Tech • AI • Robotics • Game
Mythos 5 and Google Just Broke Quantum Computing (Major Breakthroughs)
9.4/10TL;DR
Anthropic reported that Claude Mythos 5 discovered quantum circuit designs that beat published human baselines on seven of eight benchmark tasks, while Google Quantum AI separately reported a practical breakthrough in reducing quantum error rates for magic states.
KEY POINTS
Automated search beat published baselines
In a paper released on October 15, researchers in Anthropic’s STEM Fellows program said Claude Mythos 5 was given plain-language descriptions of eight quantum circuit primitives and asked to generate candidate circuits in Python. A separate software evaluator, not another model, checked correctness and scored each design by qubit count multiplied by gate count, a proxy for runtime and resource cost. Anthropic said the system improved on the best published designs in seven cases and produced a new formulation in the eighth.
Why quantum circuits fit this approach
Quantum circuit design is unusually amenable to machine-guided search because correctness can be checked cheaply and automatically. Circuits must also obey constraints that differ sharply from classical software: they cannot rely on ordinary branching, and intermediate data cannot simply be discarded because operations must remain reversible. That creates hard tradeoffs between qubits and gates, making the field well suited to iterative propose-evaluate-revise loops.
A major result on modular inversion
The headline result involved prime-field modular inversion, a costly step in elliptic-curve cryptography. Classical implementations use divsteps, but those procedures appeared impractical for quantum machines because they generate branch history that would normally need to be stored. Anthropic said the model instead derived a way to reconstruct that history from final results, proved new algebraic identities, and formally verified the decoding rule in Lean. One resulting design used 12% more qubits but 17 times fewer gates than the smallest previous circuit.
Point addition and formal verification
For elliptic-curve point addition, Anthropic said a single 53-hour run produced a new design from a brief initial prompt and occasional guidance. The system reportedly proved the relevant algebraic identities, tested them through large numerical simulations, and formalized the claims in Lean. That combination of search, proof, and verification made the result more than a heuristic optimization.
Decoder redesign cut a dominant cost
Another standout result involved decoding quantum interferometry, an optimization method whose most expensive step is a Reed-Solomon decoder. Instead of extending the branching-heavy extended Euclidean algorithm, the model switched to Berlekamp-Massey decoding, which performs the same operation each round and better matches quantum constraints. Anthropic said the new design used a simulator to study 43 million sample decodings and concluded that reserving just 19 qubits for rare outlier cases was sufficient, producing a circuit with 12 times fewer gates than Google’s earlier construction.
Hardware limits make these gains meaningful
Quantum computers remain highly error-prone, with current machines offering only hundreds to low thousands of qubits and suffering from disturbance from heat, vibration and radiation. Because long computations accumulate errors, reducing gate counts and memory demands is central to any path toward useful applications in chemistry, materials, cryptography and optimization. Anthropic’s results matter mainly because they target those bottlenecks directly.
Caution on the biggest claims
The reported gains have not yet been independently reproduced. Anthropic’s own writeup emphasized that the dramatic multiyear reductions in estimated cost for simulating FeMoco, the iron-molybdenum cofactor central to nitrogenase, were the product of many research groups over several years, not one model run. Some online summaries also overstated the outcome by saying all eight tasks were wins, when the paper describes seven better scores plus one new formulation.
The method may matter more than any one circuit
Anthropic also described a failure mode: the model initially searched too conservatively and tended to stay close to established approaches. The fix was structural rather than prompt-based, running several qualitatively different search tracks in parallel and occasionally nudging exploration. That architecture, combining a proposer with an independent checker and a numeric objective, resembles broader agentic AI systems driven by test suites rather than fluent output.
Google reported a separate quantum advance
On the same day, Emma Rosenfeld and Craig Gidney of Google Quantum AI reported in Science that magic state cultivation worked in practice on the Willow processor. Magic states are essential for universal fault-tolerant quantum computing but are notoriously expensive, with some estimates suggesting they could consume up to 95% of a machine’s capacity. Google said the method cut errors by about 40 times and achieved a record logical fidelity of 99.99% for T states, while successfully transferring purified states from a distance-3 color code to a distance-5 surface code.
CONCLUSION
The two results address different bottlenecks in quantum computing: one uses AI to shrink algorithmic resource costs, while the other improves the quality of the fault-tolerant states those algorithms require. If both lines of work hold up under replication and scaling, they could materially shorten the timeline to practical quantum machines.
Ask a question