Doing GPU programming and building Three.js projects in public. I post demos, tutorials, and deep dives.
Joined July 2025
- Tweets291
- Following35
- Followers42
- Likes518
Pinned Tweet
Everyone's doing something with the fruit fly, so I figured I'd play around with it too.
I made him a slippery race track so he could drift 🪰🏎️
I'm using FlyGym for the fly model and the "animations".
For the environment and physics I had to make my own to make it easier to allow slipping and drifting. I tried to make it work with MuJoCo at first but I figured I'm not looking for realism, I just want slipping and drifting so why even bother.
A camera on his head renders what's in front of him as a depth map and dumps that into his visual projection neurons. And information about his speed and direction gets fed into his neurons related to his Johnston and campaniform cells. Then I read his DNs to drive his legs and wing flaps.
At first I was trying to train him by updating his KC->MBON synapses to have him optimize for speed and do cooler stuff but visual input barely made a difference in how the KCs were firing so I gave up on that. I'm not a biologist so skill issue I guess lol. I'm still very happy with the result tho!
Small update on my game.
I think the horse movement and animations are good enough for me now!
I was really struggling with animation last week but I'm definitely getting the hang of it.
Based on how badly AI did the gallop for me initially, I gave up on it and made that one by hand but I figured I'd give it another go because I really don't like animation😅
So for my next animations to improve my prompt, I gave it my handmade gallop as a reference on how far the legs are allowed to bend, how the knee should react, etc and with that it was able to almost one shot walking and trotting.
I've been seeing videos of people using video generation models to generate videos of the animations they want as a reference and apparently that works really good as well.
In my own experience trying to tune the animations with AI one thing I've noticed is, it can do exactly what I ask it to do but I have to be extremely specific. Leave room open for interpretation and 90% of the time it'll interpret it a different way😂
There's a lot of implicit knowledge we don't put into words when we talk about motion but AI doesn't have that so it can make silly mistakes I guess.
To solve that I made a small rule I use where I have it explain in detail the new poses and motions I ask for so I can catch any misinterpretations early.
I decide to make a small game. I can't make 3D models or animate them but I figured AI could help.
- modeling / texture: Tripo gets it done
- rigging / skinning: I don't have an AI tool for this but I got the ARP plugin for blender and it was pretty easy to work with so no complaints
- animation: I figured maybe Astra could help but no... I mean it worked and it did quite well, I just didn't like some parts of the animations it made. I tried to get it to fix them but it didn't work so I figured maybe I should just learn how to make animations myself.
Anyways here's my first serious attempt at animating a horse🫡 (I need to fix some weights and make his neck move but overall I'm proud of it)
Why is everyone wasting resources on this? Just give me (2^2n)n bits of RAM and a couple thousand years to prepare and I'll give you multiplication in constant time 🗿
We are publishing an update to OpenAI problem #109 (integer multiplication) with further tightening.
κ = 2⁻³⁴ (tightened from κ = 2⁻¹⁸²)
The exact witness is 8.3 × 10⁻¹¹, a roughly 48 million fold improvement over the previous result and a 2¹⁴⁸ fold improvement over original OAI result.
The improvement came from removing the spacing penalty behind the quadratic bottleneck. This was done by moving compact control bits instead of entire windows.
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
Just set up the base for a small shader experiment I've been wanting to do for a while 🙂
I also used the opportunity to turn the codebase into a reusable template so in the future I can start projects much faster.
Will post updates soon!!
Finally got back to the gym after a month due to my shoulder. Man it's so demoralizing not being able to lift the same. Gotta build back up 🫡
Found an absolute gem for understanding reductions in GPU kernels 👇
developer.download.nvidia.co…
I love the way it walks through multiple optimizations added on top of a basic solution. That's the best way to build an intuition for stuff!!
#cuda #gpuprogramming
Arayz retweeted
Introducing Batmobile.
10-20x faster CUDA kernels for equivarant GNNs. The bottleneck in MACE, NequIP, Allegro? Spherical harmonics and tensor products. e3nn computes them with Python overhead and global memory round-trips.
Batmobile keeps everything in registers.
> Spherical Harmonics: 11.8x
> Tensor Product: 20.8x
> TP Backward: 20.6x
All CG coefficients baked in at compile time. No loops, no branches, no intermediate tensors.
Cool trick!
A quick breakdown of my interactive waterfall
If you’re interested in creating your own shaders, I’m writing an e-book about it (almost 300 pages written) You can wishlist The Shader Survival Guide for free here
gameslave.dev/theshadersurvi…
#shader #shadergraph #hlsl #madewithunity
Low poly procedural terrain generator I made a while ago inspired by the "stacked pyramid" approach mentioned here.
It's looks so cool!
Trying out one new silly idea for procedural noise
#indiegames #indiedev #gamedev #screenshotsaturday
Cooking up my first project I'm planning to build in public! 🚀
Man @Railway has some crazy good DX. I love how simple it was to set up everything 😁
#BuildInPublic
New post: LeetGPU Challenge #7: Reverse Array! 🚀
We're gonna go over implementing an in place array reversal GPU kernel with Triton.
#LeetGPU #GPUProgramming #Triton #CUDA
arayz.dev/blog/leetgpu-rever…
This is gold. Gotta keep it in mind when you eventually launch that SaaS you've been thinking about vibe coding 🤖
#BuildInPublic #SaaS
xitter.cf/RKRigney/status/201053…
