Claude finds quantum circuit designs that beat the best published ones

Claude cut ops by up to 17x and 12x here by rebuilding temporary data and swapping in a better algorithm. Impressive for a closed loop setup, honestly.

Feels like the universal paper clips meme, where AI builds quantum computing to make itself smarter. Also ASI is coming long before usable quantum computers are, so it's a bit moot.

I checked the arXiv PDF and the acknowledgments only say Andi Gu did part of the work under the STEM Fellows Program, with Harvard funded by IARPA, ARO, and NSF. The whole Anthropic angle looks made up to me.

This builds on Google's Kickmix quantum language, which is basically an assembly language for classically simulatable circuits. Optimizing it is code golf on assembly, so LLMs do well because they can test and toss ideas fast.

Envelope principle: AI is great at filling in blanks inside what's already known, but real progress comes from expanding the envelope itself.

The ellipsis thing is fine, the article says plainly which optimization drops the rare-input handling. That's a standard trick, and the compute part is what gives the speedup over an order of magnitude.

Weird they don't mention Bennett (1973), that seems like the obvious reference here.

I'm normally more of an AI hater but this just seems objectively good, so I'll ignore it and move on.