The American mind cannot understand this

So far my first attempts to train a keyword spotter model from scratch have failed. It’s like learning how to bike all over again. You can vibe code a huge SaaS app in a few hours but one can’t vibe train.

The untold joy of seeing loss decrease with a massive training dataset

Interesting to see that one can train a 1 014 parameters model for keyword spotting and making it work. I might have to write an article for this one

Infinite Wiki ⁕ Every word is a hyperlink. Every description is generated in real-time (in ~1 second) ⁕ Runs on Gemini 2.5 Flash Lite. ASCII diagrams using 2.5 Flash

LFM2 🤝 Outlines Constrained generation is amazing to squeeze more performance from tiny models. We just released a blog post with Rémi 📎 with a smart home control demo. We used LFM2-350M to save notes, close blinds, and handle displays.

Is there such a thing as a K8S / N8N / Redis / React-Frontend (replit-based) / Python-backend / Tilt-enabled boilerplate starter ?

Oops I did it again... Guessing what's next ?

whisper.cpp is coming to ffmpeg github.com/FFmpeg/FFmpeg/…

Shipped two new open source projects today. The coolest one isn't quiiite ready yet but already on my GitHub if you're nosy parker. 🍭

The Nano Banana 2 model is the first one to understand a complex architecture and actually draw it.

Come on ! Beautiful docs are one click away... I mean READMEs won't ever look the same now... Logan Kilpatrick Nano Banana 2

> llama-cli -hf org/model

Grok @levelsio Pat Walls Grok now argue against everything you just said. Find the weakest points in your logic.

Distilling software engineering expertise into a AI coding skill has surprising effects. It’s a freaking dopamine shot. Remember the doherty threshold story in Halt and catch fire ? The thing that get you hooked? This is it. For context: think Replit but for backend tasks