michalsustr

Not a cuda programmer, but since they’re making a new API, why would they already make it inconsistent at start? :-/ I’m referring to the examples a,b,c vs z,x,y (different ordering of output elements)

If you'd like to explore what the Navier–Stokes blow-up construction looks like visually, I vibe-coded an interactive 3D visualization based on the published result for fun :) Demo: https://minfx.ai/navier-stokes/ Source code: https://github.com/minfx-ai/navier-stokes-blowup

If you'd like to explore what the Navier–Stokes blow-up construction looks like visually, I vibe-coded an interactive 3D visualization based on the published result for fun :) Demo: https://minfx.ai/navier-stokes/ Source code: https://github.com/minfx-ai/navier-stokes-blowup

If you'd like to explore what the Navier–Stokes blow-up construction looks like visually, I vibe-coded an interactive 3D visualization based on the published result for fun :) Demo: https://minfx.ai/navier-stokes/ Source code: https://github.com/minfx-ai/navier-stokes-blowup

If you'd like to explore what the Navier–Stokes blow-up construction looks like visually, I vibe-coded an interactive 3D visualization based on the published result for fun :) Demo: https://minfx.ai/navier-stokes/ Source code: https://github.com/minfx-ai/navier-stokes-blowup

Are the materials for the class on complex system failure analysis you had available online? Thank you.

Nice article. As perf optimisation has become more like a machine learning problem, maybe there is opportunity to use other techniques for checking generalisation. Like e.g. cross validation

Very cool! Also, huh interesting. I’ve used rdtsc to measure cycle diffs but had no idea its execution takes that long. Is that common across architectures?

Hmm so why did boxed() do type erasure in the first place?

I talked with Hochreiter about 1.5y ago they should go for it and properly scale xLSTM to LLMs, not just playing around with few B params for experiments. They felt like they are too late to the game to get investment and traction. So much for European ambitions I guess...

Very interesting. What do you want to do about async?

Congrats on the launch, this must have been an insane amount of work! I noticed the headers had 2024 in them, so indeed quite some time put in.

It seems the main selling point of micron is speed. You’ve done quite a lot of benchmarking. I think to make the case of micron stronger compared STL it would be cool to publish these benchmarks (where the two are comparable).

Fantastic presentation! Thank you for sharing.

Given general software quality of the js ecosystem, the proliferation of supply chain attacks was just matter of time. I’m curious how other ecosystems will hold (eg Rust)

What language are you building this in? I’m interested but trying to stay away from js world for security reasons.

Interesting. So one way to interpret current situation is that Github is "trapped" by its open source offering. This will likely have implications soon on what they do or the direction of open source...

Why don’t they just raise GitHub actions prices? Supply and demand, that would sort itself

I recently saw a submission here that does that, by essentially implementing GC in Rust. It is not beginner material though. https://kyju.org/blog/tokioconf-2026/

Edit: also the simplest way how to do cyclic structures is to heap-allocate via Box and leak memory. Box::leak This is also mentioned in the linked article.

Next step would be optimize for performance so it’s not slow. LLMs in Rust like to use cloning a lot , which will make it super slow.

Very interesting! Im curious, how does this work, it binary patches glibc allocator? AFAIK custom allocators are only in nightly and require generics in the form Vec<T, A>

Mine too! Rust is my favourite language right now.

The complications begin with async. Outside of async it’s a beautiful world.

With async, you tend to get locked down on library ecosystem level, with the dominant approach being work-stealing Tokio runtime, which I disagree with the fundamental design, after doing a lot of research. However, the gravity field of Tokio is strong. To escape it , I had to make a copy of popular crates and dig with LLMs to rewrite them to be free of work-stealing.

It would be great to include non-io-uring bench numbers as baselines.

This would be great if it compiled to wasm! Rust has great support for it. And you’d get native architecture for free too!

Very cool. How about using caps lock for switching modes? Or vim-style?

We might be interested in the technology (not as a VC), if it shows meaningful improvement (>2x) over what we currently do.

If you could show how it compares with a simple replicable baseline, maybe something like C program that just accepts a connection? Lots of these things are hardware-dependent.

Right now we use monoio and have a draft benchmark with speed. Happy to continue talking over e-mail. Should I write to you?

Interesting. How does your benchmark look like if you compare with monoio?

Btw this blog series is a fantastic read: https://en.ihcblog.com/rust-runtime-design-1/

Hi, nice motivation! I’ve built async runtime driven by clocks on top of monoio. You can drive each thread at different speed, to simulate a distributed system faster than real time. Our motivation is outlined here: https://minfx.ai/reliability.html

It’s not published yet, as it’s a bit wired to our internal systems at the moment. But happy to chat :)

Very interesting. What is the runtime model, based on the "async trilema" https://without.boats/blog/the-scoped-task-trilemma/ ?

I assume tokio-like, i.e. work-stealing?

Very cool! I’m exploring adding DICOM support to https://minfx.ai project. The idea is to better support machine learning in medical space and make DICOM viewing in web assembly. Not right now, but in next 2/3 months. What you’ve built looks very impressive! Are you planning to release Linux binaries? Happy to collaborate!