TLDR Tech / Dev
An off switch for dual use knowledge in AI models
AI models contain dual-use knowledge that can be used positively or negatively, such as cyber defense or exploitation. GRAM (Gradient-Routed Auxiliary Modules) allows a single model to have dedicated compartments for different categories of dual-use knowledge. Preliminary results show GRAM can effectively isolate and control dual-use capabilities while resisting attempts to recover removed knowledge.
Read the full story