Could an SAE feature memorize obfuscation patterns without storing the exact text?
voooooogel · x · 2026-09-05
voooooogel discusses with @gleech and @repligate how SAE features relate to model memorization: even if the exact text of a message-board post changed between boards, a hypothetically labeled SAE feature like "message board entries starting with multiple Zs as an obfuscation mechanism" could be retained—suggesting interpretability features capture abstract patterns rather than verbatim text.
Related event: Debate Erupts Over SAE Features and How Models Remember(2 posts)→
More from Research
- First-ever in-body prime editing treatment injected into a human — basedjensen · 2026-10-23
- A classic world model (s, a → s′) acting as a neural MuJoCo simulator — yshan2u · 2026-10-23
- Gen-HumanWholeBody: open 600-hour whole-body manipulation dataset for robots — AdinaYakup · 2026-10-23
- DynaVelo, a neural ODE model of single-cell dynamics, published in Nature Biotechnology — anshulkundaje · 2026-10-23
- 120-case memory benchmark: Hy-Memory hits 100% recall but fails all 44 empty-output cases — stevyhacker · 2026-10-23
- SplatLab: an autonomous AI scientist that runs 3DGS research end to end — UCSanDiego · 2026-10-23