Mission
Build computers that understand humans.
History
In 1997, IBM's Deep Blue beat Kasparov at chess. Excellence in chess had long been the mark of intelligence, and a computer beat our best.
In 2026 these machines have a superhuman understanding of code and mathematics, which leaves the question of what falls next. It is increasingly likely that the answer is us.
How we got here
Jean started by building downstream products: AI memory, context engineering, matching, embeddings, recommendation. Each one pushed us toward the same underlying problem. None of them can be solved well without a model of the person, and every team building one was quietly rebuilding that model from scratch, badly, from whatever data it happened to hold.
So we stopped building the applications and started building the thing underneath them. Instead of a model that predicts the next word, a model that predicts the next thing a person does.
The applications did not go away. They became readouts. Matching is how we measure whether the representation is any good, because the ground truth is unambiguous: the interaction either worked or it did not.
What counts as one
“Foundation model of human behavior” is worth defining precisely enough to be checkable, otherwise it is a slogan. Four clauses, all of which have to hold.
- 01
Trained on behavioral traces
Records of what a person did, not descriptions of who they say they are.
- 02
Under self-supervised pressure
Predicting the next thing a person does needs no labels, so it scales to all recorded behavior rather than the sliver someone annotated.
- 03
Inducing a per-person representation
The artifact is a representation of an individual, not a score for one task.
- 04
Consumed by many tasks
If only one task can read from it, it is a good model of that task, not a foundation model.
The hard part is the data
There are three versions of a person available to a model. The stated self is the profile, the survey, the preferences typed into a settings page. It is cheap to collect and it is the weakest, because it is a document someone authored about themselves.
The revealed self is the trace: what was actually clicked, bought, opened, abandoned, returned to. That is what we train on.
If we are right, every meaningful interaction between a person and software runs on a model of that person. We would like it to be one they consented to, can inspect, and can take with them.