models and see what emerges. The cost is negligible compared to hiring a PhD mathematician. The hit rate? Still being determined. But the asymmetry is worth exploring.

There's a second-order effect worth noting: this validates prompt engineering as a legitimate discipline. We're moving past the era where prompt engineering was seen as a hack or a parlor trick. It's becoming methodology. The researchers who figure out how to frame problems for these models—who understand the cognitive friction points that stall human mathematicians—will have outsized leverage.

The broader trend here connects to something we're seeing everywhere: AI is collapsing the distance between problem formulation and solution. Historically, 80% of hard problems were actually hard *formulation* problems. Get the question right, and the answer often follows. AI is becoming exceptionally good at the formulation part, or at least at exploring framings humans missed.

This also raises a pragmatic question for your own work. If you're a founder dealing with computationally hard problems—whether that's scheduling, routing, matching, or resource allocation—you should probably start experimenting with frontier models on these problems now. Not because they'll always win, but because the cost of experimentation is near-zero and the upside is genuinely asymmetric. You might unlock 10x faster solutions, or you might learn the boundaries of what these models can actually do.

The math community will spend months unpacking exactly how GPT-5.6 approached this problem and what it means for the field. But for builders, the takeaway is simpler: if you have a problem that's been unsolved for years, a fresh AI perspective might be worth a try—and a well-designed prompt might be all you need to see it differently.