Strange Lab

We're building world models for work.

Frontier LLMs can reason and use tools once work is specified. Open-ended work starts earlier: choose what matters from thousands of possible actions.

A world model learns how a particular repository, drug program or company changes from its dated record. It ranks possible work and forecasts what happens next. The LLM reads the evidence and uses tools.

We have tested this in software, life sciences, cybersecurity, games and fiction. In software, one model transferred to a ninth repository without retraining. In life sciences, another predicted FDA gates for whole drug programs excluded from training.

We start with one bounded world, test transfer across related worlds, then build domain and industry models. Our world models essay explains why the same approach may help agents work in open-ended environments.

Our goal: simulate every decision, across the economy.