Some months back I started working on Blaze, a portable DSL-based kernel compiler. The goal was to prove a point: a well-designed ML compiler can deliver competitive performance across different hardware. After hundreds of hours of development across multiple models and backends, I'm ready to share some details.
To start, I'm releasing a small binary benchmark suite for Blaze's GPU backend, built on NVIDIA's PTX , a layer below CUDA, compiled down to native code for multiple generations of NVIDIA GPUs. The benchmarks show competitive performance against llama.cpp - especially LLM prefill - on both newer and older NVIDIA hardware, and demonstrate scaling across two different model families (LLMs and diffusion).
They also point to specific gaps, especially on larger models, and those gaps set the roadmap for what gets optimized next.
This is a research project for now. The compiler also targets x86, Apple CPUs (NEON and SME), AMD and Metal GPUs, but those need more device testing before I publish their numbers.
Repo (binaries, Docker environment, measurement protocol): https://t.co/TJfaZoTbBd
When I saw the @antirez excellent @MiniMax_AI h3 video generation implementation on Metal, I thought -" what better way to toruture my tiny M4 ANE chip than running a 33B parameter model on it!" 😅
fork and torture your ANE chip too! https://t.co/ubk8pujgs0
Built with SSD streaming and overlaps int8 compute on ANE for the diffusion transformer. Only occupies 2 GB of peak RAM and under 5 W of power.
PS - If you got a spare M5 Max around, please send me one for ..."experiments!" 😇
My dog Rosie was given months to live. Chemo and immunotherapy failed
I used chatgpt + computational genomics to design a personalised mRNA cancer vaccine for her tumour. Several of her tumours shrank
Today we launch Gamgee (YC S26): personalised mRNA cancer vaccines for dogs
"Yes Claude, let's pause and move on from the core goal we have been working on since last two days.
Yes, yes rest is more important. Thank you for caring about my well-being 🥰"
Imagine torturing a model so much, that's its first instinct is to defend itself against a hallucinated constraint it wrote itself.
And then lecture the users why the model deserves your kindness
Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.
During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.
That’s why we created the Open Secure AI Alliance.
I'm having Opus 5 3d build my motorcycle (Royal Enfield Interceptor 650) from scratch using Blender and 3djs , from publically available part sheets, photographs and diagrams.
It proposed starting from the engine, let's see how far it can go 🤓
I mean would you be impressed with this lol..it's not all bad but I guess the blind designer can only go so far.
Safe to say automotive designers and engineers aren't going to be replaced anytime soon 😅