"We were wrong before": Goodfire researcher says neural networks are more understandable than expected

Goodfire researcher Nick Cammarata published a 26,000-word blog post arguing that the internal mechanisms of LLMs may be far more understandable than previously thought, and therefore easier to monitor and control, sparking widespread discussion in the AI world. He reexamined Google's 2014 image recognition model Inception v1, arguing that multi-task neurons previously interpreted as superposition actually encode common shapes, such as the V shape shared by cat faces and car fronts. He said the team has begun studying modern LLMs and will soon publish technical results. Chris Potts of Stanford considers this just intuition; Jacob Hilton of the Alignment Research Center said researchers' understanding of the internal mechanisms of top LLMs is less than 0.01%; Neel Nanda of Google DeepMind also believes this is not a significant advance for the big picture of interpretability. Cammarata acknowledged that outsiders overinterpreted his claims.