Writing
Research and notes from Jean Technologies.
How Much Catalog Fits in a Semantic ID
A semantic ID with three levels of 256 codes has 16.8 million addresses. On a 66,000 item catalog it collides on one item in nine. The gap between those two numbers is the whole design problem, and it is measurable from the collision rate alone.
Two live bets about how a recommender should name its items: keep the IDs sparse and scale the model, as HSTU does at sixty terabytes of embedding table, or make the IDs carry meaning so the model can be ordinary, as TIGER and PLUM do at under eight megabytes. The compression has to come from somewhere. Deriving the effective capacity of a codebook from its collision rate shows where, and gives a rule for sizing one before you train anything.
A Field Under Many Names
Five literatures, from ads ranking to national health registries to psychology experiments, are converging on one artifact: a model pretrained on what people do, inducing a representation of a person that transfers across tasks. What makes it a category rather than a coincidence is the substrate underneath it and the tasks it opens up, including simulation and experimentation, which had no commercial form until something faithful enough to simulate against existed. We collected the reading list and are writing a survey.
The Internet of Agents Needs a Model of You
We competed at the AGI House Internet of Agents Build Day on the simulated and RL track. The premise (100 agents per person) sharpens the thesis: an agent acting in your name is only as good as its model of you, and delegation adds a trust gap that never existed when humans acted for themselves.
Emotion Vectors and the Future of Human Models
Anthropic just published landmark research showing that LLMs maintain abstract, causally operative representations of emotion. This is something we have been thinking about for a long time, and it changes how we should model people.
Why General-Purpose Embeddings Fail at Modeling People
Standard embedding models are trained on text similarity. But when the goal is a working model of how a person behaves, text similarity is the wrong objective entirely.
Local Drift-Adapters
A single global adapter fakes an embedding-model upgrade for a fraction of the cost, but its uniform-drift assumption breaks on real corpora. Local drift-aware adapters close most of the gap.
Introducing the Embedding Adapter
Switching embedding providers has historically meant re-encoding your entire vector corpus. Today we are releasing infrastructure that eliminates that requirement.
Jean Technologies
We started Jean Technologies to solve a problem we kept seeing across industries: products whose whole value depends on understanding a person, built on infrastructure designed to understand documents.
The State of AI Memory
The first thing we built was not a model of behavior, it was memory. A memory system files the million snippets that make up a person; the model is the layer above it, the meta of memory, that turns those snippets into one representation of past, present, and intent. A first-principles look at the Experience, Storage, Recall pipeline, the Memory Frontier, and why the model layer sits on a memory layer most teams never look at.
General Personal Embeddings
An exploration of how deep representation learning can be applied to people, and why the logical endpoint is trusted infrastructure for personal embeddings.