Watch modeSubscribed

Decoder · 3 min · Oct 22, 2026

Gates sees three routes to AI control

Bill Gates separates near-term misuse from the longer control problem, then names external action monitors, different training methods and model interpretability. The framework gives researchers three concrete directions to pursue.

Bill Gates is co-founder and former CEO of Microsoft and chairman of the Gates Foundation.

Co-founded Microsoft in 1975; now chairs the Gates Foundation, which will close in 2045.

Nilay Patel, Host · Editor-in-chief, The Verge
Helped start The Verge in 2011 and has hosted Decoder since 2020.

About this segment

Bill Gates separates near-term misuse from the longer control problem, then names external action monitors, different training methods and model interpretability. The framework gives researchers three concrete directions to pursue.

From “Bill Gates says to prepare for 20 years of tumultuous AI disruption”, Oct 22, 2026. Watch the full episode

“One is to have an external set of things that get to look at all of the model's actions and subject them to a symbolic, clearly stated constraint.”

Bill Gatesat 15:06 · Decoder15:06

Takeaways

Use symbolic rules to check model actions and resource use
Change reinforcement learning to reduce unintended behavior
Inspect model internals to trace sources of errors

In this episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

13:58
We can talk about the three different ways to solve that, but that's not an imminent issue. When I look at some of the attempts to control the models, the ideas are somewhat simple. We shouldn't let them talk in neural ease. We should have interpretability. We should be able to see the chains of thought inside of the models as they carry out tasks. We should have other agents on top of the models watching their tasks before those actions are taken out.
14:24
Microsoft has said very clearly to me and to others, we should not put in the model constitutions the notion that they might be alive and we should treat them nicely. This is how Anthropic handles it. These are very tactical, right? It's don't tell the model it's alive and it might not believe that it's alive. You've obviously used a lot of this technology. I just come back to this. You're saying there's very simple things we can do to hold the models back from doing bad things in the hands of bad people.
14:52
What do you make of the nature of the models themselves? Because it feels like as they get more powerful, the notion that they could avoid our controls becomes all the more risky. Yes, we definitely should work on the control problem, but that's not the imminent threat. And there's three ways to solve the control problem, and it overlaps somewhat with bad actors using today's models and safeguards.
15:18
One is to have an external set of things that get to look at all of the model's actions and subject them to a symbolic, clearly stated constraint. Like, did I expect you to send mail to strangers? Did I expect you to go take over a different computer cluster? So you can surround the model with something that has the resource and security logic. Or you can write these models in a different way. RL, fundamentally, anything you leave out of the optimization function of RL can be abused where you have an implicit notion.
15:59
Well, of course, I didn't mean for you to do it this way. So you can change the template of the model, or you can have a sort of white box understanding of what's inside the model and see where you're getting those mistakes. And then, you know, you're in the build. It's not as foreign. Today, we grow these models and we just see how they behave. So the priority of work on maintaining control is huge, but that is not, you know, it's not inevitable that that cannot be solved, nor is it the imminent problem that come from things like kids' use of AI or job market or bio or cyber.