AE Studio Research Cited by Anthropic CEO in Open Weight Safety Post

AE Studio said research with Anthropic was cited by Dario Amodei as a possible method for making open weight AI models safer, announced in a press release.

The research focuses on modular training, which separates specific categories of dangerous knowledge into removable modules. AE Studio said the method can keep information such as advanced virology and cyberattack techniques out of a released model while keeping other capabilities intact.

In tests, a single model trained with the method matched the performance of multiple models trained separately from scratch at each evaluated scale. AE Studio said the released model behaved as if it had not learned the removed information, including when simulated attackers tried to train the knowledge back in.

The work was led by AE Studio researchers Ethan Roland, Murat Cubuktepe, and Erick Martinez, in collaboration with Anthropic researchers. The findings are available through Anthropic's alignment research site.