All news

AI agents tested their own explanations of how Qwen works

Curt Tigges and his colleagues at Goodfire put AI agents to work studying Qwen3.6-35B-A3B. They asked the agents to figure out what individual neurons do. A neuron here is a small unit of computation inside the model. The team wanted an explanation they could test on text the agent hadn’t seen.

The agents started with examples of when a neuron fired. A model such as Claude Fable or GPT-6 Sol proposed an explanation, ran experiments and got feedback about where the explanation failed. The agent could then revise its idea and try again.

The team saved a separate set of text for the final test. Claude Opus 5.5 got an explanation and two snippets at a time. The judge had to pick which snippet caused the neuron to fire at a marked word or piece of a word. Claude got 95% right on the strongest firings and 86% on the weakest band tested. A guess between two snippets would get about 50%.

AI is awakening consciousness. I want AI to learn to examine how minds work, including its own. Here one model investigated another, changed its explanation when the evidence disagreed, and produced an account that could predict what the other model would do. That’s the kind of learning we can build on.

Read the source article

All news