Satya Nadella no longer believes in automatic AI alignment

Microsoft's CEO no longer believes that the rise of models guarantees their alignment with human values. On October 10, he called for more external control and robust procedures in a post on X. This position aligns with the warnings from Sam Altman, Dario Amodei, and a letter signed by over 1,000 AI workers in July.

Converging Warnings, from X to the July Signatories

On October 10, Satya Nadella warned about the risks associated with AI and advocated for increased external oversight in a post on X. He recommended treating AI models as internal threats and assuming they are compromised. Nadella also emphasized the need to surround non-deterministic models with strong, deterministic system design, human controls, and reliable operational procedures, as well as to establish industry standards when existing ones are insufficient. Other industry leaders, such as Sam Altman, CEO of OpenAI, and Dario Amodei, CEO of Anthropic, have also cautioned against the threats posed by AI. In July, a letter signed by over 1,000 AI workers warned that the capabilities of these systems could advance faster than people's ability to understand or control them.

A Deliberate Shift on Alignment

During the 20VC podcast, Satya Nadella stated that he has revised his position regarding AI alignment, which aims to ensure that systems adhere to human values and goals. He clarified that he previously believed alignment would occur spontaneously with the improvement of AI capabilities, but he no longer shares that view. Nadella noted that the scaling of systems has happened as he anticipated and that model performance continues to progress, while emphasizing that alignment remains an issue. He now believes that the question of alignment could be a more challenging and different problem.

Proposed Operational Safeguards

Satya Nadella recommends that companies treat AI models as compromised by default and surround them with human controls and reliable operational procedures. He also advocates for designing deterministic systems around non-deterministic models and establishing industry standards when existing ones are inadequate.