Do AI Neurons Follow Scaling Laws? Rosetta Neurons, Superposition, and the Black-Box Caveat

Do larger AI models develop more shared, interpretable neurons, or do their interiors stay
idiosyncratic?

This video explains Neuron Populations Exhibit Divergent Selectivity with Scale by Dravid, Bahri,
Efros, and Gandelsman. The work studies Rosetta Neurons: units whose activation patterns recur across
independently trained models. The key claim is not that the black box is solved, but that
reproducible neuron populations can be measured as a scaling observable.

We cover:

- How Rosetta Neurons are found with activation alignment and mutual nearest-neighbor matching
- Why their count grows sublinearly with model size
- The capacity-allocation picture behind the scaling law
- The “neuron polarization” effect: cleaner Rosetta neurons versus a larger mixed background
- Why the result is interesting, and where it remains observational rather than fully mechanistic

Paper: https://arxiv.org/abs/2606.03990
Project page: https://avdravid.github.io/rosetta-neuron-scaling/
Code: https://github.com/avdravid/rosetta-neuron-scaling

Видео Do AI Neurons Follow Scaling Laws? Rosetta Neurons, Superposition, and the Black-Box Caveat канала Xiaol.x