◢ ◤ // Freelance @hyperlane
// building @livebonk & @bonk_fun
// Ex- @improbableio
// F1 ⚡ Crypto ⚡ AI
// Head of Wristbands @ WestHamCapital
// ETH is money
Jenga v1.0.1 drops in 14 days
• Functions support
• Bitwise ops
• JS-like type system
Contributors: check the new files in src to get familiar with the codebase.
Dev environment: Arch Linux, i3, Vim. Nothing fancy.
Gemini Nano Banana Pro.
Prompt: First-person perspective inside a brightly lit supermarket aisle. Realistic human hands are holding a bottle of Fanta soda close to the camera. The vivid orange drink in its iconic branded bottle is surrounded by a multi-layered holographic augmented reality interface displaying nutritional data, including calorie count, sugar content, caffeine level, freshness indicator, expiration date, and recommended refreshing recipes and cocktails based on Fanta.
The UI elements smoothly shift and reorganize based on the viewer’s gaze direction, as if dynamically responding to user focus. In the left peripheral vision, a vertical semi-transparent shopping list is visible with checked-off items, where Fanta is highlighted as the currently active selection.
Hyper-realistic mixed reality, clean futuristic AR design, glass-like UI panels, soft ambient glow, realistic lighting and shadows, natural depth of field, immersive first-person interface, showcasing next-generation retail technology.
Zartbot is one of the most interesting AI researchers at the moment. Unsure why Alibaba gives him so much freedom to spread his ideas.
The breadth of his knowledge on hardware, mathematics and software is amazing.
A small selection of his posts are translated and on Github here:
https://t.co/ACChGNCDqN
But articles like this one on the relationship between Optimal Transport and Attention are only available in Chinese: https://t.co/V17upWyMwn
Devlog: C++ & inference
1️⃣ Spent time on broadcasting + topological sort for the autograd engine
2️⃣ Reading the Hugging Face ultra-scale playbook
3️⃣ Digging into the llama.cpp codebase
LFM2.5-Audio by Liquid AI coming soon to MLX-Audio.
Currently seeing ~6s to generate 5s of audio at full precision — no custom streaming, no quants.
With streaming enabled, performance should jump significantly.
React Compiler feels magical—code that used to break apps now just works.
Still, I’m not a fan of the React rendering model. Signals with a stable component tree are way easier to reason about than constant re-renders in React.
New today: we went a little crazy rewriting our @ logic in Rust — and it’s now 10× faster on large codebases 🚀
One of those changes you really feel once you hit scale.