Last winter I spent three weeks staring at the same t-SNE plot. Twelve million embeddings from a production retrieval model, projected down to two dimensions, glowing violet against the black render window. I had built the visualization to debug a recall regression. Instead I found myself asking a stranger question: what does the manifold itself encode that the training loss never measured?
The curse of interpretability
We treat distance in embedding space as semantic proximity — closer equals more similar. But the geometry of the manifold is an artifact of the loss function, not a mirror of meaning. Cosine similarity measures angular proximity on a hypersphere. Euclidean distance measures straight-line separation through regions the manifold never visits. Neither one captures what the model actually internalized during pretraining.
Distance is the first lie every manifold tells.