Show notes
In this episode of Faster, Please! — The Podcast, I talk with economist Robin Hanson about a) how much technological change our society will undergo in the foreseeable future, b) what form we want that change to take, and c) how much we can ever reasonably predict.Hanson is an associate professor of economics at George Mason University. He was formerly a research associate at the Future of Humanity Institute at Oxford, and is the author of the Overcoming Bias Substack. In addition, he is the author of the 2017 book, The Elephant in the Brain: Hidden Motives in Everyday Life, as well as the 2016 book, The Age of Em: Work, Love, and Life When Robots Rule the Earth.In This Episode* Innovation is clumpy (1:21)* A history of AI advancement (3:25)* The tendency to control new tech (9:28)* The fallibility of forecasts (11:52)* The risks of fertility-rate decline (14:54)* Window of opportunity for space (18:49)* Public prediction markets (21:22)* A culture of calculated risk (23:39)Below is a lightly edited transcript of our conversationInnovation is Clumpy (1:21)Do you think that the tech advances of recent years — obviously in AI, and what we're seeing with reusable rockets, or CRISPR, or different energy advances, fusion, perhaps, even Ozempic — do you think that the collective cluster of these technologies has put humanity on a different path than perhaps it was on 10 years ago?. . . most people don't notice just how much stuff is changing behind the scenes in order for the economy to double every 15 or 20 years.That’s a pretty big standard. As you know, the world has been growing exponentially for a very long time, and new technologies have been appearing for a very long time, and the economy doubles roughly every 15 or 20 years, and that can't happen without a whole lot of technological change, so most people don't notice just how much stuff is changing behind the scenes in order for the economy to double every 15 or 20 years. So to say that we're going more than that is really a high standard here. I don't think it meets that standard. Maybe the standard it meets is to say people were worried about maybe a stagnation or slowdown a decade or two ago, and I think this might weaken your concerns about that. I think you might say, well, we're still on target.Innovation's clumpy. It doesn't just out an entirely smooth . . . There are some lumpy ones once in a while, lumpier innovations than usual, and those boost higher than expected, sometimes lower than expected sometimes, and maybe in the last ten years we've had a higher-than-expected clump. The main thing that does is make you not doubt as much as you did when you had the lower-than-expected clump in the previous 10 years or 20 years because people had seen this long-term history and they thought, “Lately we're not seeing so much. I wonder if this is done. I wonder if we're running out.” I think the last 10 years tells you: well, no, we're kind of still on target. We're still having big important advances, as we have for two centuries.A history of AI advancement (3:25)People who are especially enthusiastic about the recent advances with AI, would you tell them their baseline should probably be informed by economic history rather than science fiction?[Y]es, if you're young, and you haven't seen the world for decades, you might well believe that we are almost there, we're just about to automate everything — but we're not.By technical history! We have 70-odd years of history of AI. I was an AI researcher full-time from ’84 to ’93. If you look at the long sweep of AI history, we've had some pretty big advances. We couldn't be where we are now without a lot of pretty big advances all along the way. You just think about the very first digital computer in 1950 or something and all the things we've seen, we have made large advances — and they haven't been completely smooth, they've come in a bit of clumps.I was enticed into the field in 1984 because of a recent set of clumps then, and for a century, roughly every 30 years, we've had a burst of concern about automation and AI, and we've had big concern in the sense people said, “Are we almost there? Are we about to have pretty much all jobs automated?” They said that in the 1930s, they said it in the 1960s — there was a presidential commission in the 1960s: “What if all the jobs get automated?” I jumped in in the late ’80s when there was a big burst there, and I as a young graduate student said, “Gee, if I don't get in now, it'll all be over soon,” because I heard, “All the jobs are going to be automated soon!”And now, in the last decade or so, we've had another big burst, and I think people who haven't seen that history, it feels to them like it felt to me in 1984: “Wow, unprecedented advances! Everybody's really excited! Maybe we're almost there. Maybe if I jump in now, I'll be part of the big push over the line to just automate everything.” That was exciting, it was tempting, I was naïve, and I was sucked in, and we're now in another era like that. Yes, if you're young, and you haven't seen the world for decades, you might well believe that we are almost there, we're just about to automate everything — but we're not.I like that you mentioned the automation scare of the ’60s. Just going back and looking at that, it really surprised me how prevalent and widespread and how serious people took that. I mean, you can find speeches by Martin Luther King talking about how our society is going to deal with the computerization of everything. So it does seem to be a recurrent fear. What would you need to see to think it is different this time?The obvious relevant parameter to be tracking is the percentage of world income that goes to automation, and that has been creeping up over the decades, but it's still less than five percent.What is that statistic?If you look at the percentage of the economy that goes to computer hardware and software, or other mechanisms of automation, you're still looking at less than five percent of the world economy. So it's been creeping up, maybe decades ago it was three percent, even one percent in 1960, but it's creeping up slowly, and obviously, when that gets to be 80 percent, game over, the economy has been replaced — but that number is creeping up slowly, and you can track it, so when you start seeing that number going up much faster or becoming a large number, then that's the time to say, “Okay, looks like we're close. Maybe automation will, in fact, take over most jobs, when it's getting most of world income.”If you're looking at economic statistics, and you're looking at different forecasts, whether by the Fed or CBO or Wall Street banks and the forecasts are, “Well, we expect, maybe because of AI, productivity growth to be 0.4 percentage points higher over this kind of time. . .” Those kinds of numbers where we're talking about a tenth of a point here, that's not the kind of singularity-emergent world that some people think or hope or expect that we're on.Absolutely. If you've got young enthusiastic tech people, et cetera — and they're exaggerating. The AI companies, even they're trying to push as big a dramatic images they can. And then all the stodgy conservative old folks, they're afraid of seeming behind the times, and not up with things, and not getting it — that was the big phrase in the Internet Boom: Who “gets it” that this is a new thing?I'm proud to be a human, to have been part of the civilization to have done this . . . but we've seen that for 70 years: new technologies, we get excited, we try them out, we try to apply them, and that's part of what progress is.Now it would be #teamgetsit.Exactly, something like that. They're trying to lean into it, they're trying to give it the best spin they can, but they have some self-respect, so they're going to give you, “Wow 0.4 percent!” They'll say, “That's huge! Wow, this is a really big thing, everybody should be into this!” But they can't go above 0.4 percent because they've got some common sense here. But we've even seen management consulting firms over the last decade or so make predictions that 10 years in the future, half all jobs would be automated. So we've seen this long history of these really crazy extreme predictions into a decade, and none of those remotely happened, of course. But people do want to be in with the latest thing, and this is obviously the latest round of technology, it's impressive. I'm proud to be a human, to have been part of the civilization to have done this, and I’d like to try them out, and see what I can do with them, and think of where they could go. That's all exciting and fun, but we've seen that for 70 years: new technologies, we get excited, we try them out, we try to apply them, and that's part of what progress is. The tendency to control new tech (9:28)Not to talk just about AI, but do you think AI is important enough that policymakers need to somehow guide the technology to a certain outcome? Daron Acemoglu, one of the Nobel Prize winners, has for quite some time, and certainly recently, said that this technology needs to be guided by policymakers so that it helps people, it helps workers, it creates new tasks, it creates new things for them to do, not automate away their jobs or automate a bunch of tasks.Do you think that there's something special about this technology that we need to guide it to some sort of outcome?I think those sort of people would say that about any new technology that seemed like it was going to be important. They are not actually distinguishing AI from other technologies. This is just what they say about everything.It could be “technology X,” we must guide it to the outcome that I have already determined.As long as you've said, “X is new, X is exciting, a lot of things seem to depend on X,” then their answer would be, “We need to guide it.” It wouldn't really matter what the details of X were. That's just how they think about society and technology. I don't see anything distinctive about this, per se, in that sense, other than the fact that — look, in the long run, it's huge.Space, in the long run, is huge, because obviously in the long run almost everything will be in space, so clearly, eventually, space will be the vast majority of everything. That doesn't mean we need to guide space now or to do anything different about it, per se. At the moment, space is pretty small, and it's pretty pedestrian, but it's exciting, and the same for AI. At the moment, AI is pretty small, minor, AI is not remotely threatening to cause harm in our world today. If you look about harmful technologies, this is way down the scale. Demonstrated harms of AI in the last 10 years are minuscule compared to things like construction equipment, or drugs, or even television, really. This is small.Ladders for climbing up on your roof to clean out the gutters, that's a very dangerous technology.Yeah, somebody should be looking into that. We should be guiding the ladder industry to make sure they don't cause harm in the world.The fallibility of forecasts (11:52)I'm not sure how much confidence we should ever have on long-term economic forecasts, but have you seen any reason to think that they might be less reliable than they always have been? That we might be approaching some sort of change? That those 50-year forecasts of entitlement spending might be all wrong because the economy's going to be growing so much faster, or the longevity is going to be increasing so much faster?Previously, the world had been doubling roughly every thousand years, and that had been going on for maybe 10,000 years, and then, within the space of a century, we switched to doubling roughly every 15 or 20 years. That's a factor of 60 increase in the growth rate, and it happened after a previous transition from forging to farming, roughly 10 doublings before.It was just a little over two centuries ago when the world saw this enormous revolution. Previously, the world had been doubling roughly every thousand years, and that had been going on for maybe 10,000 years, and then, within the space of a century, we switched to doubling roughly every 15 or 20 years. That's a factor of 60 increase in the growth rate, and it happened after a previous transition from forging to farming, roughly 10 doublings before.So you might say we can't trust these trends to continue maybe more than 10 doublings, and then who knows what might happen? You could just say — that's 200 years, say, if you double every 20 years — we just can't trust these forecasts more than 200 years out. Look at what's happened in the past after that many doublings, big changes happened, and you might say, therefore, expect, on that sort of timescale, something else big to happen. That's not crazy to say. That's not very specific.And then if you say, well, what is the thing people most often speculate could be the cause of a big change? They do say AI, and then we actually have a concrete reason to think AI would change the growth rate of the economy: That is the fact that, at the moment, we make most stuff in factories, and factories typically push out from the factory as much value as the factory itself embodies, in economic terms, in a few months.If you could have factories make factories, the economy could double every few months. The reason we can't now is we have humans in the factories, and factories don't double them. But if you could make AIs in factories, and the AIs made factories, that made more AIs, that could double every few months. So the world economy could plausibly double every few months when AI has dominated the economy.That's of the magnitude doubling every few months versus doubling every 20 years. That's a magnitude similar to the magnitude we saw before from farming to industry, and so that fits together as saying, sometime in the next few centuries, expect a transition that might increase the growth rate of the economy by a factor of 100. Now that's an abstract thing in the long frame, it's not in the next 10 years, or 20 years, or something. It's saying that economic modes only last so long, something should come up eventually, and this is our best guess of a thing that could come up, so it's not crazy.The risks of fertility-rate decline (14:54)Are you a fertility-rate worrier?If the population falls, the best models say innovation rates would fall even faster.I am, and in fact, I think we have a limited deadline to develop human-level AI, before which we won't for a long pause, because falling fertility really threatens innovation rates. This is something we economists understand that I think most other people don't: You might've thought that a falling population could be easily compensated by a growing economy and that we would still have rapid and fast innovation because we would just have a bigger economy with a lower population, but apparently that's not true.If the population falls, the best models say innovation rates would fall even faster. So say the population is roughly predicted to peak in three decades and then start to fall, and if it's falls, it would fall roughly a factor of two every generation or two, depending on which populations dominate, and then if it fell by a factor of 10, the innovation rate would fall by more than a factor of 10, and that means just a slower rate of new technologies, and, of course, also a reduction in the scale of the world economy.And I think that plausibly also has the side…
Full show notes at the publisher 
Our TOPPODCAST Picks 
Stay Connected
