You say Potato, I say 110

Tue, 30 Jun 2026 00:00:00 +0000

I was re-reading the Rectified LpJEPA paper early this week and it got me thinking - this paper allows us to regularize model outputs to match a desired global distribution. What other embedding spaces might we care about?

The first thing to come to mind was the concept of semantic hashing. This is in many ways the pre-cursor to RAG systems, and focuses on the generation of a series of bits associated with the semantic meaning of some input data, where an efficient bit-counting metric can be used (e.g. Hamming distance) to identify documents similar to a query. This use of a binary representation and an efficient metric should allow incredibly fast retrieval when combined with algorithms such as PyNNDescent (or HNSW, I just love PyNNDescent’s explainer page).

Representation-Learning on tl

You say Potato, I say 110