New blog post: Stop over-engineering static websites apoorvaj.io/stop-overeng...

https://mitchellh.com/writing/zig-builds-getting-faster

I feel that Claude Code + Sonnet writes way more tasteful Zig code than it does JavaScript. JS is littered with dumb defensive programming, with null checks or conditions where here should be asserts. My theory is that this is because of the delta in the quality of Zig vs JS human training data.

Insanely well-made video on EUV photolithography: www.youtube.com/watch?v=B248...

New blog post: Testing that works apoorvaj.io/testing-that...

This was an easy one to implement. Then I started on white balance, fell into a rabbit hole of black body radiation and color spaces, and then noped out for a bit.

Holy crap, #ziglang 0.14 -> 0.15.1 is quite the breaking change. This is literally 5% of @papaya.io's Zig codebase.

I'm really proud of this. 1. Unlike many other web-based image editors, papaya.io rasterizes text in WebAssembly. This means that text rendering looks the same across browsers or OSes. 2. This means that I had to re-implement a text box, which was not fun. ...

A new article on my site: Approximate first principal component It describes simple trick to estimate the direction of most spread of a bunch of points without running a full PCA. Includes my Python re-implementation and plots of some quick tests. 30fps.net/pages/approx...

Zig's New Async I/O https://kristoff.it/blog/zig-new-async-io/

The Zig 2026 roadmap is absolute fire. Andrew and team are playing the long game, and it's really starting to pay off big time—incremental compilation, async await, cross compilation, fuzzing, and more. www.youtube.com/watch?v=x3hO...

Mikko and I added MacOS-style keyboard navigation to the Skribidi text library. github.com/memononen/Sk...

*Tin foil hat* Has Google added a special case for Google Docs in Chrome to show the MacOS accent menu in the right spot? It renders out of the window in Firefox, which is what you'd expect to happen when using a hidden textbox to get input. I cannot find a web API do this. Looks like an inside job.

Still long way to go, but got the beginnings of IME text composition working in papaya.io. I'm super impressed that @mikkomononen.bsky.social's Skribidi library even has this feature. Here's a video of me typing in '⌥-u' and then 'o' on Mac, which results in 'ö'.

Got simplified Chinese and FreeType variable font rendering to work in papaya.io.

Multi-script rendering in papaya.io is starting to come together. Here's glyph caching and retrieval across several fonts (including an emoji font)!

Added intermingled rendering of Latin and emoji (from different font files) to papaya.io. (Also look at that automatic word-wrapping and left-alignment, courtesy of Skribidi!)

Emoji font rendering in papaya.io with @mikkomononen.bsky.social's github.com/memononen/Sk... library. Next up: intermingling font and emoji rendering.

I just published the bidi text stuff I have been working the past months. It's called Skribidi (because all text and font related things need funny name for some reason). github.com/memononen/Sk...

Started implementing text rendering for papaya.io. Got FreeType and HarfBuzz compiling to WebAssembly. Here's a demo of kerning working with Roboto.

Just relaunched papaya.io — a web-based image editor I've been building on nights and weekends. What's new: - .ppy file format for local saves - Automatic browser persistence - WebP import/export support Building fast and shipping features regularly. Try it out — feedback very welcome.

An update, a bit of retrospection, and a call for you, yes, you, to help JCGT a bit: www.realtimerendering.com/blog/jcgt-is...

I wrote a blog post about mipmap level selection. pema.dev/2025/05/09/m...

These pruned trees enable efficient evaluation of the SDF for rendering or discretization, resulting in >100x speedups for complex scenes compared to brute-force evaluation. Come to my talk next week for more details, or read the paper here: wbrbr.org/publications...

I'm glad somebody out there is brave enough to push back against the "personal ChatGPT usage is terrible for the environment" message andymasley.substack.com/p/a-cheat-sh...

Prematurely quoting papers you clearly did not read is the root of all evil.

File Pilot just came to open beta today, and it already feels way better than Windows File Explorer. filepilot.tech

This year, I'm trying to push back against tech enshittification. - Deleted my Twitter account - Deleted my Instagram account - Switched to Signal where possible - Switched to Obsidian + Syncthing for mirroring my notes and files across devices - Self-hosted visit counter for my website

Delve into the creation of the main character featured in Unity's Time Ghost, a real-time cinematic crafted with Unity 6. Watch here: shorturl.at/BvN5S Download a sample of the character, including the ML model from this link: shorturl.at/9E5F5 #Unity6 #Gamedev

NVIDIA Blackwell whitepaper is available: images.nvidia.com/aem-dam/Solu...

DeepSeek R1 is the first time I've felt a computer be actually intelligent. Given a math problem, it uses inference-time scaling to try thinking through various angles in a way I've never seen before.

Zig 0.14.0 release date has been set to mid February! https://ziglang.org/news/what-to-expect-from-release-month/

Spotted some DirectX code in the movie Mars Express. www.imdb.com/title/tt2691...

“We’re going to be the social network that lets you - encourages you - to say slurs”

You cannot get physically-based rendering right without understanding radiometry. My new blog post explains it all, relying on familiar concepts from rendering algorithms as much as possible. Part 2 will come next week. momentsingraphics.de/Radiometry1B...

Aaaand Twitter account deleted.

New blog: How to make sense of projection matrices, clip space and normalized device coordinates. Learn about which spaces GPU hardware cares about. This started out as a personal note, but I kept referring back to it, so I'm publishing it. apoorvaj.io/ndc-clip/

PyTorch is now easy enough to write (I've been practicing for a month), but I'm still lacking intuition when my training converges to a bad result. Wondering if there's any systematic literature on choosing loss functions, regularization, etc.

ghostty.org is out! It's the only third-party terminal emulator that has Just Worked™ for a terminal pleb such as myself. I don't use tmux or anything else, and I need mouse input and the usual OS keyboard shortcuts. It's very fast, too, which doesn't hurt.

I didn't see any post related to Timothy Lottes notes about GPUs here. Let's share a link : docs.google.com/document/d/1.... This is a gold mine of quirks/information/thoughts for optimizing/developing on various GPUs.

The light captured by a digital sensor is an integral, but not exact grid corresponding to the final pixel grid because of the "fill factor": photonlines.co.uk/wp-content/u... On non-BSI CMOS it is just 30-50%, BSI 70-90%. And then there are microlenses, which complicate it even more! 1/

"Moving Frostbite to PBR" is 10 years old already but still an amazing physically based rendering and lighting resource, so much knowledge packed in there. seblagarde.wordpress.com/2015/07/14/s...

✅ Toy 2, differentiable rendering ( #PBR + #ML ): Learning the diffuse and roughness map. This was way harder than learning just the diffuse map given a fixed roughness, and I learnt a LOT along the way.

Finally online, the brilliant talk Pondering Orbs by Mikkel Svendsen on the rendering of the similarly brilliant game Cocoon youtu.be/8kS17jt38XQ?...

Software(+robotics) defined optics:

😂 Slang interface docs read: "Use the accumulated experience of the programming language community to ensure that our generics are humane (in other words: not like C++)" github.com/shader-slang...

SamplingLab : a small tool to live code and visualize samples aduprat.com/tools/sampli...

PyTorch needs stricter NaN handling than HLSL. In HLSL, you can often do NaN-alicous operations like sqrt of negative numbers, as long as you clamp NaNs away in the end. In PyTorch, you need to do clamping early, otherwise you get a NaN gradient in backprop.

✅ Toy 1, PBR + ML: Learning the diffuse map Ground truth: 128x128 diffuse map fed into disney diffuse + ggx lambert Inference: 128x128 weights map fed into same model. L2 loss, learnt on a laptop GPU in 13 seconds. Visually identical results.