What we found

The numbers below come from this run. The explanations are hypotheses we haven't tested yet. Select a name to see that person's neighbours under every model.

Which models agree?

Each cell is how many of a person's 3 closest neighbours two models have in common, averaged over all 68 people. 3 means identical neighbour lists for everyone; 0 means no overlap at all. The colours span the lowest to the highest value in the grid, not 0 to 3, so small differences show.

What do the closest neighbours have in common?

The share of each person's closest neighbours who work in the same field, or are the same gender. The grey line is what you'd get by picking neighbours at random. The further a bar is above it, the more that model groups people by that trait.

Who changes most between models?

People ranked by how much the models agree on their closest neighbours, least agreement first. Select a person to compare their neighbours under every model.

Explore one model's map

t-SNE projection of one model's embeddings. Grey lines link each person to their 3 closest neighbours. Hover over a point for details; click it to compare that person above. Each map is laid out independently, so compare who is near whom, not where.

Perplexity sets roughly how many neighbours each person's position takes into account. Low values keep small, tight groups; higher values also arrange whole groups relative to each other. Compare Marie Curie at 10 and 30.

Distance on the map is an approximation. Squeezing hundreds or thousands of dimensions into two can’t keep every distance, and t-SNE gives up long distances first. The neighbour lists and similarity scores above are computed on the full vectors. When the map and a list disagree about a pair, trust the list.

Limits

• The field and gender tags were assigned by hand, one field per person, and some are judgement calls (Leonardo is tagged as art, Franklin as science). The field scores are only as good as those choices.
• Results drift between runs. The embeddings API returns slightly different vectors on each call, so neighbour lists and maps can shift. We haven't measured how large this noise is, so treat small differences, such as 0.1 in the agreement grid, with caution.
• The sample is small. For one person and one pair of models, agreement can only be 0, 1, 2 or 3 shared neighbours, so a single person's score rests on very few comparisons.
• One short biography per person, all from one source. The results describe how these texts are written as much as how the models behave.
• Similarity scores and map positions are per model. Compare rankings and who is near whom, not raw scores or positions across models.