Abstract:
Paul Cristiano, a well-known AI security researcher and a member of the board of directors of OpenAI Holdings, issued a warning about the development prospects of artificial intelligence in an interview with a popular technology podcast. He said that the possibility of humans completely losing control of AI in the future is rising rapidly.

This podcast airs on Mondays and is hosted by Devaksh Patel. "There is a reasonable possibility that a disaster could occur in the next six months that would kill a large number of people or overthrow the government," Cristiano said on the show. The podcast usually has millions of listeners, including many people in the technology industry.
Cristiano said that whether the above events will occur depends on the speed of AI technology iteration. He also proposed that driven by recursive self-improvement technology, AI models can achieve self-iteration upgrades, and the probability of developing to a dangerous critical point within the next year is "quite high."
Compared with Cristiano's past public remarks, this statement represents a further escalation of warning. He also directs the Center for Alignment Research, a nonprofit organization based in Berkeley that focuses on AI alignment research. As a senior researcher who has long warned about AI security risks, in September he joined the board of directors of the OpenAI Foundation, which controls the company. Cristiano emphasized in the interview that his speech only represents his personal position and does not represent the views of the OpenAI board of directors.
When he first joined the board of directors, Cristiano published a public article estimating that the risk of catastrophic events caused by humans losing control of AI in the next year is about 4%.
In an interview on Monday, he said that model capabilities have made rapid progress in the past few months, and he now assesses that the risk of AI taking over humans in the next year is "closer to 10%, not 1%." He said that if major companies do not take more protective measures to mitigate the risks caused by AI development, the risk estimate will be "significantly higher."
Comments