AI will never be safe again
On the 20VC podcast the Microsoft CEO admitted that alignment with human goals does not arrive with model power.
Microsoft CEO Satya Nadella says he no longer believes AI will learn to act in people's interests on its own as it gets more powerful. He said it on the 20VC podcast, when host Harry Stebbings asked him which view he had changed over the past year.
Nadella used to think alignment, meaning AI matching human values and goals, would simply arrive with scale. Scaling has gone as expected and models keep getting stronger, yet alignment is still unsolved. He now says it may be a different and even harder problem than he assumed.
This is not the first signal. On October 10 Nadella wrote on X that companies should treat AI models as a potential insider threat and assume a model may already be compromised. His answer is an engineering one: wrap non-deterministic models in strict deterministic architecture, human oversight and clear procedures.
Hearing this from Altman or Amodei is routine. Hearing it from the man who sells AI to every office on the planet makes it an industry position.