I searched for "5.1" in your blog post, and the only hits are in the benchmark numbers.

I agree with GP: without a comparison to Opus 5.5 or some other models, it's not clear how much of an advance this is. If anything, my experience is that performance optimization in existing codebases is something existing models are already great at: it's measurable (you directly profile or try experiments to see what works) and the code architecture is usually relatively fixed (at least relative to deep/greenfield work where there are large architectural questions to be addressed). Newer models might be better at generating ideas for things to try, but if you're getting 36x then it honestly sounds like you have a lot of low-hanging fruit.

Hey for once I feel qualified to reply on a HN thread (I'm one of the developers of DFTK, a similar package). Very nice result. I'm not particularly surprised, as QE is not particularly GPU optimized. I've heard that the QE developers interested in GPU are doing it in Sirius, but I don't know much about it. More generally there's a tension between new features / optimizations and maintainability. I'm not sure how to deal with this medium term. Perhaps having two code bases, one for humans and the other for AI, is not the worst solution.

This sounds very much in line with most reports from people getting early access to models over the last years. Maybe it actually is a lot faster, but it may as well be the load on it not being very high yet. I'll keep my excitement until it has been adopted by the masses.