Strange Lab

Strange Lab

We build Predictive Intelligence for every world.

Every org is already a world. It lives in a timeline, and generates a continuous record of what it does: mail, chat, docs, tickets, code, etc. We create that record and turn it into a "state": company snapshot, timeline of decisions, documentation, policy, all semantically enriched and assessed against outcomes.

With this, we build a world model for each domain we're interested in. It understands causality, enough to help learn the nuances of the world and understand what makes it tick. This provides your Codex or Claude Code with a better 'causal core' to embed within decision making. And with it, we get Predictive Intelligence.

This is useful because it means we can check counterfactuals, generate predictions, understand how various parts of the world fit together and what the outcomes of any action might be, far better than just using an LLM (we tested it).

For instance, within companies this means it helps us immediately figure out which parts of that event stream are actually useful, to figure out what workflows AI agents should target, what skills they should automate first, and how we should organise the company, humans and agents alike, to get to the future! This already works by the way, highlighting key workflows and multiple discovered skills for our design partner, updated daily!

We've tested this now on a large variety of "worlds". We've tested it with two live startups, Enron, Star Wars and Tolkien's worlds, Macroeconomic datasets, the U.S. Civil War, and a flagship effort with Project Confirm, the oncology initiative from FDA, where official accelerated-approval outcomes give us a public answer key. In each one, we see that we're able to understand the world better and get better predictive accuracy for key points or decisions.

Better predicting the world is the first step towards understanding it. Our goal is to make sure that the causal understanding of our world is made explicit. We need it if we're to get towards truly general intelligence!

Two public examples are the best place to start.

  • Bismarck: a dense historical record becomes dated events. Later history is hidden, GPT and the world model see the same past, and the forecast is scored against what happened next. The page also lets you try alternate Bismarck timelines from the same cutoff.
  • Project Confirm Oncology: FDA's accelerated-approval oncology record becomes an answer key for which drug programs verify, stall, or get withdrawn. Whole drug assets are held out; later FDA actions grade the forecast.

Essays, papers, and code

Adjacent work