Method A
Sparse autoencoders
2,560
encode
16,384
≈77 fire at once
Δ = original − reworded
decode
2,560
rebuilt
R² 0.9972
Feature 3818
the presence-framing
register, layer 17
Feature 12139
selects the answer, layer 29
register replicates with own transcoders · AUROC 0.948
How it works
‹
›
Play in sync