Let me explain what just happened.
Levent Alpöge, a mathematician at Anthropic, says he has proven the Riemann Hypothesis while working with Claude.
The Riemann Hypothesis (a 167-year-old prediction about how prime numbers are spread out) is one of the seven Millennium Prize Problems, with a $1M reward. It underpins a huge amount of modern math, and it's tied to the encryption that keeps the internet secure.
He says the same method also proves the Generalized Riemann Hypothesis, an even broader version. The whole proof is written in Lean (a programming language where a computer checks every single step of a proof). The full written-out version reportedly runs past 250,000 pages, so he released the Lean code instead.
The work started on October 6.
8 of the 20 key steps landed on a single day, November 15. He openly credits Claude for the parts that rely on areas he says he hasn't mastered.
Why it matters: problems like this used to take generations of mathematicians, if they got solved at all. This is one mathematician and one AI model in about six weeks.
So don't misunderstand it as "AI solved math." Nobody has independently verified it yet, and that has to happen first. But the Lean format means it can be checked much faster than a normal paper.
Still crazy!
If it turns out to be true and the Riemann Hypothesis really has been proven, then today history has been made. I am in shock.
But even more interesting is how fast it was solved and how much Claude did:
The first steps date to October 6. Then 8 of the 20 steps land on a single day, November 15, and the last entry on November 16 just says "GRH. (Generalized Riemann Hypothesis)"
He admits he has little mastery of one area the proof depends on (automorphic forms on higher rank groups) and credits Claude's technical skill for those parts. A written-out version ran past 250,000 pages, so he only released the Lean code.
So it actually needed one mathematician, one model, six weeks.
I expect official statement from Anthropic and Alpöge today.
Dario Amodei reveals Anthropic discovered what looks like a new gene-editing mechanism and says it may discover dozens a month within months
“I’ve been talking to our biology team several times a week. We discovered what looks like a new gene editing mechanism, and I think within a few months we may be discovering dozens of them every month. We’ve seen, with the speed at which this technology moves, these little sparks will soon turn into a firestorm.”
“I think that we can find a way to build this technology safely, in a way that benefits people, that addresses all the risks, and that delivers a world that is so much better than the world we have today. We haven’t done that yet. We haven’t earned the right to talk about how great those things are, but we’re starting down the path.”
Almost a full year ago in November, I decided to take a leap and visualize the math, physics and machine learning ideas I love.
Whats you favourite visualization so far?
https://t.co/jlQAOvn35h
Dario Amodei explains why he'd rather have a team of 100 exceptional employees than a team of 1,000 where only 200 are exceptional:
“If every time someone super talented looks around, they see someone else super talented and super dedicated, that sets the tone for everything.”
“Everyone trusts everyone else. Then you need to put a lot of processes and guardrails in place just because people don't fully trust each other.”
“It's very easy to go from 100 to 1,000 and 1,000 to 10,000 without paying attention to making sure everyone has a unified purpose.”
Today, we're launching Antinuous (@ycombinator P26), a research and experimentation platform to run 100x more experiments and simulations, 1,000x cheaper.
We combine swarm orchestration, compute clusters, and hardware-tuned simulation software to give every agent access to primitives for running thousands of experiments in parallel, and we're working on scaling this to millions.
In testing, we've built a full digital twin of an F1 car, more efficient datacenter cooling with an early partner, and a fine-tuned classifier that we use to help route requests to the right model, all within hours.
With hundreds of thousands of agent runs across just a few weeks of beta, we can scale this up to make the data flywheel for scientific research larger than anything publicly available in the space.
Today, we're launching in public beta. Try us and give us feedback at https://t.co/ZbJvGwp84I