AIForbes Innovationabout 2 hours ago
Modular AI Safety Method Targets Malicious Knowledge
1 min read
A new method, "Gradient Routed Auxiliary Modules" (GRAM), isolates dangerous knowledge in generative AI into dedicated modules during initial training, enabling an on/off switch for access control. Preliminary research by Anthropic and AE Studio demonstrates the concept on smaller models. Scalability and coherence risks remain unresolved.
Level
Hype check
Tap to vote and see what everyone thinks.
