Robert Knight

@geofflangdale Does "everything and anything else" include changing the behavior of the program (outside of undefined behavior)? IME that is the major constraint that compilers have compared to humans. Human optimizations often change behavior slightly (but in ways that don't materially matter)

predict the future - i.e. it can't look at a program and magically infer facts about unknown inputs - but it can do everything and anything else. It's interesting in that there are almost certainly cases where the 'God Compiler' is only fractionally faster than what we have - ..

Here's a question: "How much faster code does the 'God Compiler' produce"? That is, suppose there's a compiler that's "optimal" and can spend as much time as it likes. Assume it can aggressively restructure whole programs with as yet unknown optimizations. It *cannot* ...

@lemire Do you know if any of the cloud vendors are going to offer Arm chips with SME/SME2 any time soon?

Update: Microsoft is adopting Neoverse V3 later this year. This means that Microsoft and Amazon will offer powerful SIMD-enabled cores in 2026 (so lots of compute), whereas Google goes for weaker SIMD processing (less suitable for compute).

@pshufb What are the main issues with it?

When a CPU architect titles a presentation "RVV - why it is the pain in the ..." I may tell my gut feelings about RVV were right. :)

@ashvardanian Did you find a good policy for when to use SVE vs NEON, if SVE isn't always a win?

Good news: I finally SIMD-ified several UTF-8 string operations — seeing up to 10× gains. Wild that most libraries still don’t optimize for UTF-8, given it’s the dominant format in the HTML corpora Frontier AI labs train on. Bad news: SVE(2) is so underpowered on many devices that NEON often wins even when SVE is available 🤦‍♂️ Some Zen5 cherry-picks for UTF-8 tokenization: • Splitting on 8× 1–3-byte line delimiters (Chinese): 16.65 GiB/s vs 1.93 GiB/s (serial) • Splitting on 25× 1–3-byte whitespace chars (Korean): 1.88 GiB/s vs 0.98 GiB/s (Rust std) More workloads are coming. Until then: `pip install stringzilla`, `cargo add stringzilla`, or CMake-fetch it if your pain tolerance matches mine 😅 github.com/ashvardanian/S…

@gvanrossum This opens in Keynote now. Thanks!

@robknight_ Thanks for the report! Should be fixed now on the release branch. Will be fixed on main a little later.

Slides from my PyBay talk today can be downloaded here: github.com/microsoft/type… (PowerPoint only for now).

@tmikov @awesomekling @dhh One change from 15 years ago is that social media engagement-baiting has become much more optimized, so this kind of drama tends to get spread and amplified more.

@awesomekling @dhh Was I super naive at the time, or was there no politics in Open Source 15 years ago? The FSF came closest to to politics, IMHO, but I don't remember it taking any positions unrelated to software.

I stand with @dhh because they’ll come for me next. Eventually they’ll come for you, too. It’s time we put our feet down and say “no” to these losers who demand everyone share their exact opinions or get smeared, fired, etc. Let’s usher in a new era of OSS: one where everyone is welcome, and nobody tells other people what they must think and say.

@bcrow Thanks, that looks good!

@robknight_ My thoughts and prayers! github.com/robertknight/t…

Someone sent a patch for my JS implementation of the TeX linebreaking algorithm, and now I'm stumbling through the original TeX source (a 25K LOC file called tex.web) to try to figure out how it handled a particular situation. Send thoughts and prayers.

Bad news -- I'm not going to be with my employer for much longer due to team relocation demands. If anyone has any leads for roles that would allow me to continue my Rust compiler work (in New York City), they'd be greatly appreciated.

@mitchellh Which project(s) were you compiling that prompted the tweet?

Is the Rust project doing anything about the insanely slow compile times? I see various blurbs about language features, build tooling, lib enhancements, but I rarely see the same activity around compiler speed. Genuine question from an outsider, these compiler speeds are killing me as a consumer, I can't imagine what it's like as a developer.

I wrote a progress report / retrospective on the RTen machine learning runtime in 2024, looking at the added features and progress towards performance parity with more mature runtimes: robertknight.me.uk/posts/rten-202…

@laurencerowe @jordwalke My FB experience and X experiences have been similar. FB used to be a useful place to see what my IRL friends have been up to. Recently it has been full of suggested content I didn't ask to see.

@jordwalke I feel like Facebook did the same thing. Used to go there to see what my friends were doing and now I scroll down a page or two and it’s just ads and algo. suggested content and I have to view individual profile pages to see anything. Just ends up in a downward spiral.

People leave X for another more controlled platform in hopes that doing so will grow that platform, establishing it as the new town square - one that is more controlled. Not so they can speak more freely, but so that you cannot.

RTen v0.14 has been released. It adds examples for Whisper speech recognition and RMBG background removal, plus compatibility with tokenizers exported by the latest version of Hugging Face Transformers - #0140---2024-10-27" target="_blank" rel="nofollow noopener">github.com/robertknight/r…

When starring a repo, it would be cool for maintainers if users could provide context on why it is interesting ("I really need a FooBar implementation for my XYZ project")

The onslaught of unhinged political content in my Twitter feed got a bit much over the weekend. This prompted me to take another look at Bluesky (robertknight2.bsky.social). My first impressions are quite positive.

@tonofcrates The use of MathJax for rendering the HTML version has the unfortunate consequence of breaking copying and searching text amongst other things (github.com/WebAssembly/sp…, github.com/WebAssembly/sp…). I really wish that had been given higher priority.

Comparing the sentential and symbolic representations of inference rules, you can start to appreciate the value of specialized notation. (Also, rather than a new Greek letter for each type of context, just use a natural language subscript!)

Periodic reminder that we have an example of a high-quality language spec that is broadly accessible to both math-oriented and systems-oriented developers: WebAssembly! webassembly.github.io/spec/core/vali…

Additionally there is the possibility to specialize based on the static argument types, making slicing fast enough to use inside operator kernels.

I've been working on safer and more ergonomic slicing APIs for the next rten release using type-level arithmetic. chw_tensor.slice((0, .., 0)) // Statically infers result is 1D nchw_tensor.slice(0) // Statically infers result is 3D

I appreciate the Street Fighter analogy at the end.

Very easy way to improve the quality of your work: whenever you put up a PR, take a few seconds to skim through it yourself just as a reviewer would (i.e., in the GitHub UI). I find a surprising number of small issues this way that never stand out in the editor.