Essays on one question: can we actually trust AI where being wrong is expensive? Two strands — reading the mind of a single model, and governing systems of many.
Trusting and governing AI at the scale of many agents working together.
Multi-agent systems are distributed systems
The hard problems in multi-agent AI aren't new — they're the ones distributed systems spent twenty years learning to survive. The governance layer being built on top mostly isn't drawing on that heritage, and finance is where it bites first.
Interpretability at the scale of a single model, made playable.
What does a machine actually know?
A network that only ever read move-symbols built a map of its world inside — and acted on a false one when I lied to it. What that means for trusting AI, with a live demo you can poke.