
The Neuron Basis: Goodfire's Curt Tigges on Agentic Interp, Monosemanticity & Reverse Engineering
AI 摘要 Individual neurons in modern large language models, particularly within gated MLP architectures like SWIGLU, exhibit significantly higher interpretability than previously assumed. Rather than functioning as a chaotic, distributed mess, most neurons serve specific, identifiable…
- Neurons as an Interpretable Basis
- Neural Networks are Already Clean
- Sparse Autoencoders (SAE)
1 h 56 minEnglish摘要 · 转录


























