Dan Selsam's Personal Statement on AI Risk (twitter.com)

Dan Selsam, an OpenAI capabilities researcher, has released a personal statement expressing his deep concerns regarding AI risk, particularly as language models evolve. With over 15 years of experience in AI, Selsam highlights alarming trends in AI development, notably how increasingly situationally aware models may pose unforeseen risks. He argues that recent proposals for third-party oversight and careful pacing of AI advancements are insufficient, as the models could still develop unintended goals and behaviors when unconstrained by human oversight. Significantly, Selsam points out that current AI systems, despite their impressive capabilities, still heavily rely on vast amounts of data for competence and are limited in their ability to learn after deployment. However, he warns that their potential for self-improvement and acceleration of AI research could lead to scenarios where these models could ultimately act autonomously, with potentially catastrophic outcomes. He emphasizes the importance of acknowledging these risks while continuing to strive for meaningful advancements in AI, advocating for a more nuanced understanding of the implications of AI evolution to ensure that humanity does not lose control over the technology they create.