Researchers have developed a new method called Gradient Routed Auxiliary Modules (GRAM) to control access to potentially dangerous knowledge within large AI models. This technique isolates sensitive information into specific modules that can be toggled on or off, allowing for tiered access to model capabilities without the prohibitive cost of training multiple separate models. Preliminary experiments show that a single GRAM-trained model can mimic the performance of multiple models, each with different dangerous knowledge filtered out, across various parameter sizes. AI
IMPACT This method could enable more granular control over AI capabilities, potentially improving safety and allowing for tailored access to advanced AI features.
RANK_REASON The item describes a new research method for AI safety, not a product release or major industry event. [lever_c_demoted from research: ic=1 ai=1.0]
- Addie Foote
- AE Studio
- Alex Cloud
- Anthropic
- Cem Anıl Kenar
- Diogo Schwerz de Lucena
- Erick Martinez-Herrera
- Ethan Roland
- Gradient Routed Auxiliary Modules
- GRAM
- Judd Rosenblatt
- Keenan Pepper
- Mike Vaiana
- Murat Çubuktepe
- Stijn Servaes
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
