Subjective World Model

Can a model represent people as they actually differ, in taste, judgment and story, rather than as an average?

Most AI models are trained to converge on the most probable answer, and in doing so they erase the differences between people. They also learn from public text, which records what people say, not what they do. A subjective world model takes the opposite aim: its state is a single person.

Each person is built from four layers of evidence: expression (what they say in reviews, posts and surveys), story (how they narrate their own lives), cognition (the weights they actually give to price, taste, health or novelty, inferred from their choices rather than asked) and behavior (what they did). When the layers disagree, as when someone who says they avoid sugar keeps ordering sweet tea, the contradiction is kept as evidence rather than averaged away. Given a new event, the model rolls the person forward many times and returns a distribution of reactions, not a single guess.

The model is checked against people. Its predictions are compared with in-depth interviews of the same people, and AI personas play the same behavioral-economics games as humans, following Park et al. (2024). When a design reaches the market, the gap between predicted and actual response is used to recalibrate the personas.

atypica.AI puts this research into use. It builds AI personas and lets them take part in interviews, user tests and discussions, so that designers, researchers and businesses can simulate how different people would respond to a product, a message or a policy before it exists. Its premise is a design premise: people do not choose between things; they choose between descriptions of things.

atypica.ai ↗