Announcing Astar Labs

When AI started making headlines, I did something counter-intuitive: I tuned out of it. Not because I wasn't interested; quite the opposite in fact. I was too interested in the wrong direction. Whilst everyone else was asking what AI could do for them, I kept getting stuck on a more fundamental question: how does it actually work?

The only honest answer to that question, is to build one yourself.

Astar Labs started on March 12th, 2026, as a personal project with a simple goal: build a language model from the ground up, understand every decision that goes into it, and share it with whoever's willing to listen. No team, no funding, no roadmap handed down from above. Just a rented GPU, a lot of research papers, and funnily enough, some much-needed AI assistance.

The name is borrowed from a previous life: Astar Technologies. That was a project where I built self-landing model rockets. Different domain, same impulse: understand the thing by building it yourself.

The goal is modest by industry standard. We're not chasing state-of-the-art benchmarks or competing with well-funded labs. The focus is on small language models: efficient, lean, and genuinely understood by the person who built them. If they turn out to be useful to a handful of people along the way, that's a bonus.

This is only the beginning. There's much more still to come.