Nadella: I Thought AI Alignment Would Emerge Naturally. I Was Wrong
In the 20VC podcast, Microsoft CEO Satya Nadella revealed that his biggest shift in thinking over the past 12 months has been about AI alignment: he once believed that alignment would solve itself once models scaled up, and he now publicly says he no longer believes that.
In the latest episode of the 20VC podcast, Microsoft CEO Satya Nadella was asked what belief had changed the most for him over the past 12 months. His answer is worth listening to again and again.
In a tweet, host Harry Stebbings recalled that when he started the podcast at 18 from his bedroom in London, he had nothing and didn't know anyone in the tech world. This interview was something he had dreamed of for a long time. Nadella took Microsoft's market cap from $300 billion to $3.9 trillion, and in the host's eyes, he is one of the greatest CEOs of this era.
Nadella's answer mentioned several names: Dario, Ilya, Sam, Greg, and Elon. He said these people are the true believers in scaling laws, and that he himself had believed in it in 2018 and 2019. Then he admitted: "I used to think the alignment problem would just emerge on its own. I no longer believe that. I think alignment may be a harder and different kind of problem."
Some background is needed here for readers who are not familiar with AI.
Scaling laws are the idea that model performance improves predictably as the scale of parameters, data, and compute increases. Much of the rapid development of the AI industry over the past few years has been built on this belief. Dario (Anthropic CEO), Ilya (OpenAI chief scientist), Sam (OpenAI CEO), Greg (OpenAI president), and Elon (Musk) are all early representative figures of this belief.
Alignment research is about how to make AI systems' goals conform to human intent. Nadella once believed that once models reached a certain size, alignment would solve itself, as a byproduct of capability improvements. Now he publicly says he no longer thinks so.
In the video, Nadella says this himself. He is wearing a black jacket, sitting in front of a wooden wall, with a potted plant in the background.
Some online commenters noticed that Nadella's list of scaling-law believers did not include anyone from Google. Another person commented that believing scaling laws would automatically solve alignment was a very common trap back then, and that they were glad he was now speaking candidly.
Nadella's shift is worth paying attention to because he is not just a researcher in an AI lab; he is the CEO of a large technology company with enormous compute and capital at its disposal. His judgment on alignment directly affects Microsoft's investment in AI safety. If even he believes alignment will not emerge naturally, the priority of this problem may be higher than many people imagine.
In the full interview, Nadella also discussed topics such as open vs. closed models, why every company needs to own its own intelligence, and vertical integration.
发布时间: 2026-10-28 09:42