Interp debate: denser wiring at scale is the real question, not 'more stuff'
banburismus_ · x · 2026-10-18
@banburismus disentangles two conflated issues in the interpretability-at-scale debate: (1) more parameters mean denser interconnection of neurons, vs (2) simply 'more stuff'. He argues (2) is a non-issue — just spend more tokens — while evidence that larger networks are cleaner to interpret cuts against (1). He also wonders why recent papers haven't shifted the dominant transcoder/attribution-graph view.
More from Research
- Map of 171,250 RL papers shows deep RL and robotics dominate the field — CatAstro_Piyush · 2026-10-18
- KNOWS Benchmark: Best Web Agent Fully Succeeds on Fewer Than 3% of Complex Tasks — anmarasovic · 2026-10-18
- Open-source Engram memory engine, modeled on databases, matches Mem0 on BEAM — phecdaYm · 2026-10-18
- Mathematician: Pages of Routine Analysis Work Well, and AI Makes Writing Them Easier — jessi_cata · 2026-10-18
- Chinese researchers argue harness engineering is over: the eval set becomes the job — blaizedsouza · 2026-10-18
- Andrew Ng: 'AI for Science' will become meaningless as AI becomes every scientist's tool — andrewgwils · 2026-10-18