saey-dictionary-sizes-a0-a1-a2
IN premise — summaries/2026/08/24/bricken-2023-monosemanticity-chunk-5.md
Created 2026-08-25T02:57:55+00:00
The dictionary sizes across models are: A/0 = 512 features, A/1 = 4,096 features, A/2 = 16,384 features.
Summary
The three A-series models differ dramatically in how many distinct features they track: 512, 4,096, and 16,384 respectively. This sets the baseline scale for the system, meaning any comparison or reasoning between these models has to account for the fact that the larger ones carry roughly 4x to 32x more representational capacity than the smaller ones.