arpu retweeted
There's a lot of AI hate on @pouetdotnet lately, so I decided to vibecode a demo about it.
💾 xplsv.github.io/lolfuckai/
arpu retweeted
We just open-sourced the world's fastest WebGPU kernels for 200+ machine learning operations.
Specialized for your hardware. They run entirely locally in your browser. Now available on Hugging Face 🤗
Tell your agent to use @huggingface/kernels in your next project.
arpu retweeted
I’m having AI port Three.js and its examples to Rust + WebAssembly for various experiments.
github.com/takahirox/three-r…
arpu retweeted
🔥🏆Heat Kernel Textures (HKTex) received the Best Paper Award at #ECCV2026. 🏆🔥
UV-mapping anchors 3D texturing to the past. Inspired by 3DGS, we rethink textures using heat kernels and NO splatting!
circle-group.github.io/resea…
arpu retweeted
After more testing this is what I'll use for Basis Universal's PBR material support: a PCA-based or bilinear model. Way faster to decode vs. MLP (neural) and not need a GPU or even SIMD.
We can already encode to full BC7/ASTC in real-time (~8 threads=1 gigatexels/sec.).
If the big win from neural texturing is saving VRAM, there are lots of ways to do that by creating your own non-neural custom texture formats, and doing mipmapping/filtering yourself. Then: no neural-specific IP concerns. Decoding is also fast/simple.
I'm using the same ES training code I wrote for neural nets but with a bilinear model. The decode cost: 60-250 weights.
arpu retweeted
You don't need neural networks (at all) to compress PBR material textures. PCA applied to texture/image data is old prior art. The lower dimensional data is easy to handle.
"Dimensionality reduction for image and texture set compression" by Bart Wronski
bartwronski.com/2020/05/21/d…
arpu retweeted
Equal Earth projection is coming to 3DTilesRendererJS! 🗺️
Here it is rendering TerrainRGB & vector annotations. The LA streets look a bit less square but the projection looks great!
We are updating the engine so data aligns across projections, letting you render maps in three.js however you need ✨
#equalearth #GIS
arpu retweeted
Fitting a neural texture decoder with ES (no backprop):
richg42.blogspot.com/2026/09…
arpu retweeted
Furry objects render beautifully in NeRFs but crumble in real-time graphics. This method replaces volumetric blobs with explicit, anti-aliased line segments for hair and fur that rasterize, shade, and simulate properly.
📄 arxiv.org/abs/2609.00625v1
🌐 kenji-tojo.github.io/sa26-li…
arpu retweeted
Grok: "Encoding *through the filter* lets the codec spend bits on interior detail instead of fighting seams."
Note none of this work is theoretical. We have billions of ASTC decoders in hardware right now, and we have the software to unlock these large block texture units.
github.com/BinomialLLC/basis…
arpu retweeted
It's on the list. Enable it in filters :)
A ready to use link: bughunt.productcompass.pm/?r…
arpu retweeted
if you ever have to read, debug, or hand-write SDP, try explainer.pion.ly/
It's an interactive SDP explainer and debugger built by @supadotsh Great for debugging WebRTC, SIP, RTSP, or just finally learning what every SDP line actually does.
arpu retweeted
A GPT-3 moment for robotics⚡️
The first time it hit me was watching a Skild robot make pancakes, after seeing it done just once. It had never been trained on pancakes before.
One Skild Brain:
- doing tasks far OUTSIDE its training distribution from a single video prompt, with no task-specific fine-tuning
- executing 10+ min long-horizon tasks
- capable of instruction following
They have cracked in-context learning💥 The era of general-purpose robot intelligence is just beginning
arpu retweeted
Three.js + KHR_interactivity demo github.com/takahirox/threejs… #threejs
arpu retweeted
Playing around with pushing Three.js WebGPU further.
Three.js stays in JS running through Node, while the rendering is passed down to C++ using Vulkan + RTX.
Now testing native ray-traced reflections, hit radiance and lighting.
Still experimental, but pretty happy with how far this has come so far.
Has anyone built a 3D Gaussian Splatting trainer that runs directly in the browser using WebGPU or Three.js TSL?
I’m wondering how far we can push on-device training/optimization 🚀
Would love to see any experiments, repos, or demos!
arpu retweeted
Depth-aware light injection in TypeGPU
I got a 448x448 monocular depth model down to ~8 ms on my M4 Pro across ~250 dispatches, which is fast enough to use in realtime :D
Since the inference is written directly in TypeGPU, I can just feed the depth buffer straight into the lighting pass. It never has to leave the GPU or go through any extra synchronization/interop step
Inference, lighting and draw all go through the same command encoder.
arpu retweeted
It's official: Qwen3.8-27B just scored 52 on the @ArtificialAnlys Intelligence Index.
We now have an open-weight model that matches GPT-5.6 Luna (max) AND can run locally... even in your browser with custom WebGPU kernels!
What a time to be alive! 🤯