Researcher Questions SAE Interpretability with RL Training Evidence

Interpretability researcher voooooogel challenges the sparsity assumption behind SAEs, citing persistently poor results and firsthand RL training observations where two 'user turn start' features encoded differences in user types.

2026-10-18 ~ 2026-10-18 · 2 related posts

• SAE researcher questions sparse coding premise and Claude-based neuron autolabeling — voooooogel · 2026-10-18
• First-hand: two SAE features both labeled 'user turn start' encode user-type differences — voooooogel · 2026-10-18