A map of what a 27B understood

Every dot is a photograph, placed by how a Qwen3.8-27B read it. That model ran once, months ago, and is not here: what survives is about a kilobyte each. A frozen Qwen3-0.6B that has never seen a photograph says what is at any point you choose, and every region name on the map was written by it, not by us.

Click anywhere on the map, or type a sentence and watch where it lands.

How the map is made. The room has 1024 dimensions; this is a UMAP projection of 40,000 of the photographs down to two, so what you see is neighbourhood structure, not coordinates that mean anything on their own. Clicking reads the average of the ~48 photographs nearest that spot. A sentence is placed by where its nearest neighbours already sit.

The honest part. The reader is fitted to the region where photographs live, so a point far outside it drifts toward a generic scene rather than failing loudly. On 5,000 held-out photographs a sentence it writes retrieves the photograph it came from at median rank 18 of 123,287, and the same reader handed a different photograph's point scores at chance. Checkpoints and full evaluation · repository