AI: Educating for the Inevitable | Prof Benjamin Rosman

Benjamin Rosman, guest · [09:21]So what everyone's trying to build are machines that are more and more general. And it's, it's really only in the last 10, 15 years that that's started to happen at any reasonable scale. And I think that's why ChatGPT and friends have become so popular now is because of the kind of broad diversity of things they're able to do. Everybody, no matter what industry they're in, has some use case that they're really fond of and that they like, like to brag about, like, I found this cool thing I can do with it, right? And it's because of that, that's the generality that we're starting to see the glimpses of coming through, that makes it exciting. Because the tools Haven't been built with each of those use cases in place. Okay. Now this is kind of part of the way we've been thinking about machine learning in recent years. This is part of the way they're trained, but we don't have to go kind of use case by use case figuring out how to solve it, which is how things were done 15 years ago. This is where the race is now. Now there's huge upshots or like the implications to this are, I think, really profound and we'll probably talk about a fair bit. So artificial general intelligence, there's many different definitions of what this is, but it roughly talks to like a point where machines are able to do anything that a human can do at like a human level of competence. Now there's different definitions. My personal feeling is you can define it however you want. They're all, you know, within a small amount of time. If it hits one, it will hit whatever other one you want. And artificial superintelligence is kind of this point, again, many different definitions where like an AI system is equivalent in intelligence, again, for You want to define it, we can talk about these things to maybe the whole planet of humans. These are, these are rough landmarks that people talk about in different ways, and then questions are all about when do we hit these things, will we hit these things, and what does it mean when we hit these things.

Benjamin Rosman, guest · [19:32]So humans have two spheres of labor that we've always worked. We do physical labor and we do kind of cognitive labor. And kind of in previous industrial revolutions, there, there's been a general shift of humanity from the physical to the cognitive. and this is kind of correlates with what we see as our most esteemed jobs and things is like the more heady you can be, the more you sit at your desk and hurt your back, like the better, like life is for you or something. That's the realm that we kind of aspire to. So there's, there's always been this, you know, these discussions about you, you want to upskill people. Implies that there's this, this hierarchy of things get better and you want to, more training means you better. In your higher up hierarchy. But also that means you're doing more and more intellectually, you know, intellectual type tasks that maybe fewer and fewer people higher up are able to actually do. Turns out all of these things are at risk of automation, right? So now if you're building systems that the exact same system can speak all the languages and can write poetry and can write, you know, maybe it's never, gone through a breakup, but if it's read every angsty breakup letter on the internet, it probably understands it better than you do. If it's got all this kind of experience, it's read every book on how to use Excel. It's, it can write the books on how to use Excel. It can process all the YouTube videos on this. It can also do all of this two hundred thousand times faster than any human, let alone consume every bit of information ever generated. What on earth is a human going to do in the face of that? Like, so, so I'm always left wondering, what are these imaginary jobs that people in their hubris like to think humans will magically always be Able to do that these machines that we can literally see trying to zoom past us on all these dimensions already as we sit here, already it can code better than I can. It can write better poetry than I can. Like, and this has happened over the space of, we take the whole tail into account, but really over the space of the last couple of years that this has happened, right? Again, the slowest change in this technology we'll ever see. It's just happened. Exactly. And, and again, I'm left wondering, like, where, where are these, you know, 170 million new jobs that are going to be created? Because to me, that doesn't make sense.

Benjamin Rosman, guest · [34:22]Well, my, my view here is that if we do the, put our head in the sand and say, no, things will all work out fine, maybe they will in the long term. But like, you know, what I see happening is more and more people are going to get either retrenched or, or not replaced, or there's just going to be a smaller demand for various skills, and this is going to be happening across kind of every sector in parallel, very, it's already happening very slowly at first, but then people say, oh, look, but some jobs were created, and they take that to mean, no, don't worry about it, and sounds a little bit like the, it can't be global warming because it snowed somewhere, you know, it's that kind of thing. We're looking at the macro picture here. And, and if you, you have the view that like, yeah, it's all going to work out, there'll just be jobs we can't even imagine yet, right? And again, maybe, like, we don't know what the future looks like, but, you know, if, if you just kind of accept that, then it is Possible, we end up at a point where, you know, in the space of a few months in some developed countries, you end up getting an uptick of unemployment of like five to eight percent or something in just a few months, and the social safety net can't handle that, civilization could collapse. It seems not a great outcome. I'm not a huge fan of this. on the other hand, maybe we could take other views and we could say, you know, something that I've been thinking about a bit, I'm not an economist, I'm not like a sociologist, but, you know, maybe this is a good thing. Maybe working for a living is immoral, and we shouldn't do it. But we've always had to do it because Humans are the ones who get the work done. But like, if you really think about it, you've got to do a thing which is bad for you, probably emotionally or psychologically or physically. But if you don't do that, then you and your family don't get to eat next month. Right? Which is kind of barbaric, if you put it in that terms, but that's what like all of humanity's been about. So maybe we shouldn't do that. Maybe we should instead figure out how to get all these crazy machines that are going to be so good at everything to do all the shitty work for us. Instead, we figure out what that society looks like. Now, there's a difference here between kind of accepting what might be happening and trying to plan ahead and say, this might require a paradigm shift, rather than saying, let's see what happens and there might be revolutions along the way. And then you can be quite specific about questions. You can say, okay, then how will people get, you know, whatever the, their currency is to buy food? There's a very well-defined problem you can think about. Or where will people find value in life if they're not working for a living? Well, they'll Is a very specific question we can have discussions about, but it's about taking some of these framings, challenging what the framing is, what might happen, and saying, okay, how would this affect some aspect? Now let's have a very specific discussion on this while other people think about other aspects. And that's what I would like to see more happening.

Jacqueline Aitchison, host · [41:57]There's something I did at the end of last year that became another little thought experiment. The grade eleven English class got a question saying, Given the global teacher crisis, Edu Inc is struggling to get an English teacher, so we've decided that you are going to have an AI-driven robot who is going to teach your English, and this robot will have the features of a woman. She's going to be dressed as a woman, features of a woman. Will be able to turn her head when she hears your voice, so it looks like she's engaging with you. She'll be able to move her arms, but she won't be able to walk around in the room, and she's not going to have changing facial expressions. The benefit of this is that you can also have this, this robot teach you any subject at any given time. So it's not just an English teacher. Now this, this is a resource on hand for whatever subject you would like. They could choose, they could write a letter to the head, or they could write a speech about it to the parents or, or whatever. And what was really interesting was 100% of those students didn't want it. We're such a techie school. We have kids who are so, tech is just an extension of their own existence. Every single one of them said no. For most of them, what was really fascinating to me was when they said no, it was a values-based no. Their answer was, this goes against the values of the school. I'm going to complain to the board that you have now taken a decision that is not a values-based decision, and it was all because of our relationship-based teaching. This robot can't tell if I have a headache. This robot can't tell if I have a tummy ache or if I'm just really tired because I had a bad night and I didn't sleep. The robot can't say to me, just put your head down on the desk for five minutes and take a breather. The robot can't read the room and say, right, clearly you just need five minutes outside, all of you, outside, come back in when you're ready. And it was all down to I have a relationship with a human being when it comes to something as important as my education. So they were actually prioritizing the relationship with the human being above I could get straight A's because I could have this robot on call tutoring me one on one day in and day out. That was prioritized lower than the relationship with the person who could look at them and say, you've got a headache, just take five minutes.

Benjamin Rosman, guest · [57:34]What this means for us, like in the education space, I think it's useful to think at like two different timescales, right? So the one is, what does this mean for me tomorrow? And then what does it mean for me, like, into the future, this murky, uncertain future? Now, for the tomorrow, if we were these things that are able to consume, like, all of human knowledge, and are able to generate usually useful and coherent things, in this wide variety of tasks, I often liken using an AI tool to having, like, a super enthusiastic intern who's done, like, two PhDs, but has zero real world experience. And, like, if you have that, right, you're not gonna leave that with your children unsupervised, but, like, if you have the ability to bring a bunch of these into your classroom with you, and you're not doing it, you're probably doing a disservice. That means you've gotta learn, like, how do I get the most value out of this, but this should be, like, the most exciting thing you could possibly do in an educational context. What does this mean longer term? You also touched on that question there of, like, what is education for? The one aspect of it, I think, is about teaching you to, like, Like, find yourself, find your place in society, your relationship to others, and these kinds of things. But then also there's a piece about learning how to go out into the world and, like, get your income and, and that kind of thing. And maybe that bit is less useful. Maybe that's something that we should be pivoting away from. So I spent many years people asking me what should they teach their kids, of saying, like, well, they should learn to code. And then I recently saw a quote somewhere online that said, telling someone that they should learn to code is now as useful life advice as you should get a face tattoo. But I do fundamentally disagree with that. I think we're at the stage now where very soon AI will be a better coder than any human ever will be. However, something like learning to code, the reason you want to do that is to be able to learn to think through problems, kind of break it down, like problem solving, right? So I think that there are these fundamental skills that typically are just useful things to have going through the world. What I've more recently been saying to people is I think the most important thing, given these, this uncertainty, these crises around jobs, is to teach kids to find value in life beyond what they do for a living. And that to me is, is critical. It's that combined with curiosity, with some resilience, like these, these things that we, we know are valuable, but tend to get, you know, there, there's not a checkbox for them. And so they tend to fall off the, these are the things you need coming out of school kind of list in most places.

Benjamin Rosman, guest · [66:08]to be generative, it can create its own data. So that, that is a problem at the moment because that if, if it's only creating its own data, it tends to get worse and worse over time. The big bottleneck many of us talk about is more on like the computational power. And these are like what they call compute centers rather than data centers. This is basically part of the arms races around this, because what they rely on is like GPUs, which is graphics processing units, all the stuff that you want if you want to be gaming on your home computer, you just want like gazillions of these things to be running these AIs. And the power consumption's crazy, like the, the predictions of like, you know, in the next however many years, this will be ten or twenty percent of like the US's power consumption will be these things, which is why you're seeing a lot of the big tech companies that are in this race are now building, like their new compute centers have power plants attached to them, including now all the discussions around nuclear power. So you're gonna have these compute centers in places like Texas that are being built that have nuclear power plants attached to them to keep them going. The flip side of that is in the research space, there's a huge drive to build more and more efficient systems. And so like a significant percentage of what, what is happening in AI research is how do we do this more efficiently. Our Inspiration is the brain, but there's a very different mechanism from an energy usage point of view. Like, you know, the brain uses something like 20 watts of energy, whereas these things are like, you know, nuclear power plants. There's a tension there which drives a lot of research questions, which I think are, are pretty interesting. I mean, the, the other funny thing is that, you know, AI is often touted as will be a solution to global warming. Once we figure out how to stop them being a huge contributor to global warming.