yhjc2692

Programming Massively Parallel Processors is a classic (written by NVIDIA/CUDA engineers themselves), if you haven’t checked it out already

Don’t say we’re just gonna be boiling water with these?

I think there's a whole plethora of research on efficient attention implementations (similar to how DeepSeek & Kimi has made strides with theirs in LLMs) [1], [2]. And various techniques to reduce the flow-matching / diffusion iterations ("reflow") [3].

CVPR 2026 paper list (after filtering on video generation) might be a good start - asking LLMs to take a look and filter relevant papers sound even better.

[1] https://arxiv.org/abs/2609.15810 [2] https://arxiv.org/abs/2602.13515 [3] https://arxiv.org/abs/2605.23381

Wow this is really impressive. Throwing away the VAE feels counterintuitive - I'm shocked it works so well! Appreciate that you shared your training insights on the blog post!

Reminds me of this great paper where they trained a LLM that beats LLaMA2-13B from scratch with < $0.1M: https://arxiv.org/abs/2404.07413

This post wasn’t exactly what I came for (looking at the title) and a bit more stream of consciousness than expected, but nonetheless rings true. Enjoyed the bit on materialistic consumerism and Wall-E.

This is a very nice and intuitive explanation.

Here goes a terrible analogy: Maybe it’s like software. When I see issues and PRs pop up in my GitHub projects, I get the motivation to maintain, fix bugs, and upgrade them.

Indeed! I was a bit skeptical to see it is a single-author paper (thinking it might be one of those Claude-generated ones), and I appreciate that you pointed out this is somebody worth listening to.

Yep, the entire field of MEMS is both outlandishly complex and fascinating. For those who are curious, take a quick look at https://courses.washington.edu/me557/nthu/mems.pdf, which gives a nice overview of the different mechanisms and manufacturing processes involved.

"Neurons in an expanded mouse brain" (3rd place) is also stunning:

https://www.nikonsmallworld.com/galleries/2026-photomicrogra...

I didn't know brains actually look like this inside. Our world truly is beautiful, across all scales!

That actually sounds implementable even now! Except the fully consistent personality part. And maybe that's a good thing. Falling in love with beings without physical presence sounds agonizing; probably better spent on real people.

It’s incredible that you set up an in-house PCBA pipeline. I’m curious about what that looks like, and what you are building/manufacturing, if you don’t mind sharing!

I feel that as a field, we overfitted towards economic value maximization (which, to be fair, is the obvious thing to do) while leaving important problems behind.

Accessibility, sanitary infrastructure (in developing countries), generally improving health (including making it cheaper), and environment/wildlife preservation being some good examples of largely unsolved problems that technology still holds the key to dramatic improvement.

This might be the right time to start working on these things. Ironically, warfare, advertisement, and wealth concentration helped advance technology rapidly, making it easier to tackle other problems.

Woah, this is great to know. Thanks for explaining!

I would love to do that, but unfortunately don’t have one on hand (yet?). In fact, I’m curious - did you buy a TT device to work on this, or are you using a cloud provider like Koyeb?

This is super cool!

This is an interactive exhibition and a tribute to the habitats we've lost, by the architect Maya Lin (https://en.wikipedia.org/wiki/Maya_Lin) - who also helped design the Vietnam Veterans Memorial in Washington, D.C.

Note that this website was built way before the advent of LLMs (and is not "AI slop"), although it's been continually updated. I first saw it in 2022 at a lecture and was stunned, went on with life, and subsequently tried numerous times to find it again without success - until today, when ChatGPT one-shotted it from my vague description.

Hope this is an interesting find as it was for me!

Smaller target directory size would be such a lifesaver. Hoping the best for you, and Rust!

Agreed! I've been using Flutter + Rust (using https://github.com/fzyzcjy/flutter_rust_bridge , which streamlines FFI bindings management). It's a joy to work with and run on all major desktop/mobile platforms with little effort, minus system APIs like win32.

Interestingly, even the sound of the word ‘round’ feels gentle, while the word ‘spiky’ has high-frequency peaks (at p and k)! I wonder if dogs (or for that matter, humans) can accurately zero-shot classify words that represent round vs spiky for different languages.

It appears to me that onomatopoeia, at varying degrees, played a role in the spontaneous convergence of verbal communication, and thus the development of a language.

One thing you can try is reducing the encoder effort: libjxl produces something like 5% larger files, but at 3-6x faster encoding speed, if you turn it down from 7 (default) to 4. (The typical compression rate benefit vs. JPEG is 30-60%.)

Shameless plug: if you have GPUs in your server, you can put them to use to make it faster! https://github.com/yunho-c/gjxl (Granted, the speed improvement at lower efforts aren't as pronounced)

This is freaking awesome

For anyone interested, here is a recent 3Blue1Brown series on "Compression is Intelligence": https://www.youtube.com/watch?v=l6DKRf-fAAM

(For anyone not familiar with 3B1B, it's a YouTube channel that visualizes intriguing math concepts to explain them intuitively!)

Yeah, it's very cool that jixel was able to achieve such large efficiency improvements!

Haha, that's a great point! On that note - do you have any good pointers on understanding more about how Butteraugli works? I wholeheartedly agree it is a great metric (which aligns well with my subjective assessment much better than MSE or MS-SSIM), but the Guetzli paper (or the Butteraugli GitHub repo) does not explain the theory behind it in detail.

Seems like others were curious too, given the Wikipedia page on Guetzli mentions:

> Butteraugli. ... How the hundreds of parameters that model the properties of the human visual system were derived remains unexplained.

My rough understanding is its multi-scale convolution architecture was inspired by biological visual perception (perhaps how the layers of visual cortex gradually represent higher level features, ish), and involved some sort of empirical parameter tuning based on human evaluations. It'd also be a very nice resource for beginners wanting to learn more about image/video codecs (like me :)

This is a good food for thought. Semi-related: It reminds me of the fact that smartphone cameras nowadays use ML based filters to make faces prettier before saving in the first place.

Perhaps we can use clever experiment design to address it, like doing a "full-reference" evaluation (e.g., pick most similar two images, out of encoder A, encoder B, and original) rather than strictly A/B.

I feel that!

Maybe the first step to designing better encoders is to build an extremely accurate visuoperceptual error metric. Seems like ML's rapid progress haven't caught onto it just yet.

> libjxl's maximum effort is very slow. It produces files 2–7% smaller than its default for photos, and ~25% smaller for the screenshot and illustration, but takes 9–46 seconds, which is 50–180x as long as its default. jixel gets the same results in a couple of seconds.

libjxl's slow speed at efforts >7 is primarily due to an iterative evaluation-optimization loop of a perceptual error metric called Butteraugli (by tuning quantization weights). This, as well as some other computations in libjxl encoding, happen to be data-parallel and run faster on GPUs than multicore CPUs.

In case anyone is interested, I've been working on a GPU-accelerated libjxl port (supporting CUDA and Metal for now) that makes this much faster, closer to libjxl's normal speeds: https://github.com/yunho-c/gjxl - geared for encoding a large batch images into JPEG XL. It's also faster by around 2.9x at normal effort (7).

It'd be interesting to scale this up to a lot of images - and have users blind A/B test, like the LLM Arena on Huggingface which helped filter out "benchmaxxing" vs actual intelligence & utility. Because 1) images can be so diverse* and 2) an intrinsic tradeoff might exist between general efficiency and out-of-distribution robustness.

It's hard to rigorously judge a codec from a handful images, especially given the large impact they would have (with a widely adopted codec encoding bazillions of images in the coming years).

* There are many intricate details that go into codecs and visual perception. For instance, our sensitivity to visual errors are dependent on colors (all of hue, saturation, and brightness), spatial frequency of artifacts, local contrast and texture presence, et cetera.

And like OP mentioned, perceptual error metrics (including SSIMULACRA2) aren't perfectly representative of human-evaluated visual quality. For instance, libjxl deliberately makes choices that hurt SSIMU2 scores, but improve human-rated quality.

What would be a better name?

This is something that can be addressed by file managers - showing a metadata entry on whether it's lossy or lossless. File extension (along with the bifurcated mainstream adoption of JPEG and PNG) has been a great proxy to do this, but like others noted, can be deceptive, which is arguably worse.