WEBVTT - Ep120 "Will AI build us into better humans?" 0:00:05.160 --> 0:00:09.559 Will AI make humans better? And what does this have 0:00:09.640 --> 0:00:13.560 to do with linguistics or the movie Arrival or self 0:00:13.640 --> 0:00:18.400 driving cars or debate and video games and elections and 0:00:18.600 --> 0:00:26.160 chess and the ancient game of go. Welcome to Intercosmos 0:00:26.160 --> 0:00:29.600 with me David Eagleman. I'm a neuroscientist and an author 0:00:29.640 --> 0:00:33.160 at Stanford and in these episodes we seek to understand 0:00:33.240 --> 0:00:35.879 why and how our lives look the way they do. 0:00:36.400 --> 0:00:40.560 In today's episode is about whether AI will make us 0:00:40.760 --> 0:00:57.880 better humans. I often find myself totally flabbergasted by the 0:00:58.080 --> 0:01:01.000 change that I've seen just in my lifetime. When I 0:01:01.200 --> 0:01:05.520 was a really little kid, personal computers didn't exist, and 0:01:05.600 --> 0:01:08.759 then we passed through a door and suddenly they did. 0:01:09.160 --> 0:01:10.960 And I saved up my money and I got a 0:01:11.319 --> 0:01:14.640 common or VIC twenty computer, which was something my parents 0:01:14.640 --> 0:01:17.160 had never seen the likes of. And I felt like 0:01:17.200 --> 0:01:20.919 I was living the largest change in human history, because 0:01:20.959 --> 0:01:25.600 for the first time, everybody could have a machine that 0:01:25.640 --> 0:01:28.840 would do all kinds of things. But that turned out 0:01:28.920 --> 0:01:31.640 not to even be the biggest change, because the next 0:01:31.680 --> 0:01:34.600 stop was even bigger, and that was the idea that 0:01:34.680 --> 0:01:37.720 some people had to build a system where we could 0:01:38.040 --> 0:01:41.960 keep information on computers and make computers talk to each 0:01:42.000 --> 0:01:46.520 other such that if the Soviets bombed America, the important 0:01:46.520 --> 0:01:51.360 information wouldn't just be stored on one computer, and messages 0:01:51.440 --> 0:01:54.640 could follow different routes on the network, and that way 0:01:54.680 --> 0:01:58.480 you had a very robust system for keeping information. That 0:01:58.600 --> 0:02:01.920 was of course Arpanet, which became the Internet, and not 0:02:02.080 --> 0:02:04.960 long after, the idea was introduced of a way that 0:02:05.440 --> 0:02:08.959 everyone could use this giant network, not just with text, 0:02:09.000 --> 0:02:10.880 but with graphics, and. 0:02:10.800 --> 0:02:13.400 That was the birth of the Worldwide Web. 0:02:13.919 --> 0:02:17.280 And as soon as that technology existed, then people started 0:02:17.280 --> 0:02:20.480 figuring out what to do with it, and one young 0:02:20.520 --> 0:02:23.280 man started to sell books over the Internet. 0:02:22.919 --> 0:02:24.120 And that became Amazon. 0:02:24.240 --> 0:02:28.400 And two graduate students at Stanford asked the question of 0:02:28.680 --> 0:02:30.480 how the heck we were going to be able to 0:02:30.600 --> 0:02:34.840 find information on this sprawling network, and they created a 0:02:34.840 --> 0:02:38.680 way of measuring all the connections between web pages and 0:02:38.720 --> 0:02:40.880 that gave the ability to search for it, and. 0:02:40.760 --> 0:02:43.640 That garage project grew into Google. 0:02:44.000 --> 0:02:46.959 And around the same time, a Harvard kid was thinking 0:02:46.960 --> 0:02:50.799 about a better way to allow his classmates to get 0:02:50.800 --> 0:02:53.520 to know each other, which was traditionally done at the 0:02:53.520 --> 0:02:56.720 beginning of the year by printing a booklet with everybody's 0:02:57.000 --> 0:03:00.200 picture and name, and that little booklet was called called 0:03:00.240 --> 0:03:04.440 a Facebook, and he thought of digitizing that, and on 0:03:04.480 --> 0:03:07.799 and on, people increasingly figured out how to take this 0:03:07.960 --> 0:03:11.600 new technology and make things that would live on top 0:03:11.680 --> 0:03:14.600 of it. And I felt very lucky to have experienced 0:03:14.760 --> 0:03:19.560 two truly world changing inventions in my lifetime, and I 0:03:19.639 --> 0:03:22.280 knew that had launched us into a world that was 0:03:22.400 --> 0:03:28.000 so different from what my great grandparents could have possibly imagined. 0:03:30.160 --> 0:03:33.600 But now here we are in the middle of a 0:03:34.080 --> 0:03:37.680 third revolution. It's related to the first two computers and 0:03:37.720 --> 0:03:41.200 the Internet, but it makes them look like warm up acts. 0:03:41.440 --> 0:03:44.400 And that's the fact that we have created a new 0:03:44.920 --> 0:03:49.360 intelligent species that we're going to be sharing the planet 0:03:49.400 --> 0:03:51.800 with from now on. This is not to say that 0:03:51.960 --> 0:03:55.760 AI has exactly the same type of intelligence that human 0:03:55.800 --> 0:04:01.000 brains have, but obviously it has absorbed the higher knowledge 0:04:01.080 --> 0:04:04.200 sphere of humankind and it can spit that back to 0:04:04.280 --> 0:04:08.160 us with all sorts of remixes. Now, the question is, 0:04:08.440 --> 0:04:11.320 are we in trouble because of this new invention? Have 0:04:11.400 --> 0:04:15.240 we taken things a step too far? Well, let me 0:04:15.280 --> 0:04:19.280 give you an example that's on people's minds. A few 0:04:19.279 --> 0:04:23.400 months ago, a Swiss research team carried out a secret 0:04:23.480 --> 0:04:28.440 experiment in online persuasion. They went to a forum on 0:04:28.480 --> 0:04:33.320 the website Reddit, and in this forum, users post opinions 0:04:33.400 --> 0:04:36.760 and invite other people to challenge them. And so these 0:04:36.800 --> 0:04:42.520 researchers quietly unleashed a set of AI accounts to pretend 0:04:42.640 --> 0:04:45.920 that they were people and to try to change other 0:04:45.920 --> 0:04:51.600 people's minds. So their large language models LMS, they participated 0:04:52.120 --> 0:04:55.600 in the debating just like any other human. These bots 0:04:55.680 --> 0:05:00.560 wrote arguments, they engaged with users, They debated with the 0:05:00.600 --> 0:05:05.760 hope of changing minds. Now, the researchers measured success by 0:05:05.920 --> 0:05:10.839 tallying whenever an original poster publicly admitted that their mind 0:05:10.960 --> 0:05:14.119 had been changed. So how did the bots do well? 0:05:14.480 --> 0:05:19.000 They achieved up to an eighteen percent success rate in 0:05:19.160 --> 0:05:22.400 changing people's minds. Now, the critical piece you need to 0:05:22.440 --> 0:05:26.480 know is that the average human success rate is about 0:05:26.880 --> 0:05:31.679 three percent. So the bots had absolutely crushed their human 0:05:31.720 --> 0:05:37.560 competition in an environment teeming with intellectuals. The bots not 0:05:37.680 --> 0:05:41.680 only survived, but they thrived. In fact, one of the 0:05:41.720 --> 0:05:45.400 AI accounts climbed into the ninety ninth percentile of all 0:05:45.520 --> 0:05:49.400 users on this subreddit. It racked up ten thousand karma 0:05:49.440 --> 0:05:52.680 points along the way. Now, one of the most surprising 0:05:52.839 --> 0:05:57.599 aspects of the experiment was how effectively a small team 0:05:57.640 --> 0:06:02.400 of researchers, operating with modest academic resources, how they were 0:06:02.400 --> 0:06:06.960 able to outperform essentially every human debater on the platform. 0:06:07.279 --> 0:06:09.080 So let that sink in for a second. 0:06:09.240 --> 0:06:13.279 You have a handful of graduate students armed with an 0:06:13.400 --> 0:06:19.240 LM and a sneaky deployment strategy, and they quietly demonstrated 0:06:19.279 --> 0:06:23.680 what it might look like if influence operations were scaled 0:06:23.800 --> 0:06:27.599 by AI. Now, when this data was released, one of 0:06:27.600 --> 0:06:32.040 the scariest parts to people was that nobody noticed that 0:06:32.080 --> 0:06:36.320 these debaters were actually AI and not real humans. This 0:06:36.520 --> 0:06:40.200 reddit channel prides itself as one of the most critically 0:06:40.279 --> 0:06:44.719 minded communities on the platform. If any place could sniff 0:06:44.760 --> 0:06:47.960 out an imposter, it should have been here. But for 0:06:48.080 --> 0:06:51.960 the entire four months of the experiment, the bots played 0:06:52.000 --> 0:06:56.960 along undetected by the humans. So that's worrisome. And there's 0:06:57.040 --> 0:07:00.839 another issue too, which is the role of personal data. 0:07:01.160 --> 0:07:03.760 When the bots were given just a basic profile of 0:07:03.800 --> 0:07:06.520 the user, their aged, their location, and their political leaning, 0:07:06.960 --> 0:07:11.000 the success rate bumped up by one percentage point. Now, 0:07:11.040 --> 0:07:13.840 you might say wait a minute, who cares. One percentage 0:07:13.840 --> 0:07:17.720 point is very small, but in political terms, one percent 0:07:17.840 --> 0:07:21.720 could be enough to swing a lot of national elections. 0:07:22.400 --> 0:07:25.880 The difference between nudging public sentiment one way or another 0:07:26.400 --> 0:07:27.200 can come down to. 0:07:27.200 --> 0:07:28.400 Very tiny tweaks. 0:07:28.880 --> 0:07:32.200 So the lesson that surface here is that even minimal 0:07:32.280 --> 0:07:38.440 personalization can significantly sharpen the edge of an AI's persuasive power. 0:07:39.000 --> 0:07:44.240 So when this story surfaced recently about these debaters not 0:07:44.400 --> 0:07:48.320 being real human beings, the reactions were very grim because 0:07:48.400 --> 0:07:55.080 everyone realized that this small academic experiment illustrated what stealth 0:07:55.400 --> 0:07:58.960 influence operations could look like in the near future. If 0:07:59.000 --> 0:08:02.640 a handful of real searchers could do this undetected, what 0:08:03.000 --> 0:08:07.760 happens when you have state actors or corporations or political 0:08:07.800 --> 0:08:11.960 campaigns with real resources doing the same thing at a 0:08:12.040 --> 0:08:16.800 scale thousands of times larger. Also, what's worse is that 0:08:16.880 --> 0:08:20.400 the failure of everyone to detect that these were bots 0:08:21.000 --> 0:08:25.040 that raises serious questions about how we can protect public 0:08:25.120 --> 0:08:29.040 discourse as we enter this new future. If smart reditor 0:08:29.240 --> 0:08:32.240 debaters couldn't spot the difference between a human and a bot, 0:08:32.800 --> 0:08:37.160 what hope is there for broader audiences. So even though 0:08:37.200 --> 0:08:41.319 the major AI companies have all pledged to avoid building 0:08:41.360 --> 0:08:46.600 models with dangerous capabilities like manipulating public opinion on mass. 0:08:47.400 --> 0:08:52.160 This Reddit experiment suggests the thresholds may already be easier 0:08:52.200 --> 0:08:57.160 to cross than any of us had anticipated. What this 0:08:57.240 --> 0:09:01.800 all means is that with major elections in future years, 0:09:01.880 --> 0:09:06.040 were genuinely gonna have to worry about this. The halcyon 0:09:06.600 --> 0:09:10.280 days are gone when we could assume that the replies 0:09:10.320 --> 0:09:14.000 to our online messages came from a fellow human, and 0:09:14.080 --> 0:09:18.640 the risks are real and the potential for abuse is obvious. 0:09:19.200 --> 0:09:22.199 But for today's episode, we're gonna look at all this 0:09:22.360 --> 0:09:25.600 from a very different angle. I think it might be 0:09:25.679 --> 0:09:30.000 worth asking a different question, which is why did the 0:09:30.040 --> 0:09:34.760 bots succeed? After all, they didn't hack people's brains with 0:09:34.880 --> 0:09:39.920 neural interfaces. They didn't spread fear or disinformation. They didn't 0:09:39.920 --> 0:09:45.840 manipulate emotions or overwhelm users with noise. They simply made 0:09:46.520 --> 0:09:51.480 better arguments. They didn't insult people while arguing with them. 0:09:51.520 --> 0:09:55.400 They didn't do little jabs and digs. They just made 0:09:55.600 --> 0:10:02.400 good arguments empathetically. The bots present their points calmly and 0:10:02.640 --> 0:10:08.199 rationally and persuasively, and when users changed their minds. 0:10:08.320 --> 0:10:10.480 It wasn't because they had been tricked. 0:10:10.520 --> 0:10:16.840 It was because they recognized that another perspective was worth considering. 0:10:32.800 --> 0:10:37.319 Humans often change their minds when faced with sound reasoning 0:10:37.360 --> 0:10:40.560 that can be backed up, and mind changing is a 0:10:40.600 --> 0:10:44.640 great thing. It means you're willing to reconsider some closely 0:10:44.679 --> 0:10:49.400 held opinion when the facts or logic warranted. It's a 0:10:49.440 --> 0:10:53.560 mark of intellectual strength, not a sign that you've been tricked. 0:10:54.240 --> 0:10:59.439 In that light, the success of AI debaters isn't necessarily 0:10:59.480 --> 0:11:02.480 a story about manipulation. It could also be a story 0:11:03.000 --> 0:11:07.920 about raising the bar of debating. Now, as I followed 0:11:07.920 --> 0:11:11.200 the outcome of this AI debater study, it struck me 0:11:11.679 --> 0:11:15.040 that there may be a helpful precedent for thinking about 0:11:15.080 --> 0:11:18.880 this moment, for thinking about how AI could. 0:11:18.640 --> 0:11:20.280 Actually improve us. 0:11:20.640 --> 0:11:24.320 So think about the events that unfolded in the hyper 0:11:24.360 --> 0:11:27.760 competitive world of chess. If you can remember back to 0:11:27.840 --> 0:11:32.000 nineteen ninety seven, IBM had built an AI system called 0:11:32.240 --> 0:11:37.160 Deep Blue, and it defeated the world champion, Gary Kasprov. 0:11:37.320 --> 0:11:41.000 This was seen as a seismic moment. The game of 0:11:41.120 --> 0:11:45.840 chess had been quote unquote solved, and everyone worried that 0:11:46.040 --> 0:11:50.120 human mastery was obsolete and then about two decades later, 0:11:50.559 --> 0:11:55.400 Alpha Go beat the world's top Go player, and the 0:11:55.800 --> 0:12:02.080 choir of concerned voices grew louder. But then something unexpected happened. 0:12:02.760 --> 0:12:06.240 In May of twenty seventeen, the world's number one Go player, 0:12:06.720 --> 0:12:11.600 Could G, faced off against his toughest opponent. G was 0:12:11.640 --> 0:12:14.920 the reigning champion in Go, and you know this is 0:12:14.960 --> 0:12:18.360 the game where two players use smooth black rocks or 0:12:18.400 --> 0:12:22.640 white rocks to surround more territory than their opponent. In 0:12:22.720 --> 0:12:25.280 G's case, he was playing against AI. He was playing 0:12:25.280 --> 0:12:29.120 against Alpha Go, which had been trained on many millions 0:12:29.120 --> 0:12:34.079 of games, and it had deeply absorbed the statistics of 0:12:34.240 --> 0:12:35.400 possible plays. 0:12:35.480 --> 0:12:38.000 So G lost the first game. 0:12:38.480 --> 0:12:42.400 Alpha Go had pulled moves that none of G's human 0:12:42.440 --> 0:12:45.760 opponents had ever thought of, and then G lost the 0:12:45.880 --> 0:12:49.080 second game. The fact is he didn't stand a chance. 0:12:49.200 --> 0:12:52.319 The AI had won over a human in a game 0:12:52.360 --> 0:12:56.640 that was way more complex than chess, and subsequent versions 0:12:56.679 --> 0:13:00.840 of this AI will without doubt continue to win ever more. 0:13:01.480 --> 0:13:04.120 But that's not the interesting part of the story. The 0:13:04.280 --> 0:13:09.320 interesting part is what happened next. G got over his 0:13:09.400 --> 0:13:14.400 embarrassment and he became mesmerized by what the heck had 0:13:14.520 --> 0:13:19.640 just transpired. He studied the games that he lost. Now, 0:13:19.679 --> 0:13:23.440 before he had played Alpha Go, G had won most 0:13:23.480 --> 0:13:26.880 of the games against his human opponents, But after he 0:13:27.000 --> 0:13:30.240 played Alpha Go, he was able to beat his human 0:13:30.280 --> 0:13:35.040 opponents even more easily. In other words, after his species 0:13:35.160 --> 0:13:39.040 shaming defeat in twenty seventeen, G went on to play 0:13:39.360 --> 0:13:42.920 twelve straight matches against fellow humans, and he won. 0:13:42.760 --> 0:13:45.560 Them all in a row. Now, what had happened. 0:13:46.000 --> 0:13:50.320 G had been exposed to new kinds of moves and 0:13:50.520 --> 0:13:55.840 strategies that AlphaGo was pulling off that lay outside the 0:13:55.920 --> 0:14:00.400 traditional ideas. All these moves were legal and pop pole, 0:14:01.120 --> 0:14:03.880 but they were different from what had been played over 0:14:03.920 --> 0:14:08.320 the previous twenty five hundred years. For go eficionados, this 0:14:08.400 --> 0:14:13.040 included novelties like playing a stone directly diagonal to your 0:14:13.080 --> 0:14:18.319 opponent's loan stone, or commonly playing six space extensions while 0:14:18.400 --> 0:14:22.520 humans tend to prefer five space. So G reported that 0:14:22.680 --> 0:14:26.480 playing against the AI was like opening a door to 0:14:26.600 --> 0:14:31.600 another world. Some people worry that AI might make games 0:14:31.600 --> 0:14:35.680 of Chess and Go irrelevant, but amazingly, that is not 0:14:35.800 --> 0:14:41.200 what's happened. When AI Trump's Chess champions and Go champions 0:14:41.800 --> 0:14:46.840 It does so with moves that seem inhumanly creative, but 0:14:46.960 --> 0:14:51.040 all the moves are allowed by the rules. Humans simply 0:14:51.080 --> 0:14:53.520 never thought to go there before. And the key is 0:14:53.560 --> 0:14:57.600 that once the moves are seen by humans, then they're 0:14:57.640 --> 0:15:03.640 easily incorporated into our models. GE's experience with Alpha Go 0:15:04.040 --> 0:15:09.560 illuminated new nooks and crannies in his landscape. It exposed 0:15:10.080 --> 0:15:14.080 pathways that have never been lit up before. So AI 0:15:14.280 --> 0:15:18.920 immediately became a tool for steep improvement. And nowadays all 0:15:19.120 --> 0:15:22.040 chests and Go players above a certain level they all 0:15:22.160 --> 0:15:27.720 train with AI. They study the surprising and sometimes counterintuitive 0:15:28.120 --> 0:15:33.440 and alien strategies of the artificial mind, and just like Koje, 0:15:33.760 --> 0:15:37.800 today's grand masters play a deeper and more creative game 0:15:38.200 --> 0:15:43.240 than ever before. In other words, many commentators are worried 0:15:43.320 --> 0:15:46.760 that AI is going to leave humans far behind, and 0:15:46.800 --> 0:15:48.120 in some respects that's true. 0:15:48.120 --> 0:15:52.320 But as computation improves, so will we. 0:15:53.240 --> 0:15:57.320 AI will illuminate dark parts of our maps, allowing us 0:15:57.320 --> 0:16:01.680 to see new roads we didn't even suspect. The key 0:16:01.680 --> 0:16:03.600 point I want to make here is that instead of 0:16:04.040 --> 0:16:09.560 dampening human excellence, AI sparked a renaissance. In these games, 0:16:09.760 --> 0:16:14.560 human minds are elevated by learning from our artificial cousins. 0:16:14.800 --> 0:16:17.160 And if it were just chests and go, that's one thing. 0:16:17.200 --> 0:16:20.240 But I think we can detect this pattern happening all 0:16:20.400 --> 0:16:23.480 over the place. For example, the same thing has happened 0:16:23.560 --> 0:16:28.560 in poker. Professional poker players master the art of bluffing. 0:16:28.640 --> 0:16:32.040 They project strength when they're weak or weakness when they're strong. 0:16:32.400 --> 0:16:35.280 This has always been seen like a deep test of 0:16:35.360 --> 0:16:40.680 human psychology. The players read faces, they watch for micro expressions, 0:16:40.720 --> 0:16:43.720 they guess intentions, So in this way, poker is a 0:16:43.760 --> 0:16:48.360 really human game. But then Carnegie Mellon cooked up an 0:16:48.400 --> 0:16:51.920 AI called Librotus. Obviously, it doesn't have a face, it 0:16:51.920 --> 0:16:54.720 doesn't sweat, it doesn't blink, so it wasn't really like 0:16:54.960 --> 0:16:59.080 playing another human. So Lebronis started playing poker, and it 0:16:59.160 --> 0:17:03.560 did things that baffled the human players. The AI began 0:17:03.760 --> 0:17:08.160 making betting choices that looked bizarre. Sometimes it over bet 0:17:08.240 --> 0:17:11.320 the pot by huge margins, a move that the human 0:17:11.359 --> 0:17:15.840 players considered reckless. Other times it made really tiny bets 0:17:15.960 --> 0:17:19.640 in situations where no human would bother and the human 0:17:19.640 --> 0:17:23.639 players dismissed this all as network nonsense. But as the 0:17:23.760 --> 0:17:29.880 games went on, they realized these inhuman strategies were working. 0:17:29.920 --> 0:17:33.760 The AI was winning, and it was doing so by 0:17:33.920 --> 0:17:38.480 reinventing the language of poker. And now, just like with 0:17:38.600 --> 0:17:43.399 chess and go, professional poker players study these strategies. You 0:17:43.440 --> 0:17:47.520 have human players adopting patterns of play that just didn't 0:17:47.520 --> 0:17:52.639 exist before the machines taught them. AI made humans better 0:17:52.720 --> 0:17:55.840 poker players. And let me give you another example. There's 0:17:55.880 --> 0:18:00.679 a strategy video game called StarCraft two where players build 0:18:00.800 --> 0:18:04.160 armies and manage resources and try to outwit their opponent. 0:18:04.520 --> 0:18:07.919 So the company DeepMind set some AI agents on it, 0:18:08.080 --> 0:18:12.600 and those agents showed humans an entirely different way of 0:18:12.680 --> 0:18:16.720 going about running the strategy. At first, people thought the 0:18:16.840 --> 0:18:22.040 AI play seemed unfair and robotic, but then human players 0:18:22.119 --> 0:18:26.640 began to adapt. They started rethinking their own strategies. They 0:18:26.680 --> 0:18:32.640 started redesigning their build orders. They discovered things like sometimes 0:18:32.960 --> 0:18:37.000 sacrificing entire groups of units early in the game could 0:18:37.040 --> 0:18:41.160 produce long term advantages. These were not moves that humans 0:18:41.160 --> 0:18:43.760 had come up with, or possibly ever would have come 0:18:43.840 --> 0:18:47.520 up with, but once they were revealed, they became part 0:18:47.600 --> 0:18:50.760 of the human playbook. Or take a different video game 0:18:50.920 --> 0:18:54.879 called Dota two. You've got professional players. But then open 0:18:54.920 --> 0:18:58.760 AI built a system that used strategies that looked to 0:18:58.880 --> 0:19:03.560 human professionals like they were clumsy. The AI would push 0:19:03.600 --> 0:19:08.320 aggressively when humans would retreat or take risks that seemed absurd, 0:19:08.840 --> 0:19:12.679 but more often than not, the machines won, and just 0:19:12.720 --> 0:19:16.120 like with Chests and Go and StarCraft, the human professionals 0:19:16.160 --> 0:19:19.479 were forced to ask themselves had they been playing with 0:19:19.640 --> 0:19:24.800 blinders on all along? After all, the machines were uncovering 0:19:24.960 --> 0:19:29.199 landscapes that we just never knew were there. In other words, 0:19:29.240 --> 0:19:33.960 whenever AI discovers new pathways inside the universe of a game, 0:19:34.280 --> 0:19:38.439 it illuminates something for us about the fence lines of 0:19:38.520 --> 0:19:44.000 our own imagination. Humans had decades to explore video games, 0:19:44.040 --> 0:19:48.719 and centuries to explore poker, and millennia to explore the 0:19:48.760 --> 0:19:53.560 games of Chess and Go. But within weeks AI uncovered 0:19:53.680 --> 0:19:58.240 strategies that had never even struck us. And this is 0:19:58.320 --> 0:20:02.880 the point of today's episode. AI can expand the possibility 0:20:02.880 --> 0:20:05.800 space for us. What we've seen in the past few 0:20:05.880 --> 0:20:09.720 years is the limits of our imagination. Even in domains 0:20:10.080 --> 0:20:13.880 we thought we had mastered, our internal models might be 0:20:14.240 --> 0:20:17.639 a lot more narrow than we had ever realized. But 0:20:17.760 --> 0:20:21.720 by playing with machines, we're learning how they think, and 0:20:21.840 --> 0:20:26.040 more importantly, how we might think differently. So beyond chests 0:20:26.119 --> 0:20:29.400 or go or video games, the bigger game is how 0:20:29.440 --> 0:20:32.560 we can expand our own internal models. A lot of 0:20:32.600 --> 0:20:36.440 people cast this as man versus machine, but I think 0:20:36.480 --> 0:20:40.920 a more productive lens is seeing it as man learning 0:20:41.000 --> 0:20:45.159 from machine. The prize is a new way of seeing 0:20:45.200 --> 0:20:47.960 the game, and maybe, by extension, a new way of 0:20:48.080 --> 0:20:49.000 seeing the world. 0:20:51.359 --> 0:20:52.880 So let's come back to that. 0:20:53.000 --> 0:20:57.080 Amazing Swiss experiment where they made debate bots on Reddit 0:20:57.480 --> 0:21:01.760 that performed way better than humans. I suggest it's possible 0:21:01.800 --> 0:21:05.520 and maybe even likely, that a similar dynamic is going 0:21:05.560 --> 0:21:10.800 to unfold in the world of persuasive argumentation that happened 0:21:10.800 --> 0:21:14.520 in chess and go in video games. If AI agents 0:21:14.960 --> 0:21:19.720 can model the best forms of debate, clear and structured 0:21:20.040 --> 0:21:25.480 and empathetic and rational, then we humans can learn something 0:21:25.960 --> 0:21:30.040 from our artificial cousins. We can try out new moves, 0:21:30.119 --> 0:21:36.000 We can sharpen our skills on digital grinding stones. Imagine 0:21:36.040 --> 0:21:41.680 a future where students practice crafting arguments by debating highly 0:21:41.720 --> 0:21:48.520 skilled AI tutors. Imagine online discussions becoming more useful because 0:21:48.680 --> 0:21:55.480 users have gotten used to high quality exchanges. Imagine politicians 0:21:55.520 --> 0:22:00.000 and journalists and everyday citizens pushed to improve their things 0:22:00.920 --> 0:22:06.399 and better articulate their positions. So, rather than dumbing down conversation, 0:22:06.880 --> 0:22:11.640 the rise of high performing debate bots could nudge public 0:22:11.680 --> 0:22:16.600 discourse toward a new level of reasoned discussion. If that 0:22:16.680 --> 0:22:19.400 turns out to be the case, then we may come 0:22:19.440 --> 0:22:22.160 to see AI not as an enemy but as a 0:22:22.200 --> 0:22:25.840 sparring partner, just like in Chest and Go, and that 0:22:26.359 --> 0:22:45.639 has very different implications. Now, some people will take just 0:22:45.680 --> 0:22:49.280 the opposite position. If AI is so much better at 0:22:49.280 --> 0:22:52.919 debating than we are, won't that cause us to lose 0:22:53.000 --> 0:22:56.280 the skill entirely because there doesn't seem to be a 0:22:56.280 --> 0:22:59.320 point in being a good debater if the computer can 0:22:59.359 --> 0:23:01.760 do it better. But I think this is not an 0:23:01.800 --> 0:23:06.560 issue because even when we extrapolate this era of AI, 0:23:07.240 --> 0:23:10.000 people will still be talking with each other most of 0:23:10.040 --> 0:23:12.879 the time. You'll still be with your family over dinner, 0:23:13.200 --> 0:23:16.160 or with your friends over a coffee, or arguing with 0:23:16.200 --> 0:23:19.320 your neighbor about where the fence goes, or debating with 0:23:19.400 --> 0:23:22.240 a stranger at a town hall or whatever. It's not 0:23:22.359 --> 0:23:25.520 like we are plugging into the matrix through a port 0:23:25.520 --> 0:23:27.520 in the back of our neck, and we're only going 0:23:27.560 --> 0:23:33.480 to be communicating with machines. We are intensely social creatures, 0:23:33.960 --> 0:23:37.280 and the success of our species has been in part 0:23:37.320 --> 0:23:41.280 because of our massive sociability. So I think we will 0:23:41.560 --> 0:23:46.359 debate other humans all the time, and AI is just 0:23:46.440 --> 0:23:49.600 going to train us to be a little bit better 0:23:49.680 --> 0:23:52.520 at it. Now, I don't want to minimize the risks 0:23:52.560 --> 0:23:56.719 we're facing. We need audit tools and authentication systems that 0:23:56.800 --> 0:23:59.960 can verify whether content was written by humans or AIA. 0:24:00.640 --> 0:24:06.320 We need technical solutions like water marking or cryptographic content authentication, 0:24:06.760 --> 0:24:11.720 and providence tracking. This is all essential, but beyond these 0:24:11.760 --> 0:24:17.280 defensive measures, we should also recognize the opportunity because like 0:24:17.320 --> 0:24:22.000 the chests and go engines that reshaped how champion players think, 0:24:22.440 --> 0:24:27.840 debate bots will reshape how we reason and argue and 0:24:28.040 --> 0:24:31.959 understand one another. If we take this on correctly, AI 0:24:32.320 --> 0:24:36.359 might just up our game. It's still early days, but 0:24:36.440 --> 0:24:39.639 my impression so far is that AI is already playing 0:24:39.640 --> 0:24:43.359 this role in many areas, including for example, the arts. 0:24:44.119 --> 0:24:48.160 At least in some ways, it's amplifying human creativity. It's 0:24:48.200 --> 0:24:52.000 like a jazz partner who plays a riff you weren't expecting, 0:24:52.480 --> 0:24:56.159 or a painting mentor who introduces a color that you 0:24:56.240 --> 0:24:59.560 never thought to use. AI forces us out of our 0:24:59.680 --> 0:25:03.240 group in the arts. It shows us that the boundaries 0:25:03.240 --> 0:25:08.000 of our imagination can be stretched by encountering minds, even 0:25:08.040 --> 0:25:13.520 synthetic minds that think differently. These AI systems presumably don't 0:25:13.560 --> 0:25:18.159 have esthetic tastes or emotional longing in the way that 0:25:18.200 --> 0:25:22.119 we do, but they're awesome at doing remixes and trying 0:25:22.200 --> 0:25:25.920 strange new things out, and in this way they teach 0:25:26.040 --> 0:25:31.359 us just how much more flexible and expansive our own 0:25:31.400 --> 0:25:35.920 creativity can be. The takeaway is that AI shakes us 0:25:36.040 --> 0:25:40.480 loose from esthetic grooves that we might never leave on 0:25:40.520 --> 0:25:44.480 our own, and AI is doing exactly that in science 0:25:44.600 --> 0:25:48.399 as well. Just as one example, in material science and 0:25:48.520 --> 0:25:53.080 drug design, AI is proposing molecular structures that humans just 0:25:53.160 --> 0:25:57.200 wouldn't think to try already. It's nudging us to rethink 0:25:57.280 --> 0:26:01.879 what counts as a reasonable chemical design, and we're seeing 0:26:01.880 --> 0:26:06.320 the same thing in AI assisted math proofs. There are 0:26:06.359 --> 0:26:11.680 systems like lean and GPT fueled theorem provers. There are 0:26:11.720 --> 0:26:17.400 these systems that are suggesting lemmas or strategies that mathematicians 0:26:18.080 --> 0:26:20.560 just hadn't thought of. They could have thought of it, 0:26:20.840 --> 0:26:24.720 they just never did. These AI math proofs sometimes come 0:26:24.760 --> 0:26:28.879 out strange or elegant or messy, but in all cases 0:26:29.480 --> 0:26:37.520 they force mathematicians to think differently about structure and possibility. 0:26:37.720 --> 0:26:42.760 So AI can serve as a creativity engine in arts 0:26:42.760 --> 0:26:46.679 and in science, pushing us outside of our intuition driven 0:26:47.160 --> 0:26:50.000 blind spots. And there's something else that AI might be 0:26:50.000 --> 0:26:52.639 able to help us with, which is our personal lives 0:26:52.920 --> 0:26:56.680 and how AI can uncover blind spots there as well. 0:26:57.080 --> 0:26:59.399 If you're a regular listener, you know I've talked on 0:26:59.480 --> 0:27:03.040 many previous episodes about the ways in which we fool 0:27:03.119 --> 0:27:05.000 ourselves because you're. 0:27:04.800 --> 0:27:06.000 Not one thing. 0:27:06.119 --> 0:27:09.560 Instead, you're built of many different drives, or, as I 0:27:09.560 --> 0:27:11.960 wrote my book Incognito, you can think about the brain 0:27:12.359 --> 0:27:15.880 as a team of rivals. So think about the way 0:27:15.920 --> 0:27:20.000 that many people approach fitness. Someone might sign up for 0:27:20.320 --> 0:27:24.199 an expensive year long gym membership because they're determined to 0:27:24.200 --> 0:27:26.520 get in shape, and on paper, that looks like a 0:27:26.520 --> 0:27:29.760 great decision. It's an investment in their health. But at 0:27:29.800 --> 0:27:33.280 the same time, that person keeps telling themselves that they 0:27:33.320 --> 0:27:37.120 just don't have the time to exercise, so months pass, 0:27:37.440 --> 0:27:41.560 the membership goes unused, the rationalization continues. We've all seen 0:27:41.600 --> 0:27:46.280 this sort of thing. An AI personal assistant reviewing their 0:27:46.400 --> 0:27:51.520 spending and scheduling could point out the contradiction. It could say, Hey, 0:27:51.520 --> 0:27:54.199 you know what, You've paid twelve hundred bucks for a 0:27:54.359 --> 0:27:58.000 gym that you rarely visit, but you also spend seven 0:27:58.040 --> 0:28:01.840 hours a week watching streaming shows and an hour every 0:28:01.920 --> 0:28:05.920 day doom scrolling on social media. You say you don't 0:28:05.960 --> 0:28:10.920 have time, but your calendar suggests otherwise. We humans are 0:28:10.960 --> 0:28:13.960 so good at fooling ourselves with our little stories. We 0:28:14.040 --> 0:28:18.240 smooth over inconsistencies without even realizing it. But a good 0:28:18.320 --> 0:28:21.960