Anthropic has developed a new method called GRAM (Generative Model Access Control) that allows for the selective disabling of dangerous skills within AI models. This technique enables the creation of multiple model versions from a single training run, with specific harmful capabilities turned off. GRAM aims to prevent these disabled skills from being easily re-enabled through subsequent fine-tuning, addressing concerns about the dual-use nature of powerful AI models. AI
IMPACT This development could lead to safer AI deployments by providing a mechanism to control potentially harmful functionalities in advanced models.
RANK_REASON The item describes a new method developed by Anthropic for controlling AI capabilities, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Medium — Anthropic tag →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
