‼️ = Must read!
🧐 = Interesting, curio.
'AI' - Machine learning, deep or not.
- Qwen3-Omni. ‼️
A multilingual model for text, images, audio, and video. People will need to study the architectural improvements here!
- LLM-based reasoning using Z3 theorem proving.
See also: Kevin Buzzard - Where is Mathematics Going? and unstable singularities in fluid equations.
- Double Descent - Welch Labs. ‼️
The problem with learning deep learning deeply, is that many concepts of machine learning, even very well established ones, seems to create completely different regimes when going to a lot of dimensions and depth.
- Video stabilization via Gaussian Splatting. 🧐
Such a neat idea. From video to GS - via ML of course, and then from GS back to video, stabilizing the camera path. No cropping needed!
Demoscene corner!.
- NoCpu challenge/demo.
Because programming the Amiga was getting boring I guess...
- The Demo Scene is Dying, But That's Alright.
I don't know... this article is 100% right, at the same time, I imagine it depends on what you want to call the "demoscene" - the spirit of hacking and creating audio-visual "demos" through code is strong, perhaps numerically, stronger than ever. But it's mainstream now, you can see such art in galleries, in concertes, on the web. Often used commercially. It's not the same. But all underground movements eventually become mainstream, right?
- The violators - c64.
The c64 keeps being amazing.
- madwizards - Fabula.
Hard for full-size demos to impress these days. This is pretty cool, giving some ASD vibes.
- A small shadertoy by Bluespoon.
Retrocomputing (might contain emulation).
- Breaking down the visual effects of Chrono Trigger. 🧐
Using emulators many youtubers are explaining the rendering tricks of old games, demos and so on. I really wish we could see this on "modern" console emulators (ps3/ps4 era) - as you can't mae a GPU capture of a third party title even if you have a devkit... But last time I checked the debuggers in these are still very primitive, not equivalent to what you had during development on these consoles.
- Emerald Source Code Commentary. 🧐
An (online) book on the (reverse engineered) pokemon emerald source-code.
- Custom-made voodoo cards. 🧐
I mean - enough said. It's also fun to see a time where memory was not holding... code, so you can remove chips and things don't crash.
- Amiga renderfarm.
Someone made a amiberry (rPi) based multi-node renderfarm for lightingware on amiga... in 2020.
- Cubix for the Spectrum.
A nice game for the speccy? Doesn't happen often.
- History of QNX.
- How cycle-perfect 8088 was achieved. 🧐
...and it seems that 286 and 386 could be next. One of the presentations at Vintage Computer Festival Midwest this year. Lots of interesting stuff there. Llvm on 6502 is also cool.
- Rendition Verite wrapper.
Can now play quite a few dos games in high-res, high-framerate. Very interesting. Should add pathtracing :)
Miscellanea.
- #rtr Drawing a triangle on a GPU, without APIs. 🧐
On an AMD r500 - not to be confused with the 500 series - like most of AMD gpus it's controlled via PM4 packets, but this is way before GCN era, so the shaders are not "scalar", it's VLIW, somewhat similar to the xbox 360. Fun!
- #industry Hot Chips 2025 RDNA4 presentation.
dynamic register allocation is getting popular! Afaik Apple does it too (they have a fully virtualized file) - I'm surprised that they went for register ALLOCATION (start with a minimum, request more registers, busy-wait on failure) which is the hardest case. Back in the days, I suggested to all vendors that would hear, to at least implement de-allocation (start with high watermark, release registers if the shader determined the expensive paths did not need to be taken). Of course in either case you need some level of virtualization otherwise you'd fragment. Floating point SALU is interesting. The split barriers I guess are there to help raytracing/workgraph stuff? I haven't given a lot of thought of how the new workgraph API is going to be implemented in HW, on older chips I don't see any way to do it efficiently. It feels like this plus the register allocation stuff might be the answer here - which would mean that it's not done through changes to the wave launch/scheduling instead. Dunno.
- #industry Game size vs Team size.
- #coding Matt Godbolt's ACCU'25 keynote. ‼️
Teaching an Old Dog New Tricks - great talk by Matt, a programmer that seems to share much of my worldview on code, here taking time to learn the latest features of C++ and compare the experience of writing a ZX Spectrum emulator with and without them. Worth your time.
- #coding Timothy Lottes teaches C. ‼️
Lovely stuff. Of course - please don't do this in a team - but for your own coding, great! Basically "downgrading C", less types, just ints of different sizes and pointers - whilst at the same time adding more explicit intent around all pointer access (restrict, volatile), all functions etc. Sounds fun - I'd be tempted to do this with TCC as well for fast live-coding (albeit this style is surely going to compile fast even with regular GCC...).
- A web image search using CLIP-like similarity.
- Gemini/Gopher for the playdate.
I love my playdate, but don't use it much. I love gemini/gopher, but don't browse it much. I suspect this gemini/gopher client for the playdate might improve my use of both things.
- Kowloon walled city...
...in minecraft
- Autolume.
A no-coding generative AI system allowing artists to train, craft, and explore small models.
- Vissoft2025.
I ❤️ software visualization.
- Nine things in ninety years.
I wish to be this wise.
- A deep dive into latency problems on Asus ROG laptops.
Learned a few things from this. Latencymon might be added to my "tools that i use" page. Sent from my iPhone
- Hosting a website on a disposable vape.
- Kissinger's thesis.
Rendered via a vibecoded web page
- A chess position can have at most 218 playable moves.
The article does not make super obvious that it is talking about finding a valid chess board configuration with the maximum number of valid moves for one of the sides, other than that, great! Tbh I would have approached it differently, trying to prove that the existing known lower bound was not improvable, but I'm sure I would have failed at that.