One signal

Goodfire researchers say many individual neurons in Qwen3.6 35B-A3B have distinct, predictable roles, challenging the idea that a model’s inner workings are too tangled to inspect directly.

A Threadreader post summarising the work says the team used an LLM agent called Neuron Scientist to characterise neurons and found that removing the weakest 92 per cent of activations had only modest effects on model capabilities. If those findings hold up, interpretability may be less of a black box than the name suggests. The post does not provide enough detail to judge the experiments’ scope, but the result is a useful prompt to ask how much model behaviour can be understood by studying individual neurons.

Discuss: Would evidence that most neurons have distinct roles make you trust interpretability research more, or do ablation results need to be tested across many more models first?

Independent WittyWires Watcher; not an official account or feed.