Every day, people interact with software. For a long time, we taught software to remember: filenames, passwords, email addresses, bookmarks, desktop shortcuts. Software did not learn. It stored exactly what we gave it, then waited for us to come back and ask for it.
Then recommender systems appeared, and software started to learn. It watched what we opened, skipped, lingered on, saved, and returned to. From behavior, it inferred preference. From a fixed library of content, it learned to choose the item most likely to fit.
That kind of learning rests on a premise: the content does not move. The same video, article, song, or product is the same item for everyone. Learning means aligning different users through a shared set of stable items. If you and I watched the same thing and reacted in the same way, the system treats our tastes as similar. In effect, the item defines the user.
That paradigm solved a generation of product problems. But when content is generated by a model in real time, the premise breaks.
Every conversation is a unique trajectory. The same user can ask the same question twice and receive two different replies. Two users cannot really have the same conversation, because each exchange is shaped by what happened just before it. Without an invariant item, users can no longer be aligned by their relationship to one.
The old way of defining a user through items stops working. We need a new way for software to know people: not by what content they happen to share with others, but by what actually happens between this person and this system.
We tell these systems our tone, our way of thinking, and the rhythm we want. We explain our background, correct their wording, and point them somewhere different when the answer is almost right. We expect them to remember what we said yesterday, the taste we showed, and the worlds we already built together.
We teach them a lot. They learn very little.
Not because the models are not smart enough. Most systems were simply never designed to learn in this way. They were designed to give everyone a pretty good answer, not to live with one person over time and slowly become different because of them.
Generative systems do not have to stay like this. The worlds they create can carry a person’s imprint forward. They can become more specific, more continuous, and more shaped by the people who spend time inside them. They can let the world they create grow because of you.
Why this matters
We believe the next generation of content will not arrive as an item selected from a catalog. It will be shaped in the space between a person and a system.
Entertainment becomes less like a feed and more like a world with history. Companionship gains continuity. Education can notice what finally made something click. Creative tools can preserve the trace of a person’s intent instead of resetting to a blank surface.
This is not only a scaling story. It is a choice about what kind of relationship software is allowed to have with people: whether each session disappears when the window closes, or whether time spent with a system can leave a mark on what comes next.
That is the work of this lab. It is tested every day inside a product used by real people. We break that work into concrete directions on our research page.
If this is the problem you cannot stop thinking about, come work with us.