TasteGraph

Taste is not a vector.

Building AI systems that model human taste.

Structured. Contextual. Evolving.

Explore our research

A wide relief model of terrain built from many stacked layers of cut grey board, forming nested terraces. A single taut plum thread is stretched low across it, casting one thin shadow. The model recedes into darkness at the back.

Research

Explore our research notes.

View all research
Read note 01, Taste Is Not a Vector
A relief model cut from many stacked layers of grey board, its terraces nesting into a broad summit. A single small steel sphere rests on the highest terrace, casting a small precise shadow. A plum thread lies along a lower terrace, far from the sphere.

Note 01 · Representation

Taste Is Not a Vector

Vectors are exceptional tools for computation. The difficulty begins when a representation built for one task is treated as a sufficient model of the person behind it.

Read note 02, Preference Has a History
A row of slender steel pins standing at irregular heights on a grey board. A single continuous plum thread is wound between them, looping high above the row near the middle before returning to the line. Each pin casts a long hard shadow.

Note 02 · Temporal Models

Preference Has a History

An observation is not only a value. It carries a date, a context and a relation to everything observed before it. Every preference has provenance.

Read note 03, Context Is Part of the Preference
A smooth curved plaster surface, pale grey and matte, with one shallow depression worn into it. A matte grey sphere rests in the depression, held by the curvature. A plum thread is laid across the plaster along the line of steepest slope.

Note 03 · Context

Context Is Part of the Preference

Preference is inferred from evidence produced in situations. Removing context does not remove noise. It changes what the observation says about the person.

Preference leaves traces. Traces are not understanding.

Every search, choice, rejection, purchase and discovery carries some information about preference. None of it explains itself. Understanding taste means reasoning about what those signals mean, how they relate, and how their meaning changes over time.

Taste has structure.

Patterns persist across choices that look unrelated, and they hold even when a single choice looks out of character. What a pattern does not carry is the situation it applies in.

Taste is contextual.

What feels right depends on the situation in which a choice is made: the occasion, the constraint, the person it is for. An object can be entirely to someone's taste and still be wrong for the moment it appears in. Context is not noise around a preference. It changes what a choice tells you about the person. So does when the choice was made.

Taste evolves.

People accumulate experiences, and taste develops with them. Some tendencies hold for decades. Others belong to one period of a life, and the evidence rarely says which is which. A model that reads only the present cannot tell them apart, and neither can one that keeps everything at equal weight.

Each of the three has been modelled well on its own. Our view is that a useful model of taste has to represent all three at once. A model that represents one and not the others will be right often enough to be misleading.

A representation layer for human taste.

We develop representations that help AI systems reason about a person beyond individual clicks, prompts or stated preferences. The goal is not simply to remember what someone liked. It is to model enough structure to understand what may matter next.