Making Sense - Sam Harris - April 22, 2022


#280 — The Future of Artificial Intelligence


Episode Stats

Length

37 minutes

Words per Minute

160.62764

Word Count

6,091

Sentence Count

319

Hate Speech Sentences

3


Summary

Eric Schmidt is a technologist, entrepreneur, and philanthropist. He joined Google in 2001 and served as its CEO and Chairman from 2001 to 2011, and as Executive Chairman and Technical Advisor thereafter. In 2017, he co-founded Schmidt Futures, a philanthropic initiative that bets early on exceptional people who are helping to make the world better. And most recently, he is the co-author of a new book, The Age of AI and Our Human Future. In this conversation, we cover how AI is affecting the foundations of our knowledge, and how it raises questions of existential risk. We talk about the good and the bad of AI, both narrow and ultimately AGI, artificial general intelligence. We also talk about cyber-war, autonomous weapons, and other concerns about the risk posed by autonomous weapons and how our thinking about containing the risk here by analogy to the proliferation of nuclear weapons probably needs to be revised. In an important conversation, which I hope you will find useful, I bring you Eric Schmidt. I am here with Eric Schmidt, and I am so grateful to be with him. I think we have a hard-out at an hour here. -Sam Harris If you are not yet a subscriber, you re not currently on our subscriber feed and would like to become one, please consider becoming a supporter of what we re doing here, then you re gonna want to become a subscriber. We don t run ads on the podcast, and therefore you won t miss out on the benefits that come with the podcast! Thanks to our sponsorships, we don t get any ad-free version of The Making Sense Podcast, which is made possible entirely through the support made possible by the support of our listeners, you get 10% off the making sense Podcasts, and you get 20% off of the podcast only by becoming a member of the Making Sense Community. You get a better podcast listening to the podcast and a better chance of getting 10% discount, and a discount on future episodes, and 5% of the Podcasts only, too good at making sense? Subscribe to Making Sense. Thanks for listening and Subscribe to the Podcast! - Sam Harris Thank you, Sam Harris and I hope that you enjoy what we're doing this podcast makes you think about what we do here? - Your feedback helps us make sense of the things we re making sense, right there, and we re listening to it, too, and it helps us all make sense.


Transcript

00:00:00.000 Welcome to the Making Sense Podcast. This is Sam Harris. Just a note to say that if
00:00:12.180 you're hearing this, you are not currently on our subscriber feed and will only be hearing
00:00:16.320 the first part of this conversation. In order to access full episodes of the Making Sense
00:00:20.820 Podcast, you'll need to subscribe at samharris.org. There you'll find our private RSS feed to
00:00:26.420 add to your favorite podcatcher, along with other subscriber-only content. We don't run
00:00:31.100 ads on the podcast, and therefore it's made possible entirely through the support of our
00:00:35.000 subscribers. So if you enjoy what we're doing here, please consider becoming one.
00:00:46.420 Okay, jumping right into it today. Today I'm speaking with Eric Schmidt. Eric is a technologist,
00:00:53.460 entrepreneur, and philanthropist. He joined Google in 2001 and served as its CEO and chairman from
00:01:02.180 2001 to 2011, and as executive chairman and technical advisor thereafter. In 2017, he co-founded
00:01:10.940 Schmidt Futures, a philanthropic initiative that bets early on exceptional people who are helping to make
00:01:16.540 the world better. He is the host of Reimagine with Eric Schmidt, his own podcast. And most recently,
00:01:23.740 he is the co-author of a new book, The Age of AI and Our Human Future. And that is the topic of today's
00:01:30.780 conversation. We cover how AI is affecting the foundations of our knowledge and how it raises
00:01:37.680 questions of existential risk. So we talk about the good and the bad of AI, both narrow AI and
00:01:45.360 ultimately AGI, artificial general intelligence. We discuss breakthroughs in pharmaceuticals and
00:01:54.060 other good things. But we also talk about cyber war and autonomous weapons and how our thinking about
00:02:01.780 containing the risk here by analogy to the proliferation of nuclear weapons probably needs to
00:02:09.600 be revised. Anyway, an important conversation, which I hope you find useful. And I bring you Eric Schmidt.
00:02:24.080 I am here with Eric Schmidt. Eric, thanks for joining me.
00:02:27.680 Glad to be with you.
00:02:28.880 So we have, I think we have a hard out at an hour here. So amazingly, that's a short podcast for me.
00:02:35.100 So I'm going to be, there's going to be a spirit of urgency hanging over the place. And we will be
00:02:40.020 efficient in covering the fascinating book that you have written with Henry Kissinger and Daniel
00:02:46.660 Huttenlacher.
00:02:47.800 That's right. And Dr. Kissinger, of course, is this former Secretary of State. And Dan Huttenlacher
00:02:52.020 is now the Dean of Artificial Intelligence and Computer Science at the Schwartzen Center at MIT. He's a
00:02:59.160 proper computer scientist.
00:03:01.080 Yeah. And that book is The Age of AI and Our Human Future, where you cover, you know, most of what I have
00:03:07.980 said about AI thus far, and every case where I have worried about our possible AI future has been focused
00:03:16.460 on the topic of AGI, artificial general intelligence, which you discuss briefly in the book, but it's not your
00:03:23.180 main focus. So I thought maybe we could save that for the end, because I would love to get your take on AGI.
00:03:29.080 But there are far more near-term concerns here and considerations that we could cover. And you are
00:03:36.640 quite well-placed to cover them, because if I'm not mistaken, you ran Google for, what was it, 10 years?
00:03:44.500 That's correct.
00:03:45.460 What was your background before that? How did you come to be the CEO of Google?
00:03:50.320 Well, I'm a computer scientist. I have a PhD in the area. And I worked for 45 years in tech in one way or
00:03:57.040 the other whole bunch of companies. Larry and Sergey brought me in as the early CEO of the company, and we built
00:04:03.860 it together. After a decade, I became chairman, Larry became CEO, and then he replaced himself with Sundar,
00:04:09.920 who's now doing a phenomenal job at Google. So I'd say collectively, this group, of which I'm a member, built one of
00:04:16.620 the great companies. I'm really proud of that.
00:04:19.200 Yeah. And obviously, Google is quite involved in developing AI. I just saw just the other day that
00:04:25.780 there's a new, I think it's a 540 billion parameter language model that is beating the average human at
00:04:34.440 something like 150 cognitive tests now. And it seems like the light is at the end of the tunnel there.
00:04:41.100 It's just going to be a larger model that's going to beat every human at those same tasks. But before
00:04:46.840 we get into some of the details here, I just want to organize our general approach to this. There are
00:04:53.320 three questions that Kant asked in his critique of pure reason, I think it was, which seem unusually
00:05:01.000 relevant to the development of AI. The first is, what can we know? The second is, what should we do?
00:05:09.340 And the third is, what is it reasonable to hope for? And I think those really do capture almost every
00:05:19.260 aspect of concern here. Because as you point out in the book, AI really promises to, and it has already
00:05:26.820 begun to shift the foundations of human knowledge. So the question of what we can know and how we can
00:05:32.900 know it is enormously salient now. And maybe we can talk about some of those examples. But obviously,
00:05:38.760 this question of what should we do and what can we reasonably hope for captures the risks we're
00:05:45.180 running in developing these systems. And we're running these risks well short of producing anything
00:05:52.740 like artificial general intelligence. And it's interesting that we're on a path now where we're
00:05:57.680 really not free to decline to produce this technology. I mean, to my eye, there's really no brake
00:06:04.680 to pull. I mean, we're in a kind of AI arms race now. And the question is how to put that race for
00:06:12.060 more intelligence on a footing that is not running cataclysmic risk for us. So before we jump into the
00:06:20.660 details, I guess I'd like to get your general thoughts on how you view the stakes here and where
00:06:27.200 you view the field to be at the moment. Well, of course, we wrote the book Age of AI
00:06:32.540 precisely to help answer the questions you're describing, which are perfectly cast.
00:06:39.340 And what's happened in the book, which is written roughly a year ago and then published,
00:06:44.720 we described a number of examples to illustrate the point. One is the development of new moves in
00:06:51.520 the game of Go, which is 2,500 years, which were discovered by a computer, which humans had never
00:06:57.680 discovered. It's hard to imagine that humans wouldn't have discovered these strategies, but
00:07:02.740 they didn't. And that calls the question of, are there things which AI can learn that humans cannot
00:07:08.740 master? That's a question. The second example that we use is the development of a new drug called
00:07:15.020 Hallicin, which is a broad spectrum antibiotic, which could not be done by humans, but a set of
00:07:21.680 neuroscientists, biologists, and computer scientists put together a set of programs that ultimately searched
00:07:28.260 through 100 million different compounds and came up with candidates that were then subsequently tested,
00:07:34.760 advancing drugs at an enormous rate. That's another category of success in AI. And then the third is what
00:07:41.340 you've already mentioned, which is large language models. And we profile in the book GPT-3, which is
00:07:46.220 the predecessor of the one you described. And it's eerie. On the back cover of our book, we say to the
00:07:53.400 GPT-3 computer, are you capable of human reasoning? And it answers, no, I am not. You may wonder why I
00:08:02.140 give you that answer. And the answer is that you are a human reasoning machine, whereas I am a language
00:08:09.740 model that's not been taught how to do that. Now, is that awareness or is that clever mimicry? We don't
00:08:17.680 know. But each of these three examples show the potential to answer Kant's questions. What can we
00:08:24.660 know? What will happen? What can we do about it? Since then, this past few weeks, we've seen the
00:08:31.500 announcement that you mentioned of this enormous large language model, which can beat humans on many
00:08:36.100 things. And we've also seen something called DAL-E, which is a text-to-art program. You describe roughly
00:08:43.740 what you want, and it can generate art for you. Now, these are the beginnings of the impact of
00:08:49.000 artificial intelligence on us as humans. So Dr. Kissinger, Dan, and myself, when we looked at those,
00:08:54.820 we thought, what happens to society when you have these kinds of intelligence? Now, they're not human
00:09:02.200 intelligence. They're different kinds of intelligence in everyday life. And we talk about all the
00:09:07.160 positives, of which there are incredible positives. Better materials, better drugs, more efficient
00:09:13.540 systems, better understanding, better monitoring of the earth, additional solutions for climate change.
00:09:20.380 There's a long, long list which I can go through. Very, very exciting. And indeed, in my personal
00:09:25.800 philanthropy, we are working really hard to fund AI-enabled science discoveries. We recently announced
00:09:32.160 a grant, a structure with a guy named James Manyika, who's a friend of mine, of $125 million
00:09:39.360 to actually go and fund research on the really hard problems in AI, the ones that you're mentioning and
00:09:47.480 others, and also the economic impacts and so forth. So I think people don't really know.
00:09:51.400 The real question is, what happens when these systems become more commonplace? Dr. Kissinger
00:09:58.840 says, if you look at history, when a system that is not understandable is imposed on people,
00:10:06.180 they do one of two things. They either invent it as a religion, or they fight it with guns. So my
00:10:14.760 concern, and I'll say it very directly, is we're playing with the information space of humans.
00:10:21.420 We're experimenting at scale without a set of principles as to what we want to do. Do we care
00:10:28.500 more about freedom? Do we care more about efficiency? Do we care more about education? And so forth.
00:10:33.620 And Dr. Kissinger would say, the problem is that these decisions are being made by technical people
00:10:39.460 who are ignorant of the philosophical questions that you so ably asked. And I agree with him,
00:10:45.040 speaking as an example of that. So we recommend, and indeed, I'm trying to now fund that people begin
00:10:54.580 in a multidisciplinary way to discuss the implications of this. What happens to national security?
00:10:59.560 What happens to military intelligence? What happens to social media? What happens to your children when
00:11:06.940 your child's best friend is a computer? And for the audience who might be still thinking about
00:11:13.820 the killer robot, we're not building killer robots, and I hope we never do. This is really about
00:11:20.760 information systems that are human-like, that are learning, they're dynamic, and they're emergent,
00:11:26.400 and they're imprecise, being used and imposed on humans around the world. That process is unstoppable.
00:11:35.780 It's simply too many people working on it, too many ways in which people are going to manipulate it,
00:11:41.500 including for hostile reasons, too many businesses being built, and too much success for some of the
00:11:47.980 early work. Yeah, yeah. I guess if I can just emphasize that point, the unstoppability is pretty
00:11:54.400 interesting, because it's just anchored to this basic fact that intelligence is almost by definition
00:12:02.160 the most valuable thing on Earth, right? And if we can get more of it, we're going to, and we clearly
00:12:08.460 can. And all of these narrow intelligences we've built thus far, you know, all that are effective,
00:12:14.940 that come to market, that we pour resources into, are superhuman, more or less right out of the gate,
00:12:21.620 right? I mean, it's just, it's not a question of, I mean, human level intelligence is a bit of a
00:12:27.400 mirage, because the moment we get something that's general, it's going to be superhuman. And so we can
00:12:32.780 leave the generality aside, all of these piecemeal intelligences are superhuman. And I mean, the example
00:12:40.360 you give of the new antibiotic, Hallicin, I mean, it's fascinating, because it's not just a matter of
00:12:47.320 doing human work faster. If I understand what happened in that case, this is an AI detecting
00:12:54.380 patterns and relationships in molecules already known to be, you know, safe and efficacious as
00:13:02.640 antibiotics, and detecting new properties that human beings very likely would never have conceived of,
00:13:10.100 and may in fact be opaque to the people who built the AI and may remain opaque. I mean, one of the
00:13:17.260 issues you just raised is the issue of transparency. Many of these systems are built in such a way as to
00:13:23.720 be black boxes, and we don't know how the AI is doing what it's doing in any specific way. It's just
00:13:32.220 training against data and against its own performance, so as to produce a better and better result, which qualifies as
00:13:40.100 intelligent and even superhumanly so. And yet, it may remain a black box. Maybe we can just close the
00:13:47.360 loop on that specific problem here. Are you concerned that transparency is a necessity when decision-making
00:13:57.120 is important? I mean, just imagine the case where we have a something like an AI oracle that we are
00:14:04.660 convinced makes better decisions than any person or even any group of people, but we don't actually
00:14:12.160 know the details of how it's making those decisions, right? So this is, I mean, you can just multiply
00:14:18.400 examples as you like, but just, you know, questions of, you know, who should get out of prison, you know,
00:14:23.100 the likelihood of recidivism in the case of any person, or, you know, who's likely to be, you know,
00:14:29.140 more violent, you know, at the level of conviction, right? Like, what should the prison sentence be?
00:14:34.400 I mean, it's very easy to see that if we're shunting that to a black box, people are going to
00:14:41.380 get fairly alarmed that in any differences in outcome that are not transparent. Perhaps you
00:14:48.960 have other examples of concern, but do you think transparency is something that, I mean, one question
00:14:54.360 is, is it technically feasible to render black boxes transparent when it matters? And two, is
00:15:01.720 transparency as important as we intuitively may think it is? Well, I wonder how important transparency
00:15:07.820 is for the simple fact that we have teenagers among our midst, and the teenagers cannot explain
00:15:13.940 themselves at all, and yet we tolerate their behavior with some restrictions because they're not
00:15:19.340 full adults. So, but we wouldn't let a teenager fly an airplane or operate on a patient. So I think a
00:15:26.940 pretty simple model is that at the moment, these systems cannot explain how they came to their
00:15:32.600 decision. There are many people working on the explainability problem. Until then, I think it's
00:15:37.520 going to be really important that these systems not be used in what I'm going to call life safety
00:15:42.620 situations. And this creates all sorts of problems, for example, in automated war, automated
00:15:48.040 conflict, cyber war, those sorts of things, where the speed of decision-making is faster than what
00:15:53.120 humans can, what happens if it makes a mistake? And so, again, we're at the beginning of this process,
00:16:00.200 and most people, including myself, believe that the explainability problem and the bias problems
00:16:05.580 will get resolved because there's just too much money, too many people working on it,
00:16:10.580 maybe at some cost, but we'll get there. That's historically how these things work. You start off
00:16:14.380 with stuff that works well enough, but it shows a hint of the future, and then it gets industrialized.
00:16:19.820 I'm actually much more focused on what's it like to be human when you have these specialized systems
00:16:26.400 floating around. My favorite example here is Facebook, where they change their feed to amp it
00:16:33.400 using AI. And the AI that they built was around engagement. And we know from a great deal of social
00:16:40.980 science that outrage creates more engagement. And so, therefore, there's more outrage on your feed.
00:16:47.460 Now, that was a clearly deliberate decision on part of Facebook. Presumably thought it was a good
00:16:52.560 product idea, but it also maximized their revenue. That's a pretty big social experiment, given the
00:16:58.380 number of users that they have, which is not done with an understanding, in my view, of the impact of
00:17:04.340 political polarization. Now, you sit there and you go, okay, well, he doesn't work at Facebook. He
00:17:09.940 doesn't really understand. But many, many people have commented on this problem. This is an image
00:17:17.120 of what happens in a world where all of the information around you can be boosted or manipulated
00:17:22.460 by AI to sell to you, to anchor you, to change your opinion, and so forth. So, we're going to face some
00:17:29.240 interesting questions. In the information space, the television and movies and things you see online
00:17:35.440 and so forth, do there need to be restrictions on how AI uses the information it has about you
00:17:42.120 to pitch to you, to market to you, to entertain you? These are questions. We don't have answers.
00:17:48.680 But it makes perfect sense that in the industrialization of these tools, the tools that I'm
00:17:53.680 describing, which were invented in places like Google and Facebook, will become available to everyone,
00:17:59.240 in every government. So, another example is a simple one, which is the kid is a two-year-old and gets
00:18:07.320 a toy. And the toy gets upgraded every year and the kid gets smarter. The toy is now, the kid is now 12
00:18:13.480 and there's the 10 years from now, there's a great toy. And this toy is smart enough in non-human terms
00:18:20.820 to be able to watch television and decide if the kid likes the show. So, the toy is watching the
00:18:27.340 television and the kid, the toy says to the kid, I don't like this show, knowing that the kid's not
00:18:33.500 going to like it. And the kid goes, I agree with you. Now, is that okay? Probably. Well, what happens
00:18:41.180 if that same system that's also learning learns something that's not true? And it goes, you know,
00:18:49.620 kid, I have a secret. And the kid goes, tell me, tell me, tell me. And the secret is something which
00:18:56.720 is prejudicial or false or bad or something like that. We don't know how to describe,
00:19:02.820 especially for young people, the impact of these systems on their cognitive development.
00:19:08.020 Now, we have a long history in America of having school boards and textbooks which are approved at the
00:19:14.060 state level. Are the states going to monitor this? And you sit there and you say, well, no parent would
00:19:20.400 allow that. But let's say that the normal behavior of this toy, it's smart enough, understands the kid
00:19:26.480 well enough to know the kid's not good at multiplication. So, the kid's bored and the toy says, I think we
00:19:33.540 should play a game. Kid goes great. And of course, it's a game which strengthens his or her
00:19:38.640 multiplication capability. So, on the one hand, you want these systems to make people smarter,
00:19:45.420 make them develop, make them more serious adults, make the adults more productive.
00:19:50.340 Another example would be my physics friends. They just want a system to read all the physics books
00:19:54.580 every night and make some adjustments to them. Well, the physicists are adults who can deal with this.
00:19:59.800 But what about kids? So, you're going to end up in a situation, at least with kids and with
00:20:04.960 elderly who are isolated, where these tools are going to have an out-of-proportion impact
00:20:11.680 on society as they perceive it. We've never run that experiment. Dynamic, emergent, and not precise.
00:20:20.140 I'm not worried about airplanes being flown by AI because they're not going to be reliable enough
00:20:24.760 to do it for a while. Now, we should also say for the listeners here that we're talking about a term
00:20:32.860 which is generally known as narrow AI. It's very specific, and we're using specific examples,
00:20:39.160 drug discovery, education, entertainment. But the eventual state of AI is called general intelligence,
00:20:46.600 where you get human kind of reasoning. In the book, what we describe that as the point where the
00:20:53.820 computer can set its own objective. And today, the good news is the computer can't choose its objective.
00:21:00.740 At some point, that will not be true.
00:21:04.540 Yeah. Yeah. Well, hopefully, we'll get to AGI at the end of this hour. But I think we should talk
00:21:11.160 about the good and the bad in that order, and maybe just spend a few minutes on the good. Because
00:21:17.700 the good is all too obvious. Again, and intelligence is the most valuable thing on Earth. It's the thing
00:21:25.120 that gives us every other thing we want, and it's the thing that safeguards everything we have. And if
00:21:32.560 there are problems we can't solve, well, then we can't solve them. But if there are problems that can
00:21:37.560 be solved, the way we will solve them is through greater uses of our intelligence. And insofar as we
00:21:46.080 can leverage artificial intelligence to solve those problems, we will do that, more or less regardless of
00:21:52.440 the attendant risks. And that's the problem, because the attendant risks are increasingly obvious,
00:21:57.920 and it seems not at all trivial. And we've already proven we're capable of implementing
00:22:05.720 massive technological change without really thinking about the consequences at all. You cite
00:22:12.920 the massive psychological experiment we've performed on all of humanity with no one really consenting,
00:22:19.880 that is, social media. And it's, you know, the effects are ambiguous at best. I mean, there's
00:22:26.240 some obviously bad effects, and it's not even straightforward to say that democracy or even
00:22:33.220 civilization can survive contact with social media. I mean, that remains to be seen, given how divisive
00:22:39.520 some of its effects are, I consider social media to be far less alarming than the prospect of having
00:22:46.300 an ongoing nuclear doctrine anchored to a proliferating regime of cyber espionage, cyber terrorism,
00:22:57.660 cyber war, all of which will be improved massively by layering AI onto all of that. So before we jump
00:23:06.740 into the bad, which is, you know, really capturing my attention, is there anything specifically you
00:23:12.540 want to say about the good here? I mean, if this goes well, what are you hoping for? What are you
00:23:18.600 expecting? Well, there are so many positive examples that we honestly just don't have time to make a list.
00:23:25.340 I'll give you a few. In physics and math, the physicists and mathematicians have worked out the
00:23:32.220 formulas for how the world works, at least at the scientific level. But many of their calculations
00:23:37.580 are not computable by modern computers. They're just too complicated. An example is how do clouds
00:23:44.160 actually work is a function of something called the Navier-Stokes equations, which for a normal-sized
00:23:49.700 cloud would take 100 million years for a computer to figure out. But using an AI system, and there's a
00:23:56.200 group at Caltech doing this, they can come up with a simulation of the things that they care about.
00:24:02.220 In other words, the AI provides enough accuracy in order to solve the more general climate modeling
00:24:10.460 problem. If you look at quantum chemistry, which is sort of how does, how do chemical bonds work
00:24:17.020 together? Not computable by modern methods. However, AI can provide enough of a simulation
00:24:23.740 that we can figure out how these molecules bind, which is the halicin example.
00:24:28.760 In drug discovery, we know enough about biology that we can basically predict that if you do
00:24:37.320 these compounds with, you know, this antibody, we can make it stronger, we can make it weaker,
00:24:44.080 and so forth, in the computer, and then you go reproduce it in the lab. There's example after example
00:24:50.400 AI is being used from existing data to simulate a non-computable function in science. And you say,
00:24:59.560 what's he talking about? I'm talking about the fact that the scientists have been stuck for decades
00:25:05.320 because they know what they want to do, but they couldn't get through this barrier. That unleashes
00:25:11.080 new materials, new drugs, new forms of steel, new forms of concrete, and so forth and so on.
00:25:17.440 It also helps us with climate change, for example, because climate change is really about energy and
00:25:23.160 CO2 emission and so forth. These new surfaces, discoveries, and so forth will make a material
00:25:27.680 difference. And I'm talking about really significant numbers. So that's an example.
00:25:33.060 Another example is what's happening with these large language models that you mentioned earlier,
00:25:38.180 that people are figuring out a way to put a conversational system in front of it so that
00:25:41.900 you can talk to it. And the conversational system has enough state that it can remember what it's
00:25:46.820 talking about. It's not like a question, answer, question, answer, and it doesn't remember.
00:25:51.140 It actually remembers the context of, oh, we're talking about the Oscars, and we're talking about
00:25:55.260 what happened at the Oscars, and what do I think? And then it sort of goes, and it gives you a thoughtful
00:26:00.160 answer as to what happened and what is possible. In my case, I was playing with one of them
00:26:06.840 a few months ago. And this one, I asked the question, what is the device that's in 2001,
00:26:14.940 A Space Odyssey, that I'm using today? There's something from 1969 that I'm using today that
00:26:20.180 was foreshadowed in the movie. And it comes right back and says, the iPad. Now that's a question that
00:26:26.840 Google won't answer if you ask it the way I did. So I believe that the biggest positive impact
00:26:34.400 will be that you'll have a system that you can verbally or by writing, ask it questions,
00:26:40.920 and it will make you incredibly smarter, right? That it'll give you the nuance and the understanding
00:26:46.660 and the context. And you can ask it another question, and you can refine your question.
00:26:50.940 Now, if you think about it in the work you do, or that I do, or that a scientist does,
00:26:55.280 or a politician, or an artist, this is enormously transformative.
00:26:59.140 So example after example, these systems are going to build scientific breakthroughs,
00:27:07.160 scalable breakthroughs. Another example was that a group at DeepMind figured out the folding
00:27:13.180 structure of proteins. And proteins are the way in which biology works. And the way they fold determines
00:27:19.300 their effectiveness, what they actually do. And it was thought to be not really computable.
00:27:23.960 And using these techniques in a very complicated way with a whole bunch of protein scientists,
00:27:29.700 they managed to do it. And their result was replicated in a different mechanism with different
00:27:34.180 AI from something called the Baker Lab in University of Washington. The two together have given us a map
00:27:40.380 of how proteins work, which in my view is worthy of a Nobel Prize. That's how big a discovery that is.
00:27:46.480 All of a sudden, we are unlocking the way biology works, and it affects us directly.
00:27:50.460 But those are some positive examples. I think the negative examples...
00:27:55.760 Well, let's wait, because I'm chock full of negative examples.
00:27:58.880 Okay.
00:27:59.300 But I'm interested in how even the positive can disclose a surprisingly negative possibility,
00:28:09.340 or at least it becomes negative if we haven't planned for it ethically, politically, economically.
00:28:15.320 And so you imagine the success. You imagine that more and more... So what you've just pictured was
00:28:22.260 a future of machine and human cooperation, right, and facilitation, where people just get smarter
00:28:30.740 by being able to have access to these tools, or they get effectively smarter. But you can imagine,
00:28:38.340 just in the limit, more and more getting seeded to AI, because AI is just better at doing these things.
00:28:45.480 It's better at proving theorems. It's better at designing software. It's better, it's better,
00:28:50.020 it's better. And all of a sudden, the need for human developers at all, or human mathematicians at
00:28:56.160 all, or you just make the list as long as you want. It seems like some of the highest status jobs
00:29:05.540 cognitively might be among the first to fall, which is to say, I certainly expect at this point
00:29:12.820 to have an AI radiologist, certainly, before I have an AI plumber. And there's a lot more above and
00:29:24.120 beyond the radiology side of that comparison that I think is going to fall before, you know,
00:29:29.980 the basic manual tasks fall to robots. And this is a picture of real success, right? Because
00:29:37.900 in the end, all we're going to care about is performance. We're not going to care about
00:29:42.000 keeping a monkey in the loop just for reasons of sentimentality. You know, if you're telling me
00:29:49.240 that my car can drive a thousand times better than I can, which is to say that, you know, it's going
00:29:55.020 to reduce my risk of getting in a fatal accident, you know, killing myself or killing someone else
00:29:59.840 by a factor of a thousand if I just flip on autopilot, well, then not only am I going to flip it on,
00:30:07.060 I'm going to consider anyone who declines to do that to be negligent to the point of criminality.
00:30:13.100 And that's never going to change. Everything is going to be in the position of a current chess
00:30:18.480 master who knows that the best player on earth is never going to be a person ever again, right?
00:30:25.920 Because of AlphaZero. So take that wherever you want.
00:30:29.880 I disagree a little bit, and I'll tell you why. I think you're correct in about 30 years,
00:30:35.620 but I don't think that argument is true in the short term.
00:30:38.180 Yeah. No, I was not, just to be clear, I'm not suggesting any timeframe there. I'm just saying,
00:30:43.260 ultimately, if we continue to make progress, something like this seems bound to happen.
00:30:49.760 Yes. But what I want to say is, I defy you to argue with me that making people smarter is a bad
00:30:58.500 thing. Okay. So let's start with the premise of the human assistant, that is the thing that you're
00:31:06.760 using, will make humans smarter. It'll make it deeper, better analysis, better choices.
00:31:14.840 But at least the current technology cannot replace essentially the free will of humans.
00:31:23.240 They sort of wake up in the morning, you have a new idea, you decide something, you say,
00:31:26.960 that's a bad idea, so forth and so on. We don't know how to do that yet. And I have some speculation
00:31:31.880 on how that will happen. But in the next decade, we're going to not be solving that problem. We'll
00:31:38.180 be solving a different problem, which is how do we get the existing people doing existing jobs to do
00:31:43.380 them more efficiently, that is smarter, better, faster. One of the, when we looked at the funding
00:31:50.100 for this AI program that I've since announced, the funding 125 million, a fair chunk of it is going
00:31:56.040 to really hard computer science problems. Some of them include, we don't really understand how to
00:32:01.240 explain what they're doing. As I mentioned, they're also brittle. When they fail, they can fail
00:32:06.540 catastrophically. Like, why did it fail? And no one can explain. There are hardening, there are resistance
00:32:12.340 to attack problems. There are a number of problems of this kind. These are hard computer science problems,
00:32:17.520 which I think we will get through. They use a lot of power, the algorithms are expensive, that sort of
00:32:22.340 thing. But we have also focusing around the impact on jobs and employment and economics.
00:32:28.480 We're also focusing on national security. And we're focusing on the question that you're asking,
00:32:33.940 which is, what's our identity? What does it mean to be human? Before general intelligence comes,
00:32:41.120 we have to deal with the fact that these systems are not capable of choosing their own outcome,
00:32:46.940 but they can be applied to you as a citizen by somebody else against your own satisfaction.
00:32:53.600 So the negatives before AGI are all of the form, misinformation, misleading information,
00:33:02.580 creating dangerous tools, and for example, dangerous viruses. For the same reason that we built a
00:33:09.100 fantastic new antibiotic drug, it looks like, you could also imagine a similar evil team of producing
00:33:16.500 an incredible number of bad viruses, things that would hurt people. And you could imagine in that
00:33:22.120 scenario, they might be clever enough to be able to hurt a particular race or particular sex or
00:33:26.960 something like that, which would be totally evil and obviously a very bad thing. We don't have a way
00:33:32.620 of discussing that today. So when I look at the positives and negatives right now, I think the
00:33:38.160 positives, as with many technologies, really overwhelm the negatives, but the negatives need to be
00:33:45.140 looked at. And we need to have the conversation right now about, let's use social media, which is an easy
00:33:51.740 whipping boy here. I would like, so I'm clear what my political position is, I'm a very strong proponent
00:33:59.000 of freedom of speech for humans. I am not in favor of freedom of speech for computers, robots, bots,
00:34:07.260 so forth and so on. I want an option with social media, which says, I only want to see things that a human
00:34:13.160 has actually communicated from themselves. I want to know that it wasn't snuck in by some Russian
00:34:19.320 agent. I want proof of providence and I want to know there's a human. And if it's a real human who's
00:34:25.440 in fact, you know, an idiot or crazy or whatever, I want to be able to hear their voice and I want to
00:34:30.800 be able to decide I don't agree with it. What's happening instead is these systems are being boosted.
00:34:36.700 They're being pitched, they're being sold by AI. And I think that's got to be limited in some way.
00:34:43.820 I'm in favor of free speech, but I don't want only some people to have megaphones.
00:34:49.220 And if you talk to politicians and you look at the political structure in the country,
00:34:53.820 this is a completely unintended effect of getting everyone wired. Now, is it a human or is it a
00:35:01.180 computer? Is it a Russian, a Russian compromise plane, or is it an American? Those things need
00:35:07.540 to get resolved. You cannot run a democracy without some level of trust.
00:35:12.400 Yeah. Yeah. Well, let's take that piece here. And obviously it extends beyond
00:35:16.800 the problem of AI's involvement in it, but the misinformation problem is enormous.
00:35:23.640 What are your thoughts about it? Because I'm just imagining we've been spared thus far the worst
00:35:30.780 possible case of this, which is just imagine under conditions of where we had something like perfect
00:35:39.120 deep fakes, right, that were truly difficult to tell apart from real video, what would the
00:35:45.640 controversy around the 2020 election have looked like or the war in Ukraine and our dealings with
00:35:52.080 Putin at this moment, right? Like just imagine, you know, a perfect deep fake of Putin declaring a
00:35:59.480 nuclear first strike on the U.S. or whatever. I mean, you just, you know, just imagine essentially a
00:36:05.220 writer's room from hell where you have smart, creative people spending their waking hours figuring out how
00:36:12.200 to produce media that is shattering to every open society and conducive to provoking international
00:36:21.360 conflict. That is clearly coming in some form. I guess my first question is, are you hopeful that
00:36:29.820 the moment that arrives, we will have the same level of technology that can spot deep fakes? Or is there
00:36:36.500 going to be a lag there of months, years that are going to be difficult to navigate?
00:36:43.660 We don't know. There are people working really hard on generating deep fakes, and there are people
00:36:48.640 working really hard on detecting deep fakes. And one of the general problems with misinformation
00:36:54.320 is we don't have enough training data. The term here is, in order to get an AI system to know
00:37:01.160 something, you have to give it enough examples of good, bad, good, bad, and eventually you can say,
00:37:06.300 oh, here's something new, and I know if it's good or bad. And one of the core problems in
00:37:10.960 misinformation is we don't have enough agreement on what is misinformation or what have you.
00:37:15.640 And the thought experiment I would offer is, President Putin in Russia has already shut down
00:37:21.260 the internet and free speech and controls the media and so forth. So let's imagine that he was
00:37:27.880 further evil.
00:37:30.320 If you'd like to continue listening to this conversation, you'll need to subscribe at
00:37:34.400 samharris.org. Once you do, you'll get access to all full-length episodes of the Making Sense podcast,
00:37:39.600 along with other subscriber-only content, including bonus episodes and AMAs, and the conversations
00:37:45.800 I've been having on the Waking Up app. The Making Sense podcast is ad-free and relies entirely on
00:37:51.340 listener support. And you can subscribe now at samharris.org.