The Representation Layer

Field notes on a young field applying mechanistic interpretability to retrieval models: where relevance signals live inside a bi-encoder, how a cross-encoder rebuilds BM25's ingredients unprompted, and why a dense model encodes more than its dot product c…

Abliteration deletes a model's refusal with a difference of means and some linear algebra.

A tour of the micro debate on representational convergence, from lines and categories to circles, helices and spectra, and the gap between a structure being present, readable, and actually used.

A tour of the macro debate on representational convergence, from the Platonic Representation Hypothesis to scaled critiques and the Aristotelian view.

a new encoder-only model

continuing the journey of embeddings, exploring the llm era

an overview of the journey of embeddings

what am i trying to do in the series words and their mappings?

a brief introduction to the series words and their mappings

Elasticsearch is a fast, scalable search and analytics engine built on Apache Lucene, known for real-time search and powerful querying.