Every engagement plotted by what it actually is — deal judgment to the left, frontier research to the right; advised below, built and shipped above. Recolour the field below by how the work was done, its sector, or who it's for — then click any point to open it.
Deep in one lab is common. Range across the whole field — and the judgment to pick the right model per problem — is rarer.
Experience with illustrative models below:
Production safety & evaluation work at the frontier.
Deployed in production, at scale, under adversarial conditions.
Replace expensive frontier calls for repeated, specific tasks.
World-model-safety research — “the geometry of incomplete world models.”
Building proprietary ML since the early 2010s — including the black-box oversight models I presented to the FCA: mechanistic interpretability before it had the name.
— among others; the point is the range, and knowing which to reach for.