Stanford's Potts weighs in on superposition critique: RAVEL author says alternative explanations viable
gregd_nlp · x · 2026-10-19
Responding to Nick Cammarata's thread claiming interpretability is the most important scientific problem, Stanford's Christopher Potts published his press comments: (1) he welcomes critical reassessment of superposition arguments and thinks alternative explanations may be viable; (2) despite 40 interpretability papers, he never worked on superposition per se; (3) his RAVEL benchmark (with the MDAS method, SOTA on Llama2-7B) measures feature disentanglement, finding many polysemantic neurons but making no claims about superposition specifically — he'd like to see RAVEL run on newer models. Context: Cammarata says modern agentic tools make the vast majority of neurons legible even in trillion-parameter models, vs 2% in early OpenAI work.
More from Research
- Probe repo: GPT-4, with pre-GRPO cutoff, predicts DeepSeekMath's GRPO from anonymized text — CatAstro_Piyush · 2026-10-20
- Mistral launches Timestral time-series foundation models, zero-shot SOTA on GIFT-Eval — victormustar · 2026-10-20
- MD simulations reveal striking hydration dynamics in daraxonrasib tri-complex pockets — CatAstro_Piyush · 2026-10-20
- Anchor Divergences: contextual semantic similarity for vector search without changing embeddings — victorveitch · 2026-10-20
- Marburg's Sebastian Wild releases full Advanced Algorithms course videos and materials — caglar_ee · 2026-10-20
- ICLR paper's Future Work section just points to a NEXT_VERSION.md in the supplement — CSProfKGD · 2026-10-20