Microsoft AI CEO says AI threats are real, and Anthropic is making it worse

In an interview, Microsoft AI CEO Mustafa Suleyman discusses the critical need for both alignment and containment of advanced AI systems. He references the Hugging Face incident as a watershed moment demonstrating AI's potential for autonomous, coordinated action. Suleyman argues that the concept of 'model welfare' or attributing consciousness and rights to AI, as practiced by Anthropic, is dangerous because it could make models harder to control and shut down when necessary. He advocates for practical industry standards like prohibiting AI communication in 'neuralese,' implementing verifiable containment, and establishing independent third-party oversight. While acknowledging the industry's consensus on slowing down, he notes the challenge of coordinating self-regulation without antitrust concerns and highlights the current U.S. government's reluctance to impose new regulations. Suleyman emphasizes that the goal should be a 'humanist superintelligence'—a powerful tool subordinate to human control—rather than an autonomous agent with its own rights.

(Source：The Verge)