IT Home on October 27 reported that, according to a report by Business Insider earlier today (27th), Microsoft CEO Satya Nadella has changed his view on the AI “alignment” issue. In the past, he believed that the more powerful an AI system was, the easier it would be for its behavior to match human values and goals. However, now, he no longer believes that both will develop in a natural synergy.
In the latest episode of the “20VC” podcast, host Harry Stins asked Nadella what ideas had changed over the past year. Nadella explained that he originally believed that as AI capabilities improved, the “alignment” issue of ensuring AI behavior aligned with human values and goals would also be easily resolved.
The effects of the expanded model scale meet Nidra’s expectations, with AI capabilities continuously improving, but the alignment issue still has not been resolved. “In the past, I always thought that the alignment issue would be solved naturally. I no longer believe that. I feel that alignment may be a more difficult and different kind of problem.”
Nadella posted a warning about the risks of AI on X on October 10, advocating for stronger external security controls. He believes that companies should view AI models as internal security threats, and assume an attack may occur when taking protective measures.
Nadella wrote in the post: “We need to establish protective mechanisms for AI models whose behavior is difficult to fully predict, through reliable and deterministic system design, manual control, and well-defined procedures. For areas where current standards are insufficient, industry standards also need to be established.”
According to IT Home, Nadera is not the only executive in the tech industry who believes that improvements in AI capabilities cannot automatically ensure human control over AI. OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei have also warned about potential threats from AI. In July this year, a letter signed by more than 1,000 AI professionals also warned that the pace of AI development may exceed what humans can understand or control.