I make them visible and playable — one small, provable demo at a time.
I turn frontier AI ideas into things you can actually see and poke — backed by reproducible evidence, instead of hand-waving or unreadable math. First series: what's really happening inside the model.
A 155,000-parameter network is trained to do one thing: predict the next move-symbol. It is never shown a grid, a map, or a coordinate — yet to do its job it builds a model of the world and hides it in its activations. A linear probe reads that world out, live. Then you reach in and lie to it — and it acts on the false belief, even hallucinating things that aren't there.
The model runs locally in your browser — no model inputs are sent to a backend. Every claim backed by a measured test.
Open the live demo →Two neural agents start with no shared language and invent one, live, to win a game — a real vocabulary crystallizing out of noise. Genuine emergent communication, trained in your browser.
A reversible cipher: type anything and it becomes an alien glyph-field — illegible at a glance, yet decoded back perfectly by its rules. A playful study in encoding (no AI here — that's the other two).
If a toy network can build and use an internal world model, it becomes easier to see why larger AI systems may carry internal representations we need to understand — at a scale we can't fully see. That's the whole game in interpretability and AI governance: in high-stakes domains like finance, "the model said so" isn't enough — we need to know what it represents and why.
These demos are my way of arguing that case in public — not with opinions, but with things you can run. Read the full write-up →
Interactive deep-dives into what AI models actually know — every issue is something you can open and play with, not just read. Next up: rendering the real-world map of place and time hidden inside a large language model.
I've emailed you a link to confirm your subscription to Inside the Model. Click it and you're in. (Peek in spam or Promotions if it's not there in a minute.)
I'm a principal engineer and AI architect. For close to two decades I've designed and shipped systems that have to work under real load — and for the last stretch of that, AI systems that have to be trusted, not just demoed. Most explanations of what happens inside a model are either hand-waving or unreadable math. I build the third thing: arguments you can see, poke, and reproduce. This site is where I do that in the open — one provable demo at a time.