Mikhail Belkin
An exciting article from (the always insightful) Mikhail Belkin and colleagues got its first release in Science Magazine a few days ago (paywall). A preprint version from May 2023 is available (link).
The main concept: "features extracted by a given neural network layer are proportional to the average gradient outer product (AGOP) with respect to the input to this layer". The acronym "AGOP" does not appear in the preprint but is the central mathematical operator that unifies feature learning across many neural network architectures, including transformers, kernel machines, CNNs, RNNs, ...
What strikes me is the simplicity of the concept and operator, which makes this result even more beautiful.