Helen Toner在播客访谈中重申了对人工智能首席执行官缺乏信任的观点。1 Helen Toner是Georgetown大学CSET执行主任,曾作为OpenAI四名董事之一投票解雇Sam Altman,该事件被称为“the blip”。1 Toner指出此次事件暴露了OpenAI治理机制的失效及问责缺失。1 她强调,若将未来寄托于任何个人的可信度都将陷入困境,因此需要更广泛的结构和多方视角。1
针对当前AI技术的发展前景,Toner深入探讨了前沿AI公司(如OpenAI和Anthropic)正试图实现的“递归自我改进”(RSI),即利用AI加速研发以最终实现超越人类的超级智能。1 对于AGI是否已经到来的争论,Toner认为该术语目前变得毫无帮助,因为现有模型虽满足部分定义,但远未达到全面通用或意识觉醒的标准。1
在地缘政治与监管层面,Toner分析了Anthropic即将上市(IPO)带来的市场压力,以及美国与中国在AI监管上的威胁认知出现初步趋同迹象,特别是在网络攻击和失控风险方面。1 面对技术加速带来的挑战,Toner不主张完全停止训练新模型,而是建议设定更高的安全信心门槛,让企业在达到标准前投入更多资源进行对齐和控制研究。1
Helen Toner, the executive director of Georgetown University’s Center for Security and Emerging Technology (CSET), reiterated her skepticism toward relying on the integrity of individual artificial intelligence chief executives. As one of four former OpenAI board members who voted to dismiss Sam Altman—an event often referred to as "the blip"—Toner argued that placing humanity’s future in the hands of any single leader is a flawed strategy 1. She emphasized the necessity for broader structural safeguards and diverse perspectives to ensure accountability within the industry 1.
The discussion centered on the governance failures at OpenAI and the emerging technical trajectory known as recursive self-improvement (RSI). Toner noted that frontier companies like OpenAI and Anthropic are pursuing RSI, a method where AI systems accelerate their own research and development to potentially achieve superintelligence beyond human capabilities 1. While some models meet partial definitions of Artificial General Intelligence (AGI), Toner described the term itself as increasingly unhelpful because current systems have not achieved comprehensive generality or consciousness 1. She pointed out that the industry currently exhibits blind optimism regarding these developments 1.
Beyond internal governance, Toner addressed geopolitical risks and market pressures facing AI firms. With Anthropic preparing for an initial public offering (IPO), the company will face new market demands that could influence its safety priorities 1. Concurrently, there are early signs of convergence between the United States and China in recognizing threats related to cyberattacks and loss-of-control scenarios in AI regulation 1. Rather than advocating for a complete halt to training new models, Toner proposed establishing higher confidence thresholds for safety, requiring companies to invest more resources in alignment and control research before deploying advanced systems 1.
评论
还没有评论,欢迎留下第一条。