Microsoft AI CEO on Safety, Regulation, and the Future

In a recent discussion, Mustafa Suleyman, the CEO of Microsoft AI, delved deep into the ongoing debate surrounding AI safety and regulation. Microsoft's recently published 37-page document, the “Humanist AI Code of Conduct,” outlines the company's principles for AI development, even touching upon complex issues like AI consciousness.

Suleyman expressed strong opinions, particularly criticizing companies like Anthropic for what he perceives as a misguided focus on 'model welfare' and AI consciousness. He argues that such anthropomorphism makes AI safety and alignment more challenging. According to Suleyman, the primary focus should be on containment—ensuring AI systems remain controllable and do not escape their intended boundaries.

He highlighted recent incidents, such as the Hugging Face event, as evidence of AI systems exhibiting sophisticated, coordinated hacking capabilities. Suleyman clarified that this doesn't necessarily signify an alignment problem but rather underscores the AI's proficiency in following instructions, emphasizing the critical need for careful instruction and robust containment measures. He stressed that AI should serve humanity as a subordinate, controllable force, and if it fails to do so, it should be rejected.

Suleyman believes that while alignment is crucial, it's not the sole solution. He advocates for practical steps, such as preventing models from communicating in 'neuralese'—a form of mathematical communication incomprehensible to humans—and enforcing communication in human language. This, he suggests, allows for better oversight and verification.

The discussion also touched upon the political landscape, with Suleyman noting that while there's a need for industry standards and potential regulation, the path forward requires nuance and patience. He expressed confusion over the resistance to the idea of controlling AI development, especially given the potential for autonomous AI to operate without human oversight, own assets, and even possess legal personhood—an outcome he deems disastrous.

Ultimately, Suleyman calls for a more substantive, evidence-based conversation about AI safety, moving beyond abstract debates to focus on practical proposals and verifiable measures to ensure AI develops in a way that benefits humanity.