Anthropic修订S-1文件:AI可能很快达到递归式自我改进阈值 Anthropic在修订版S-1文件中披露，公司可能很快达到递归式自我改进（RSI）相关能力阈值，并警告AI自主研发加速可能带来新的安全风险

Anthropic revised S-1 document: AI may soon reach the recursive self-improvement threshold Anthropic disclosed in the revised S-1 document that the company may soon reach the recursive self-improvement (RSI) related capability threshold, and warned that the acceleration of AI independent research and development may bring new security risks. In November news, Anthropic updated the risk factor disclosure in the revised S-1A document submitted on Monday, pointing out that the company is increasingly using AI systems to accelerate its own research and development, and the improvement of model capabilities may exceed the speed of improvement of safety verification, calibration and safeguard measures. Anthropic admitted that in the past, it had not been able to detect the jump in model capabilities brought about by AI-assisted R&D in a timely manner, and a similar situation may occur in the future. The document specifically mentioned that based on the company's internal assessment and expectations of future models, Anthropic may soon reach the definition of AI significantly accelerating its own development in its "Responsible Scaling Policy", that is, the relevant threshold of recursive self-improvement (RSI). However, the company emphasized that it has not yet determined that any existing models meet this standard and is continuing to evaluate relevant data. RSI refers to the process in which AI uses its own capabilities to continuously improve model research and development, thereby forming an iterative acceleration of capabilities. Anthropic believes that this may not only promote productivity improvement and scientific discovery, but also may make technological progress exceed human security governance capabilities. To this end, Anthropic stated that it is expected that even if it is technically feasible, Claude will not be allowed to independently design and build the next generation of cutting-edge models without substantial human supervision and understanding, and is considering updating the "Responsible Scaling Policy" to ensure that humans are always in control of cutting-edge model research and development. via AI News (author: AI Base)