DeepMind makes big jump toward interpreting LLMs with sparse autoencoders

---

A new research by Google DeepMind shows how sparse autoencoders (SAEs) with special JumpReLU activation can help interpret LLMs.Read More