I spent two years building something in obscurity. A quiet, obsessive side project that only I thought mattered. Then one day, an AI—Claude—used it to crack a puzzle mathematicians had been wrestling with for centuries. The Jacobian conjecture, a problem that has haunted algebraic geometry since 1939, fell to a counterexample discovered not by a human, but by an autonomous agent running on my pipeline.
Let that sink in. The AI didn’t guess the answer. It didn’t brute-force its way through. It used a structural framework I had designed—a pipeline of mathematical transformations, constraints, and search heuristics—to navigate the problem space and find exactly what the conjecture was missing. We spend so much time trying to make AI answer our questions. We should be spending it building the scaffolding that lets AI ask its own.
Here’s the part that still gives me chills: I’m not a mathematician. I’m a developer who got obsessed with a paper in Nature (https://www.nature.com/articles/s41467-026-73900-9) and decided to model its logic as a reusable pipeline. I released it under CC0 on GitHub, shared it on LinkedIn, and one of the original paper’s authors actually approved. I didn’t expect anything to come of it. But the mathematical community around this counterexample is now thinking through structures I spent years developing. What a ride.
This is the paradigm shift nobody’s talking about. The bottleneck in AI-driven scientific discovery isn’t just smarter models—it’s human-engineered pipelines that constrain and guide the AI’s search space effectively. We’ve been asking the wrong question. It’s not ‘Can AI think?’ It’s ‘Can we build the environment where AI thinks better than we ever could?’
The skeptic in you might ask: ‘But how do you verify the counterexample? How do you know the pipeline didn’t introduce errors?’ Great questions. The beauty of the pipeline is that it’s modular, transparent, and open-source. Every step is logged, every assumption is explicit. The same structure that guided Claude to the counterexample can be audited, extended, and reused by anyone. That’s the point.
So what’s the practical takeaway? If you’re a researcher, a developer, or just someone who loves building things: stop trying to teach AI to answer your questions. Start building the pipes that let it explore the questions you haven’t even thought to ask. The future of discovery isn’t a smarter oracle. It’s a clever scaffold. And you don’t need a PhD in math to build one—you just need two years of obsession and the courage to share it.
FAQ
Q: What if the pipeline is flawed or introduces bias into the AI's search?
A: That's a valid concern. The key is transparency—every step in the pipeline is logged, open-source, and auditable. Flaws can be found and fixed, just like any scientific tool. The real danger is treating AI as a black box; pipelines make the reasoning visible.
Q: What's the practical implication for someone who isn't a mathematician?
A: You don't need to be a mathematician to build pipelines. This developer wasn't. The skill is in modeling logic, designing modular systems, and sharing them openly. If you can code and think structurally, you can contribute to AI-driven discovery in fields you're not an expert in.
Q: But isn't this just a fancy way of saying 'garbage in, garbage out'? The AI is still limited by the pipeline's quality.
A: Exactly the opposite. The pipeline doesn't limit the AI—it guides it. Think of it like a telescope: it doesn't limit what you can see, it focuses your view. A well-built pipeline expands the AI's reach by narrowing the search space to the most promising areas. The AI can still surprise you, as Claude did here.