INTERP ATLAS

About

An orientation device

The Interpretability Atlas maps the visual and computational tools people use to inspect, explain and steer the inner workings of AI models. By showing what forms these tools take and what they make visible, the Atlas helps people understand the field, identify gaps, build on existing work and find meaningful places to contribute.

The visual is the entry point, not the illustration. Not "here is the field, and here are some pictures of it," but: someone can lead with visual thinking, navigate by what things look like and what they let you see, and get oriented that way rather than by reading forty papers first.

The position

"Grow visual ways of building with interpretability" is not neutral cataloguing. It is a position: that interpretability should be more visual than it currently is, and that laying out what exists makes the absences obvious enough to act on. An index with a thesis is more useful than one without, as long as the thesis is stated rather than smuggled in.

Five pillars

  1. Orient

    see the shape of the field, what exists, how it is distributed, where the work has clustered.

  2. Understand

    grasp what a method does by looking at what it produces, not only by reading about it.

  3. Apply

    find something usable and take it into their own work.

  4. Contribute

    see the field clearly enough to add to it.

  5. See the gaps

    spot where visual approaches are thin, or done only once, or missing entirely.