A researcher claims large language models may be far easier to understand, and so to monitor and control, than people thought.

Nick Cammarata of AI start-up Goodfire made the case in a 26,000-word blog post published on 15 October.

Titled “Neural Networks Are Already Clean,” it has sparked both wild speculation and cautious celebration, MIT Technology Review reported.

If true, it would matter hugely for AI safety, since explaining why a model acted matters for accountability.

Cammarata said he spent about ten hours a day, seven days a week, for most of a year looking at neurons.

He uses new tools and teams of AI agents to help make sense of what he sees inside networks.

He says he mapped whole circuits of activity, finding interpretable structure where others saw only noise.

His strongest results come from image models and early language-model audits. He has not yet fully audited frontier systems.

Read the full article for more details on the neural network legibility claim.