Neel Nanda on the interp debate: neuron attribution graphs are practical, not a reverse-engineering breakthrough
Jsevillamol · x · 2026-10-16
Neel Nanda weighed in on the current interpretability discussion: he endorses exploratory, neuron-based interpretability (e.g. neuron attribution graphs) for pragmatic questions like "why did the model do a sketchy thing?", saying this work plus earlier Transluce research shows the approach works and has advantages over SAEs/transcoders. But he rejects grander inferences that this constitutes major progress toward ambitious reverse engineering or "solving" interpretability.
More from Models
- Artificial Analysis Optima now lets you build custom LLM benchmarks in 3 steps — ArtificialAnlys · 2026-10-16
- Stanford CS336 lecture notes unpack MoE scaling: 64 experts stored, still 2P active per token — stanfordnlp · 2026-10-16
- I tested Jev's new image input on 48 COCO photos — 408 calls, key gotchas — SeveralSeat2176 · 2026-10-16
- Chandra 2.5 cuts OCR script confusions 7x, beating Gemini 3.8 Flash and Opus 5.5 — VikParuchuri · 2026-10-16
- Script confusion bug confirmed in Gemini Flash 3.8 and Claude Opus 5.5 — VikParuchuri · 2026-10-16
- Why VLMs sometimes output 罪 instead of 'vain': an OCR investigation — VikParuchuri · 2026-10-16