Modular Pretraining Enables Access Control

According to LessWrong, researchers at AE Studio, working with Anthropic, have proposed a new way to lock down the dangerous knowledge inside AI models. The problem: today's safeguards, like teaching a model to refuse harmful requests or screening queries with classifiers, sit on top of knowledge the model still possesses, so they can be jailbroken. Their method, called Gradient-Routed Auxiliary Modules, or GRAM, instead tries to wall off risky know-how into separate switchable modules baked into the model during training. Flip a module off and the model effectively forgets a topic, whether that's virology, cybersecurity, or nuclear physics. The appeal is cost: rather than training separate filtered models for every trust level, a single GRAM model can be reconfigured to mimic several. The team reports the effect holds across models from fifty million to five billion parameters, and even improves with scale. They stress this is preliminary work that has not been applied to Anthropic's production models.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton