Technical interest · T-05
Representation learning
What a model's internal space captures, and whether that space is useful for anything downstream.
At MIT CSAIL I designed a temporal and cross-asset representation for stock data and benchmarked it against eight others on downstream targets. The inter-agent misalignment project asks the same question of an LLM's activations. This site is one too: every point is a sentence embedding projected onto two principal components.
Nearest neighbors · any layercosine distance in the full space