Anthropic develops GRAM to remove risky knowledge from AI
Technology
Anthropic has come up with a new way for AI to forget risky information, such as hacking tricks or sensitive scientific details.
Its system, called GRAM, lets the company pull out certain knowledge from an AI after it is trained, so the model can't share potentially harmful information but still works as usual.
GRAM remains research stage, tests promising
Right now, GRAM is just in the research stage and has not made it into Anthropic's main products, such as Claude, yet.
But early tests look promising, and Anthropic hopes this technology will help make advanced AI a lot safer as concerns about misuse keep growing.