CambridgeGPT-6 Astra edited another AI model and boosted its maths.

Interpretability researchers hope this makes AI safer, but Oxford’s Fazl Barez says that is largely untested.

Astra read the raw internal values of some other models, predicted what their parts do, then edited them.

Barez would be surprised if Astra reads models in ways beyond existing human abilities.

The Massachusetts Institute of Technology team’s 2 November preprint has not been peer reviewed.

How each outlet framed it
Nature
reports an AI model (GPT-6 Astra) reading and editing another network's internal values to boost its maths, framing it as a mechanic-fixes-an-engine interpretability route

Sources: Nature