Huh. Didn't hear about that, but I'm hard pressed for cases where I wouldn't value correctness over agent doing seemingly random changes in a loop (which still happens with Sonnet, and especially infuriating when I see it trying exactly same things it tried before, and just burning time and money).

It does seem to integrate a bit worse as a standalone agent than Sonnet, but I guess it's a matter of time.

When I need a surgical improvement / implementation, it works a lot better.

Yuppp, it's pretty unhinged one. Like most weird instructions, apparently motivated by cryptography.

This is also one of the reasons I strongly prefer "massaging" data to make it autovectorizer-friendly and only dropping down to raw SIMD when necessary as suggested by the profiler & generated asm.

Autovectorizer is pretty powerful nowadays and has lots of hidden goodies I wouldn't think of.

Reduced example of my code similar to where I found it - reversing bits in 8 u64s at the same time: rust.godbolt.org/z/PYjYoxeYz

I tried so many prompts, and eg ChatGPT outright refuses to stop asking "Do you want me to blah blah?" at the end of each response. Best case, it apologises, and in the next answer uses "If you want, I can blah blah." instead.

I don't need your random suggestions that ruin my train of thought!

And yes, I know about personalisation prompts, yet way too often models ignore them either because of the bias in the training data, the weight of the built in prompts, or because context window ran out and it forgot all about my personalisation.

As a user I shouldn't have to even know why.

Damn. Another variation of my question was going to be "how can you tell there's something wrong with that cuppa just from a picture" but this answer is so comprehensive it covers it too.

I'd ask what's wrong with it, but afraid to give Reform more arguments 😂

(of course, it can't be as accurate as function-level info, but it does help find severe outliers)

Ability to see timing per instruction (when choosing sufficiently high hardware counter precision) also doesn't hurt.

Getting there... Not on par with C++ original that uses handwritten SIMD but getting pretty close, and the microarchitecture usage is looking a lot better now.

Huh isn't that slow as hell?

Kinda sounds like the thing that Relooper solved for Wasm. Maybe reusable here too somehow 😅

LLVM to VCL when

Let me get this straight. Heat dissipation is the hardest part of space, AND the hardest part of data centers, so their solution is to put data centers in space.

All this time I thought it only supports dependency-free Rust and had to drop to either play.rust-lang.org or local cargo-show-asm otherwise, both of which don't provide quite the experience that godbolt does.

This makes it infinitely more useful.