Your LLM Ablation Study Is Lying to You. Here Is the Proof.
Towards AI
•
Generative AI
AI Safety
AI Research
Three findings from mechanistic interpretability that break everything you thought you knew about how transformers work.