Foundation models of human behavior
A spec preview, and partly how we are working out what the product should be. If a page is wrong about your problem, tell us.
Language models are foundation models of text. We are building the equivalent for behavior: models pretrained on what people actually did, which transfer to a new product or catalog without a year of collecting your own labels first.
Four products, in the order they depend on each other.
1. Data
Sourcing and structuring the behavioral data a model of human behavior needs. Scoped with the team.
2. Foundation models
Models of human behavior. First release is a foundation model for recommendation. Closed beta.
3. Semantic IDs
Turn a catalog into discrete semantic tokens a generative recommender can generate. Spec preview.
4. Simulation
Rehearse a decision against a simulated population before it meets the real one. Research direction.
The shape of the argument
-
Behavior is the missing corpus. The public corpus is full of what people said and thin on what people did. The sequential, consequential, content-linked data that a behavioral model needs mostly sits unused inside companies. That is product one.
-
Pretrain on it. Train broadly enough on real interaction sequences and you get transfer, the same way you do for language. A downstream team starts strong instead of starting from zero. That is product two.
-
Items need a vocabulary. A sequence model over items only transfers if items have meaningful identifiers. Random integer IDs do not transfer and cannot cold start. Semantic IDs can. That is product three, and it is the most concrete surface here.
-
Then run it forward. A model that predicts what a person does next can be run against a population to rehearse a decision. That is product four, and it is the least built.
Where to start reading
If you want the argument, Semantic IDs is the most developed product: codebooks, collaborative signal, and a full API sketch.
If you want to call something, jean-rec-1 is a pretrained recommender you can point at a catalog without building a model first, and Models covers what we serve and which endpoint answers which question. The API reference has every endpoint with its request and response shape.
Get access
Nothing here is self-serve yet. Every engagement starts with a conversation about your data and what you want the model to predict.