Are people at Microsoft already using this new version of Copilot?
Yeah, everyone uses it. We just went to general availability company-wide a few days ago. So everyone is using it, giving tons and tons of feedback. Our comment channels are absolutely criticized. It’s a lot of fun.
People will remember Clippy, Microsoft’s latest AI assistant for Windows. Do people there see any parallels?
Ha, well, I saw Bill Gates the other day, and he said to me, do you realize you misnamed this whole AI thing? It should be called Clippy. I was like, man!
But I mean, it shows how amazing people like Bill are. People who see, not just, you know, two years ahead, but 20 years.
Are the new features a step toward so-called AI agents, which perform useful tasks on a computer?
Yeah, absolutely. The first step is for AI to process the same information as you: see what you see, hear what you hear, consume the text you consume. The second phase is [AI having] a persistent long-term memory that creates shared understanding over time. And the third step is for AI to interact with third parties by sending instructions and taking actions: buying things, booking things, planning a schedule. And we have both of these features in an experimental R&D mode that we’re working on.
Wait, you have an AI agent for Windows that can go buy things for you?
It’s far from being the case, but yes, we’ve come full circle, we’ve made deals. The problem with this technology is that you can get it to work 50-60% of the time, but to achieve 90% reliability it takes a lot of effort. I’ve seen some amazing demos where it can boot independently and make a purchase etc. But I also saw moments in car accidents where he didn’t know what he was doing.
Tell me more about a car accident. Did he buy a Lamborghini with Bill’s credit card?
If he used Bill’s credit card, I think it would be pretty funny. But no, like I said, we’re still figuring it out step by step. He is still deep in the laboratory doldrums. There is a long way to go in this area, but you can count it in quarters, not years, I would say.
What will be the biggest challenge for you in realizing the kind of AI future you describe?
The key here is to figure out how to create reliable technology, because it will feel like a very intimate and personal experience. We must of course succeed on the security side, and of course on the confidentiality side. But I think the real thing is to try to design the conversation in a way that the agent is able to articulate boundaries, so that they can say this is not something I’m ready for to engage me.
If we can solve that problem, that’s the basis of a trust experiment, and then I think we can really get into the complicated side of things, which is how to let him buy things for you on your behalf, or negotiate on your behalf, or enter into a contract on your behalf, or schedule a schedule for you that includes three or four different stops over the course of a Saturday afternoon. And you say to yourself, I trust you, Copilot, you understand, right? That’s really what we’re working towards.
