Ex-OpenAI researcher says neural networks may be far more interpretable than thought

Nick Cammarata of Goodfire, who helped establish mechanistic interpretability at OpenAI, published a 26,000-word post arguing the "black box" view of neural networks is overstated. He reports that roughly 90% of neurons in Qwen3-32B track a single main concept, and that the famous cat-car neuron actually responded to a shared V shape.

• Cammarata published a 26,000-word post on the nature of neural networks
• About 90% of Qwen3-32B neurons track one main concept
• The famous cat-car neuron responded to a shared V shape
• Critics note the post lacks rigorous empirical evaluation