This is one of those AI stories that isn’t about a chatbot at all.
What Anthropic did
On October 16 Anthropic published research where Claude Mythos 5 was given plain-language specs for eight quantum circuits (chemistry, optimization, physics simulation, number theory and numerical computing) plus an automated evaluator that checks each design and scores it on qubits times gates. Claude ran in a closed loop: propose, get scored, improve, with no human input between rounds.
- Across the eight tasks, Anthropic says it found designs that improve on the best published circuits, using fewer qubits and fewer gates, in some cases by an order of magnitude or more.
- For elliptic-curve point addition, after a 53-hour run, it came up with the first reversible version of the “divsteps” modular inversion method that doesn’t store branch history. It proved the math and checked it in Lean. One design uses 12% more qubits but 17x fewer gates than the smallest previous one.
- For the decoder inside decoded quantum interferometry (DQI), it switched to Berlekamp-Massey decoding and cut gates about 12x versus Google’s previous construction.
- The work was led by Andi Gu in Anthropic’s STEM Fellows Program, with Harvard’s Susanne Yelin and Mikhail Lukin (Harvard and QuEra).
Anthropic was upfront about the limits, too. Early on Claude was “conservative by nature,” second-guessing good results and sticking to familiar approaches. They had it develop several different approaches in parallel and sometimes nudged it to think bigger.
My take
The pattern is the part to steal: a clear spec, an automated checker and an honest score, then let the agent grind. That’s the same loop that makes coding agents useful, and it works anywhere you can write a verifier. One thing to keep an eye on as a security person: point addition is the core of quantum attacks on elliptic-curve crypto, so cheaper circuits there are one more reason to keep post-quantum migration on the roadmap.
Have a blessed day! And Happy Hacking!