An AI Tried the Riemann Hypothesis. It Didn't Win, But It Did Something Weird

A Swing and a Miss? Not Quite.

Anthropic's research version of Claude just took a crack at one of math's oldest puzzles, the Riemann Hypothesis. It didn't knock it out of the park. But it wasn't a total strikeout either. While trying to prove the hypothesis, the model managed to lift a related lower bound from about 41.6% to 67.2%. That means it showed at least 67.2% of the relevant zeros sit on the critical line—a noticeable jump, even if the full conjecture is still out of reach.

For those who haven't thought about this since college, the Riemann Hypothesis says all non-trivial zeros of the Riemann zeta function lie on the line where the real part equals 1/2. It's a big deal in number theory, with deep implications for how primes are distributed. Mathematicians have been chipping away at it for over 150 years. A full proof would be a landmark. Claude didn't get there, but it did something arguably more interesting: it discovered a new angle to attack the problem.

What Claude Actually Did

Here's the behind-the-scenes. The research Claude tested roughly 650 different ideas. It coordinated about 60 smaller AI agents to handle numerical checks, search through papers, and verify arguments. Two rounds of research consumed around 31 million output tokens. That's a lot of digital ink, but the real breakthrough came from recombining existing mathematical techniques in a novel way—not inventing a whole new theory from scratch.

Anthropic's internal mathematicians have checked the core arguments. The result was also formalized in Lean, a proof assistant, and validated with verification tools. A couple of number theory experts gave it a preliminary once-over. But the work hasn't gone through the long, rigorous scrutiny that comes with formal publication. Anthropic is careful to note they don't think this approach can directly prove the full Riemann Hypothesis—yet.

Why This Matters (Beyond the Math)

The Riemann Hypothesis is a monster. It's been taunting mathematicians for generations. It won't fall easily. But this progress shows AI can be a useful partner in math, not just a calculator or a theorem-proving machine. The way Claude coordinated multiple agents, tested dozens of hypotheses, and verified results gives a glimpse of how future mathematicians might work alongside AI.

It also reminds me of something from my own creative work. I've spent hours on a painting that turned out terrible, but in the process I discovered a color palette I ended up using for years. That's what happened here. Claude was trying to prove the conjecture, didn't, but found a better bound. That's not failure—that's a win, just not the one you aimed for.

The Creative Side of Problem-Solving

This episode highlights something we often forget about math: it's deeply creative. Finding a proof isn't just about following rules; it's about imagining new possibilities, combining ideas in unexpected ways, and having the persistence to try 650 different approaches. That's the essence of creativity, whether you're in a studio, a lab, or a server room.

For artists and crafters, there's a lesson here. The creative process is messy. You don't always get the result you want. But every failed attempt teaches you something, and sometimes the byproduct is more valuable than the intended goal. It's like trying to paint a perfect sunset and ending up with an abstract composition that sells for thousands. The journey matters as much as the destination.

Don't Hold Your Breath for a Proof

So, will AI crack the Riemann Hypothesis anytime soon? Probably not. Anthropic is clear that this method isn't a direct path to a full proof. The problem is too complex, and there's likely some missing insight that no one has found yet—human or machine. But this experiment shows that AI can make meaningful contributions to mathematical research, even if it doesn't solve the big ones.

For now, mathematicians can enjoy the improved bound. It's a small but real step forward. And for the rest of us, it's a fascinating glimpse into how AI might change the way we do creative work, whether that's proving theorems or making art.

Beyond the Riemann Hypothesis

This isn't just about one problem. It's about the potential for AI to assist in creative and intellectual pursuits across the board. We're seeing AI generate music, write stories, and now make mathematical discoveries. The creative potential is enormous, and we're only scratching the surface.

But there's also a cautionary note. AI isn't a magic wand. It needs direction, data, and a clear goal. The researchers at Anthropic gave Claude a specific problem to work on, and they guided it through the process. That's the kind of collaboration that yields results—human creativity combined with machine processing power.

As we look to the future, it's exciting to think about what other creative breakthroughs might come from this partnership. Maybe AI will help us unlock the secrets of the universe, or maybe it'll just help us make better art. Either way, we're just getting started.