OpenAI Offers a Peek Inside the Guts of ChatGPT

OpenAI has released a research paper on reverse engineering AI models in response to criticism about the potential risks of their technology. This comes days after former employees accused the company of being reckless. The paper explains a method for understanding how AI models store certain concepts, which can help identify misbehaviors in systems like ChatGPT. The research also underscores recent internal turmoil, including the disbanding of OpenAI's "superalignment" team and the departure of notable figures. This new technique, demonstrated on GPT-4, aims to make AI models more interpretable and controllable, although further refinement is needed. The research aligns with similar efforts in the field to make AI systems safer and more transparent.

Opens on Wired