There's this super talented young Korean piano player, Yunchan Lim, right now who's my favorite pianist in the world. And the thing that makes him so special is not that he is technically virtuosic, it's that he chooses to emphasize things in the score that I've never heard anyone emphasize before. I don't even know how he makes those decisions. He may not know how he makes those decisions, but the result that he gets is really extraordinary.
Yeah, so I think about this all the time because I spend almost all of my time when I'm not working thinking about craft and art. We're sitting in a pottery studio and right next door is a wood shop. That's one of the earliest things that I got my hands busy doing when I was young. In woodworking, there's this constant debate about tools and the legitimacy of the crafts person and the craft depending on which tools you choose to use. The question is, what are the tools allowing me to do? So there might be things that are artisanal through and through but that I would not be able to imagine without AI, that I would not be able to actually do without AI.
As you know, one of the things I find fascinating, intriguing, also slightly confusing about AI is that I think it's philosophically new. It sort of partly, not fully, but partly exceeds the concepts we already have available to think about things in the world. If okay with you, I would like to tell a little story of one of our earlier conversations with Reed on Sunday mornings. I was like, 'AI is so much more than a machine,' and you politely waited till my speech wreck was over and was like, 'Actually, it's a machine.' So me being hyper sensitive, next conversation we had another Sunday morning, I was like, 'AI is sort of a machine,' and you interrupted me and said, 'Actually, it's much more than a machine.' So there is a tension there that I found fascinating and that in retrospect made me think, yeah, the vocabulary we have is really not good enough.
Yeah, look, I think the vocabulary is super problematic. It just even starts with the term artificial intelligence and intelligence. It's not as if we understand our own intelligence particularly well. So you are inventing a new thing nominally inspired by or predicated on this existing thing that is ill-understood. So I think whether or not AI is a machine depends on what you think the definition of AI is, and it probably depends a lot on your predisposition and your point of view. It's very hard for me to look at a piece of software and at the end of the day, AI is just a bunch of code running on computer chips and networks and data storage. It looks kind of like just a program on a von Neumann machine the same way that we've had for a hundred years now. It's hard to say that it's not a machine. And yet, I'm persuaded by not just how you've been thinking about things but by how people are actually using it and how they perceive it that maybe it is something more than that.
Yeah, I find that so. I mean, this goes straight to the things that I spent literally every day thinking about. I think that AI systems can learn from data and can generalize to some degree, can learn concepts that allows them to navigate situations they have not really seen before. That is something that up until that moment, which for me was AlexNet, only organisms knew how to do, or DQN and Atari games, some form of agency or agentive skill in the sense of doing something that we did not explicitly program it to do, but it's not random, so there's some systematicity there. The fact that there was a long tradition of structuralist linguistics that say language is just a system and you can learn it, but very few people really bought into this because language is deeply subjective. So that AI systems can do all of these things makes them exceed the machine notion.
Yeah, I mean, I tend to think of AI and AI software more in terms of tools. I'm sure that this machine/not-machine dichotomy is important on some level, but the interesting framing for me for a very long while now has just been, okay, it's a tool, it's an instrument that we use to do something. And I do agree that it's a very interesting tool. We've got this pottery sitting here on the table, and there are things that potters have done for a very long time with the tools of their craft to try to get a result that is something that they didn't explicitly express. You glaze this thing in a particular way with a bunch of materials and you put it in a wood-fired kiln and it's this huge dynamic process, and you get something out of it that is very difficult to predict. Part of that is because the environment, the tool, is itself inherently random, and part of it is because the processes involved are very complex, beyond an ordinary understanding of what's going on. These AI tools are kind of interesting because you can get them to do things that are interesting, that are not exactly the thing that you have expressed, or they can get to a result in a way that is not the way that you expected or intended.
It's a great analogy. When any artisan who uses tools, in this case pottery, you have this kiln, lots of stuff is going on inside. You cannot possibly control it. You control for a lot of things, but you actually don't want to control that part because that's the part where the surprise can happen. And yet it's not just 'I control everything and I click the random button.' There is some deep intuition that is difficult to articulate where you play with chance. There's this beautiful phrase by Max Weber: 'Ideas come when it pleases them, not when you need them.' So you have to bring yourself in a certain condition. You have to become a virtuoso of playing with this irreducible, uncontrollable something. And AI does some of that, which is fascinating for me because you put AI in a tradition of artists and artisans. For a long time, these two things were not different. At that point, I have two diverging thoughts. One is that is so cool and is sort of different from how most companies or labs that I know build AI, namely as an automation tool, as an opportunity to push automation a little further. So that AI does stuff on its own. But what you just described, AI is a tool. So whatever it does on its own is not very interesting. What's interesting is what you can do with it.
Yeah, and look, that's certainly my point of view. I'm one of those people who could not be less interested in what AI does or is capable of doing outside of the context of humans who are using the tool to accomplish a goal and the beneficiaries of whatever that product was. Even when we're going to overuse these metaphors because we've got these objects sitting here on the table in front of us, the interesting thing about these dynamic processes, the random and dynamic way that people use tools to make objects like this, is that there's a whole story about this notion of being able to wield the whole process like a virtuoso and to get into the ballpark of what might be an interesting set of results, but you still are relying on an audience to actually tell you which of the things here are valuable and which aren't. I was actually watching a YouTube video last night about a pottery in Japan that makes, as many potteries do, a very narrow range of a particular style of teapot. They were saying that for this one particular style, you're so dependent on the dynamics of what happens in the kiln that only 30% of the things that go into the kiln actually become sellable products. The rest of them are not good enough or didn't achieve an interesting result. You have a 70% experimentation space. It's arbitrary, because if you were not a pottery aficionado, if you weren't entrenched in that deep human story space that is both people who make the things and people who appreciate them, I don't think you'd be able to tell the difference between a good and a bad pot at the same level of distinction they could.
I think that's true for most arts. There is an imprecision, so to speak, in what is the right outcome, and that imprecision is really difficult to understand. The space of possibility that is handed to the artisan by what is a good product is very difficult for us to assess. Why is this better than that? But for them, that's the world, that's the space where they can experiment, and they may actually resist the process of understanding or over-constraining or imposing an understanding on top of what they're doing. So let me get back to AI from here for a second. There are two layers I find fascinating. At one point I developed the habit to say 'there are two things' or 'there are three things' and I just list them. One is I have this ambition to list things or name things that AI systems can do that are different from what a classical machine could do. Mostly because I want engineers to feel liberated to not build AI as a machine only, but to also optimize it for these things that one can do. There's a story by Karl Marx about life in a workshop of an artisan before the steam engine was invented. It was complicated and hard, but the artisan built the product from beginning to end and was a virtuoso. How good the product is depends on how developed the virtuosity is with the arts and tools. The reason why Marx didn't like the factories is that he literally describes the machine as virtuoso that humans have to serve.
He was also discounting that the factory itself was also a really interesting product. There was a huge amount of virtuosity involved in constructing the factory itself. I think Marx just didn't really appreciate that aspect of what was going on, which is fine.
Yeah, I agree with that. I think he was mostly focusing on labor, on the living condition of the worker, which is a very good thing to focus on. Also very complicated because the prior dignity of those workers was that they were serfs, and it took a long time to develop that space of dignity. The reason why I mentioned that is there is something beautiful about AI being one of the possibilities of bringing virtuosity to lots of people back. How you use it, what you can do with it, and how you can develop it.
Yeah, and that's certainly how I hope it will get used. I think one of the interesting things that's happened as we sit here having this conversation just over the past 12 months is that the capability of the current generation of frontier models are honestly best gotten at by people who have expertise. I'm not talking about AI experts or computer scientists, but if you're a domain expert in something, if you're a philosopher, a chemist, a biologist, an artist, and you look at this and say, 'My goodness, what a powerful and fascinating tool,' you're able to squeeze things out of it that people without that particular expertise or point of view are able to do with it at the moment.
There's definitely... my domain expertise is history of thought. The number of people who speak these various languages and can trace words and how they develop is very small, and oh my god, is that a difference with these AI systems. It's magnificent. One of the things that I like about your push of getting people to sort of reframe things, like trying to understand a tool in terms of 'is it going to make the existing set of things that we're doing more efficient and increase the productivity of people who are doing those things?' I don't think is the most interesting question to ask. The interesting questions are what brand new things that were impossible before get enabled by this new thing.
Yeah, I mean, that speaks my heart. I'm not at all opposed to increasing efficiency; that's a good thing in itself. But I think it's even a necessary thing in some places where we have very zero-sum constraints on problems that are very important to society. There's just a bunch of things going on right now where we don't have enough of a thing to go around. It's ridiculous wishful thinking to think that inside of the zero-sumness of all of those constraints, we're going to be able to shuffle things around in a way where everybody's going to be happy and have enough. The thing that you actually have to have is some kind of massive change in the shape of productivity in order to create abundance so everybody can have more.
Yeah, no, there's one of these bifurcation points. But I'm not a good conversation partner here.
No, this is good. This means the conversation is flowing. One of the things I find fascinating about AI is that it can help me think things that I don't quite know how to think alone. It can help me think things on a level of subtlety and scale that is unusual, that is very difficult to achieve alone. A pen can help me think things clearer, like equations I need to solve, but that's not exactly what I mean. By thinking things through, I create a space of possibility, I order things, I form a judgment. The more of this ordering and thinking things through I can do, the more orientation I gain, the more capable in navigating I am, the more freedom I have to make decisions.
I think also the more paths you can explore. All of us feel constrained in a whole bunch of ways. It's like, 'I don't know this' or 'I'm not good at this thing.' Being able to just relax those constraints, even if not completely, just a little bit, gets you down paths that sometimes are super interesting that you would not even bother exploring before.
I totally agree. I think the AI systems also got so... one fascination I have is there is for me an intuitive connection between set theory and twelve-tone music. This is not something that most people even care about, but I was always having a very hard time articulating what that is. So I asked Claude, 'Is there a connection?' and the answer was straightforward: no. So I explained why I think that there is a connection, operating on the level of abstractions, on a level of highly determined abstraction that creates music. Claude allowed me to then find connections and articulate connections and fill a gap in a way where there was a seamless continuity between the two. I would not have known how to do that myself. It's just one example, which I think is super interesting. The other association you once said to me, I don't remember the exact number, but in the 1920s there were machines in the world that could do the work of let's say 10 billion actual workers if you would not have machines. That was a productivity increase that was overall very good and important. AI can be a further step in that direction.
Absolutely, I completely agree with that. Going back to the tool, one of the things that is fascinating about AI for me is that if I have a classical tool, let me use a violin, whether it sounds good or not depends on the virtuosity of the human who uses it. It's kind of reducible in its value to how a human knows how to do it. Whereas AI appears to be slightly different from a classical tool in the sense that it can do stuff on its own to some degree, that it has some level of agentic capabilities or agency.
Yes, probably. I mean, you picked an especially interesting example in the violin. I do think a violin is pretty binary, but I'm more familiar with piano. With piano, it's less binary. You can, without an enormous amount of effort, and this used to be a pretty common thing because it was the way that you actually had music in your life, relatively normal people, folks you wouldn't call virtuosos, could learn how to play reasonable amounts of music. I grew up in rural central Virginia and I went to a small Southern Baptist church every Sunday when I was a kid. We had a piano and an organ player in the church. They didn't go to Juilliard, they didn't go to college to study music, they just sort of learned how to play and they were good enough to produce the effect that they needed to produce, which is to engage this big group of people into a shared experience around music. I rather admire those players. There's a continuum there. You go all the way to the very extreme ends and you have people who both have spent exorbitant amounts of time training and they also have some talent. Some of it is purely neuromuscular; they just have a nervous system that functions differently than a normal person. There's this super talented young Korean piano player, Yunchan Lim, right now who's my favorite pianist in the world. The thing that makes him so special is not that he is technically virtuosic. It's that he chooses to emphasize things in the score that I've never heard anyone emphasize before. I don't even know how he makes those decisions. He may not know how he makes those decisions, but the result that he gets is really extraordinary.
This is a bit of a jump, but let me go from there to put AI in that context. What is the meaning of AI in such a context?
So I think about this all the time because I spend almost all of my time when I'm not working thinking about craft and art.
I just want to pause here because this is really important. You're Kevin Scott, the CTO of Microsoft, EVP of AI. So that signals on the surface of things cutting edge frontier AI models, as techy as possible, software. But here we are in your studio. You're a craftsman, you're a builder, you're a maker with your hands. It's very important to you that this continues into the idea that there was a time when art and artistry were not differentiated, they were sort of the same way. And this is how you approach AI and this is what you bring AI into. This is deeply fascinating. I think that we need to practice AI as if it were a craft and we need to build AI that enables people to do things they didn't know how to do otherwise.
