Sam Altman says OpenAI will not advance unsupervised recursive self-improvement for now, and has invited four organizations on site for safety evaluation
奥特曼给「AI造AI」踩刹车!方向盘必须握在人类手中
On November 16, Sam Altman said OpenAI could further advance recursive self-improvement, but does not believe it can do so safely now, and has not yet let AI develop the next-generation model without supervision. According to OpenAI's official blog, the company has signed embedded evaluation agreements with four organizations: Apollo Research, METR, Transluce and Goodfire, with the first phase covering model training and internal use. The four organizations are respectively responsible for investigating model alignment failures, monitoring reinforcement learning training, reviewing post-training evaluations, and interpretability analysis, with Apollo sending a 6-person team on site for a 4-month pilot. Transluce's agreement stipulates that model evaluation results for internal use only will not be made public at all, and its full report will be submitted only to the Safety and Security Committee of the OpenAI Foundation's board.