πΏπ¦ English, 6 November 2026, interpretability
AI's Black Box May Not Be So Black After All
A former OpenAI researcher claims AI models are far easier to understand than previously thought.
Nick Cammarata, a researcher at AI start-up Goodfire, argues that AI models are not as complex as previously assumed, and that individual neurons may have one main job, rather than juggling multiple unrelated concepts.
He points to a 2014 Google image-recognition model, Inception v1, where a neuron responds to both cats and cars, but claims that the common feature is a shared V shape, rather than the concepts themselves.
Preliminary audits of Alibaba’s Qwen3-32B model suggest that roughly 90% of neurons have one main job, with only 2-4% remaining uninterpretable.