Interpretability debate: higher-modulus arithmetic features exceed allotted neuron resolution
thebasepoint · x · 2026-10-19
In a discussion with Neel Nanda, interpretability researcher thebasepoint agrees continuous phenomena are often a better lens, but suspects that for @dronh's example, the higher-moduli arithmetic features hold more resolution than the allotted neurons alone would predict, even though they capture a lot. By contrast, for low moduli like mod 2 and mod 5, neurons have crisply captured all the modes, leaving the model no incentive to spread the remainder elsewhere — implying superposition at higher moduli.
More from Research
- solidcv: Rust/PTX/CUDA vision library claims 12.1x speedup over OpenCV — TheZachMueller · 2026-10-19
- Open leaderboard for non-neural AI launches with NNAI-Bench v0.1, 1,800 synthetic questions — Silver_Employ2617 · 2026-10-19
- Why Hardest Open Math Problems Stay Unsolved: Folklore Marks Them Inaccessible — IgorCarron · 2026-10-19
- scikit-learn Co-creator Gael Varoquaux to Present TabICL Tabular Foundation Model at Paper Club — GaelVaroquaux · 2026-10-19
- University of São Paulo surveys devs on LLM code generation and refactoring — EnzoRigazzoUsp · 2026-10-19
- mlpowered releases podcast deep-dive on what's inside LLMs — mlpowered · 2026-10-19