Microsoft CEO's reminder: Models will get stronger, but won't automatically become trustworthy

This episode of 20VC features Microsoft CEO Satya Nadella. Harry Stebbings started the podcast at age 18 in his London bedroom; this time he sat down opposite Satya. Microsoft’s market cap went from $300 billion to $3.9 trillion, a figure repeatedly mentioned in the tweet. Original tweet link: https://x.com/HarryStebbings/status/2114791168932229414

In the video, Satya puts it directly. Models are getting stronger. He had originally assumed alignment would emerge naturally with scaling, but it seems it hasn't. If we want to continue using these complex models, the best stance is to put them inside a trust architecture. If agentic capabilities are stronger but not necessarily better aligned, we need a trustworthy system outside the model providers.

The statement isn't complicated; the hard part is implementation. A trust system cannot grow on the model itself. One user wrote in a reply that a trust layer cannot come from something you don't fully trust; it can only be outside the model. Another asked: who audits the auditors—another model? Someone else joked that the trust system is still a person drinking coffee and reading logs at 7 a.m.

The episode agenda also includes open models vs. closed models, whether to trust frontier model providers, why companies need to own their own intelligence, and vertical integration from chips to models. Open vs. closed is often reduced to a contest over capability. The cost of trust does not disappear; it just shifts to someone else. Open source gives you inspectability; closed source gives you convenience and scale. Audits, permissions, rollbacks, and sign-offs ultimately need someone to take responsibility. At the operational level, a company owning its intelligence means holding data, permissions, audits, rollbacks, and sign-offs in its own hands. It is unrealistic for every company to train foundation models.

A user, Ryu1TS, offered a more concrete approach: checks the model cannot change, files that cannot be opened at the OS level, hooks that reject bad commits, and human sign-off before sending. The model can reason its way around things written in instructions. This detail is more useful than the phrase 'trust architecture.' At a bank or insurance company, this system would fall to the risk committee that approves new vendors. The process is not fast.

The model capability curve is shiny. The trust system curve is unglamorous. Inference cards are expensive. What is more expensive is the person willing to look at logs at 7 a.m.

Episode links: Spotify https://open.spotify.com/episode/3OfkdIG5IkU3kGOJuMUmAF , YouTube https://youtu.be/24XUcWhqY_4 , Apple Podcasts https://podcasts.apple.com/us/podcast/20vc-satya-nadella-on-open-vs-closed-models-who-wins/id958230465?i=1000798128221 . Quoted tweet: https://x.com/HarryStebbings/status/2114734401674228113