Researcher says much SAE work added little to understanding how models actually work
thebasepoint · x · 2026-09-27
- In a discussion thread, the author advises against hillclimbing a method for its own sake: focus on what a method tells you about how models work — what information is accessible in what arrangement, and how to exploit it to explain phenomena or build new methods.
- He explicitly criticizes a sea of SAE (sparse autoencoder) work as having added little to our understanding of models and their mechanisms.
Related event: Researchers Call SAE Work Limited, Urge Mechanistic Understanding(3 posts)→
More from Research
- Rethinking on-policy distillation: researchers propose OLIVE, letting students learn from teacher continuations — May_F1_ · 2026-09-27
- CoRL 2026 workshop on continually self-improving robots opens call for papers, due Sep 28 — PeterStone_TX · 2026-09-27
- Martin Casado recommends the best talk on in-context learning, a first-principles view of LLMs — AccBalanced · 2026-09-27
- Functional Gradient Descent with Adaptive Representations accepted at NeurIPS — CatAstro_Piyush · 2026-09-27
- Tailored ASR for Japanese speaking assessment cuts mora error rate from 12.3% to 7.1% — tkasasagi · 2026-09-27
- kalomaze proposes testing which nanogpt tricks survive causal NTP over DCT coefficients — kalomaze · 2026-09-27