All writing
Reading the Residual Stream
2 posts

Sparse Autoencoders Can't Measure Generation-Time Behavior. That's Not a Bug.
Why sycophancy SAE features have Cohen's d=9.9 but hallucination detection fails. The answer turned out to be deeper than measurement timing.
aiinterpretabilitysparse-autoencoders
Capital-of Is Not a Single SAE Feature. So I Built a Mutation Loop to Find What Is.
SAE features can't isolate relations in Gemma-2-2B. I built a mutation-selection loop that can. The bottleneck was tokenization.
aiinterpretabilityllmsparse-autoencoders