WEBVTT - The AI that Might Take My Job (And Yours) 0:00:15.356 --> 0:00:15.796 Pushkin. 0:00:22.036 --> 0:00:24.476 After every interview we do for the show, we upload 0:00:24.516 --> 0:00:28.236 the audio to a piece of software called descript. Descript 0:00:28.316 --> 0:00:31.476 turns the audio into a transcript, and then I can 0:00:31.756 --> 0:00:35.556 edit that transcript, cut out the boring parts, move sections around, 0:00:35.916 --> 0:00:39.876 and when I do that, descript edits the underlying audio 0:00:40.036 --> 0:00:45.436 to match. As software, Descript is pretty janky, buggy. It's 0:00:45.476 --> 0:00:47.996 constantly changing in ways that can make it hard to use, 0:00:48.356 --> 0:00:51.556 and sometimes it just blows stuff up. But we use 0:00:51.596 --> 0:00:55.716 it anyway because Descript is an incredible. 0:00:54.956 --> 0:00:56.676 Advance over what came before. 0:00:57.396 --> 0:01:01.956 Before descript audio software represented audio files not as words 0:01:01.996 --> 0:01:05.436 that you can read and edit, but as waveforms, as 0:01:05.676 --> 0:01:07.676 squiggly lines presented. 0:01:07.316 --> 0:01:09.876 On a timeline. So when does the script came along? 0:01:09.996 --> 0:01:12.396 Being able to edit audio by editing words on a 0:01:12.436 --> 0:01:15.956 screen was this huge advance, and it was an advance 0:01:15.956 --> 0:01:21.876 made possible by artificial intelligence. Eventually, Descript expanded to allow 0:01:21.916 --> 0:01:25.036 people to edit not just audio but also video, and 0:01:25.116 --> 0:01:28.636 last fall, open AI, the company that makes chat GPT, 0:01:29.196 --> 0:01:30.956 led a fifty million dollar. 0:01:30.796 --> 0:01:32.396 Investment round in Descript. 0:01:32.876 --> 0:01:35.636 It's a sign that Descript is moving out to the 0:01:35.676 --> 0:01:40.036 new AI frontier, the frontier of generative AI. AI that 0:01:40.076 --> 0:01:44.076 creates words and pictures. This is of immediate interest to me, 0:01:44.796 --> 0:01:47.836 as in is AI gonna help me do my job? 0:01:48.236 --> 0:01:50.756 Is AI gonna do my job? 0:01:51.276 --> 0:01:54.116 But there is also a bigger question here what is 0:01:54.196 --> 0:01:55.076 AI going to mean? 0:01:55.236 --> 0:01:58.956 More broadly for people whose jobs involve writing things and 0:01:58.996 --> 0:02:02.116 creating visuals, which is to say, what is AI going 0:02:02.196 --> 0:02:11.316 to mean for almost all white collar workers. I'm Jacob 0:02:11.356 --> 0:02:13.676 Goldstein and this is What's Your Problem, a show about 0:02:13.716 --> 0:02:17.556 people trying to make technological progress. My guest today is 0:02:17.636 --> 0:02:21.516 Andrew Mason, founder and CEO of descript or maybe it's 0:02:21.836 --> 0:02:25.156 descript by the way, I've always said descript and I'm 0:02:25.156 --> 0:02:28.356 pretty sure that's wrong, right, it's a descript like dtour. 0:02:28.436 --> 0:02:30.196 Weird noncommittal on the issue. 0:02:30.316 --> 0:02:31.756 Let's do the subjective version. 0:02:31.756 --> 0:02:33.676 You're just one man. How do you say the name 0:02:33.716 --> 0:02:34.356 of your company? 0:02:34.676 --> 0:02:37.836 Yeah, I've kind of cultivated the ability to flip between 0:02:37.876 --> 0:02:38.876 them as I speak. 0:02:39.716 --> 0:02:40.396 You're killing me. 0:02:40.556 --> 0:02:42.036 The world still needs a little mystery. 0:02:42.196 --> 0:02:44.436 Okay, how about this? Say your name and your job. 0:02:45.156 --> 0:02:48.476 My name is Andrew Mason. I work at descript. That's 0:02:49.036 --> 0:02:53.316 dscript dscript. 0:02:53.556 --> 0:02:54.476 Well played. 0:02:55.836 --> 0:02:58.516 Earlier in his career. Andrew Mason was the co founder 0:02:58.556 --> 0:03:01.596 of groupunk. He took the company public and then got 0:03:01.636 --> 0:03:05.076 fired after its stock fell by something like seventy five percent. 0:03:05.596 --> 0:03:08.676 After that, he started a company called Dtour or maybe 0:03:08.676 --> 0:03:12.236 it's I don't know. The company made these highly produced 0:03:12.316 --> 0:03:15.356 audio walking tours that you could listen to on your phone. 0:03:15.396 --> 0:03:18.276 In that job, Andrew saw the challenges of working with 0:03:18.356 --> 0:03:20.236 the old waveform. 0:03:19.716 --> 0:03:21.276 Based audio editing software. 0:03:21.596 --> 0:03:25.116 At the same time, AI generated transcripts were getting better 0:03:25.276 --> 0:03:28.476 and cheaper, and new technology was making it possible to 0:03:28.636 --> 0:03:32.476 automatically match a transcript to an audio file. Andrew looked 0:03:32.476 --> 0:03:35.076 at those two developments and thought, we should make an 0:03:35.116 --> 0:03:38.396 audio editor that works like a word processor, which he 0:03:38.516 --> 0:03:41.036 admits was a distraction from what he was supposed to 0:03:41.036 --> 0:03:43.276 be doing, which was making walking tours. 0:03:44.916 --> 0:03:47.396 If I'm being honest, it was a bit of an indulgence. 0:03:47.636 --> 0:03:50.436 It just felt like an incredibly cool problem to work on. 0:03:51.076 --> 0:03:53.956 I went to school for music technology and worked in 0:03:53.996 --> 0:03:58.156 a recording studio after I graduated, and just always loved tools, 0:03:58.236 --> 0:04:02.436 and audio visual tools in particular. It was just so 0:04:02.756 --> 0:04:05.076 fun to start thinking about this puzzle. 0:04:05.636 --> 0:04:06.316 Uh huh. 0:04:06.356 --> 0:04:10.676 So we told ourselves it was kind of way of diversifying, 0:04:10.716 --> 0:04:13.636 but that's just like a ridiculous way for a product 0:04:13.676 --> 0:04:15.796 that at like a startup that just hasn't even found 0:04:16.236 --> 0:04:19.716 product market fit in their core product to be thinking 0:04:19.796 --> 0:04:22.196 about the world. You know, all of the advice textbooks 0:04:22.236 --> 0:04:24.836 will tell you not to do that, and it's probably 0:04:24.876 --> 0:04:28.956 generally good advice, but it was just it was just irresistible. 0:04:28.596 --> 0:04:28.796 You know. 0:04:28.876 --> 0:04:31.516 So I am a fan of descript. I started using 0:04:31.596 --> 0:04:35.036 it around when it came out several years ago. Certainly 0:04:35.636 --> 0:04:38.996 I think it's great. It is kind of janky, and 0:04:39.076 --> 0:04:42.836 it's always kind of janky, right, And my guess is 0:04:43.396 --> 0:04:46.556 jankie meaning like a little bit unstable, things don't quite work. 0:04:46.596 --> 0:04:48.156 It's always telling you to restart. 0:04:48.676 --> 0:04:50.356 By the way, if I'm not sure, if I have 0:04:50.516 --> 0:04:52.476 your side, so I may ask you to send me 0:04:52.596 --> 0:04:54.676 this entire portion of the interview so I can share 0:04:54.676 --> 0:04:55.236 it with the team. 0:04:55.756 --> 0:04:58.156 So the thing is, like, I wonder, why is it 0:04:58.156 --> 0:05:00.436 always kind of janky? Why is it never just like 0:05:00.476 --> 0:05:02.916 stable and it works? And my guess is it's because 0:05:02.916 --> 0:05:05.436 you're pushing forward really fast, right, You're trying to make 0:05:05.476 --> 0:05:08.476 it better and better and better, and presumably there is 0:05:08.516 --> 0:05:11.236 some trace off, right, like the faster you try and 0:05:11.276 --> 0:05:14.796 push it forward, the more janky it's gonna be. You could, 0:05:14.836 --> 0:05:17.236 I'm sure, just perfect the way it was four years ago, 0:05:17.316 --> 0:05:18.836 but then it would never get better, but it would 0:05:18.836 --> 0:05:21.076 be stable, right, And so this is like a big 0:05:21.116 --> 0:05:25.236 whatever startup founder type question, like, is that some balance 0:05:25.236 --> 0:05:27.276 you're always trying to figure out how fast do we 0:05:27.756 --> 0:05:29.756 iterate versus how much do we try and make it 0:05:29.876 --> 0:05:31.036 just stable and work. 0:05:31.876 --> 0:05:36.796 Yeah, that's an astute observation, not the fact that it's janky. 0:05:36.996 --> 0:05:38.596 That doesn't take a genius. 0:05:38.316 --> 0:05:41.236 Respectful but respectfully as a fan. 0:05:41.076 --> 0:05:41.916 I'm telling you it's. 0:05:43.396 --> 0:05:45.716 But I think like your your attempt to make sense 0:05:45.756 --> 0:05:48.116 out of it, I think like a good story to 0:05:48.196 --> 0:05:51.116 tell here is maybe like going back to the to 0:05:51.196 --> 0:05:56.116 the very beginning of Descript. So when we became Descript, 0:05:56.476 --> 0:06:01.236 we sold off Detour to Bows and we decided to 0:06:01.356 --> 0:06:05.796 just focus on building out this media word processor thing. 0:06:06.996 --> 0:06:10.956 And some of the public radio producers who had worked 0:06:10.956 --> 0:06:14.916 at Detour went on back into public radio and they 0:06:14.996 --> 0:06:21.596 became some of the earliest customers of dscript. And what 0:06:21.756 --> 0:06:27.316 we found was that they pushed it so much farther 0:06:27.876 --> 0:06:29.756 than we were ready for. 0:06:30.236 --> 0:06:32.876 Ah, so quickly, what do you mean by that? Like, 0:06:32.916 --> 0:06:34.036 what is an example of that. 0:06:34.316 --> 0:06:37.076 Yeah, I mean specifically in the case of like some 0:06:37.116 --> 0:06:40.836 of these shows, it means putting together three to five 0:06:40.916 --> 0:06:45.756 hour cuts of tape from many different files, with lots 0:06:45.876 --> 0:06:49.716 like tons and tons of edits and notes mixed into 0:06:49.716 --> 0:06:54.916 the edits, and just like stuff that we hadn't pressure 0:06:54.956 --> 0:06:58.396 tested from a performance just giant. 0:06:58.476 --> 0:07:01.276 The files are really big, right, Like a three hour 0:07:01.636 --> 0:07:04.276 audio file is actually a giant file, right, And if 0:07:04.276 --> 0:07:05.836 you're stacking up a bunch of those, so you have 0:07:05.836 --> 0:07:08.556 all these giant files and you're making tons of cuts, 0:07:08.636 --> 0:07:14.196 that's just like computationally intensive, that kind of thing storage intensive. 0:07:15.036 --> 0:07:17.396 Yeah, it was just something that we hadn't that we 0:07:17.436 --> 0:07:20.876 hadn't optimized for. It's it's an eminently solvable problem, but 0:07:20.916 --> 0:07:25.396 it was something that in the earliest versions we hadn't done. 0:07:26.036 --> 0:07:28.676 And so that is kind of in many ways been 0:07:28.796 --> 0:07:32.316 the story of descript up to this point, where there's 0:07:32.396 --> 0:07:35.476 there's been that element of it, and there were kind 0:07:35.556 --> 0:07:39.156 of realities of needing to make quick progress that we 0:07:39.196 --> 0:07:46.636 had to balance against stability and what we had for 0:07:46.716 --> 0:07:50.436 our customers in terms of like the core product idea 0:07:50.996 --> 0:07:53.756 of being able to edit by text was still for 0:07:53.916 --> 0:07:58.476 them so much better than the alternative that there was 0:07:59.276 --> 0:08:04.516 just a tolerance of the stability issues that honestly made 0:08:04.596 --> 0:08:06.596 us sick to our stomachs that we had to put 0:08:06.636 --> 0:08:10.036 people through. And it's not like wegn but it was 0:08:10.116 --> 0:08:13.596 like we had to make trade offs there. So all 0:08:13.636 --> 0:08:16.516 of this pushing kind of culminated with this release of 0:08:16.556 --> 0:08:20.316 a pretty major overhaul that we did at the end 0:08:20.316 --> 0:08:25.996 of last year and since then, since last November and 0:08:26.556 --> 0:08:29.996 really like through the first half of this year, is 0:08:30.036 --> 0:08:32.036 when we think we start to get to a good place. 0:08:32.796 --> 0:08:35.396 Our goal is that if we're having this conversation, like 0:08:35.436 --> 0:08:38.516 we're not going to be having the same conversation in 0:08:38.716 --> 0:08:41.436 say July, for sure at the very latest, like the 0:08:41.436 --> 0:08:44.196 conversation we'll be having with someone like you will be Wow, 0:08:44.236 --> 0:08:46.756 it's gotten. It's like not an issue anymore. 0:08:46.876 --> 0:08:50.516 So you say all that, but also, you just got 0:08:50.516 --> 0:08:53.916 this big investment from open Ai. You got a thing 0:08:54.116 --> 0:08:57.756 on descript that says sign up to try GPT four 0:08:57.876 --> 0:09:00.836 with Descript, which I just signed up for and I'm 0:09:00.916 --> 0:09:04.636 very curious about That doesn't sound like, oh, we've arrived 0:09:04.676 --> 0:09:06.316 and now we've got our product and we've just got 0:09:06.316 --> 0:09:08.316 a hone it that sounds like there's this whole giant 0:09:08.396 --> 0:09:10.356 new universe of things you were about to try and 0:09:10.356 --> 0:09:10.916 figure out. 0:09:11.556 --> 0:09:13.916 That's true, And that's the funny thing about all of 0:09:13.916 --> 0:09:16.996 this are is that at the same time that we're 0:09:16.996 --> 0:09:21.076 turning to focus on quality, it's a moment where generative 0:09:21.076 --> 0:09:25.076 AI has arrived at a scale and with a force 0:09:25.316 --> 0:09:28.156 that no one really saw coming this quickly. 0:09:28.676 --> 0:09:31.396 So so okay, I know from the beginning Descript was 0:09:31.396 --> 0:09:35.076 built on top of AI, you know, the technology for 0:09:35.116 --> 0:09:39.516 transcription for matching audio to text, but was Descript itself 0:09:39.596 --> 0:09:40.716 an AI company. 0:09:41.516 --> 0:09:44.236 So we had some really smart people on the team 0:09:44.596 --> 0:09:47.636 in UH with machine learning experience, but I wouldn't say 0:09:47.636 --> 0:09:50.196 in the early days we were like a company that 0:09:50.676 --> 0:09:53.956 was with anybody that was doing like original AI research 0:09:54.036 --> 0:09:59.436 or anything like that. We saw that as a gap 0:09:59.836 --> 0:10:05.036 that we wanted to solve. And so I forget exactly 0:10:05.636 --> 0:10:07.956 what year it was, it was maybe about four years 0:10:07.956 --> 0:10:12.916 ago we saw this company called Liarbird. It was a 0:10:12.916 --> 0:10:16.436 a company out of y Combinator with some really smart 0:10:17.076 --> 0:10:21.036 PhD candidates. They had built model that would build a 0:10:21.076 --> 0:10:23.836 clone of your voice based on I think about three 0:10:24.076 --> 0:10:27.476 minutes or five minutes of training data. Of just talking 0:10:27.556 --> 0:10:27.796 to it. 0:10:27.876 --> 0:10:31.756 Let me just say, I know Liarbird is spelled l yri, 0:10:32.596 --> 0:10:34.476 but I assume they're aware. 0:10:34.236 --> 0:10:35.036 Of the hominem. 0:10:35.156 --> 0:10:37.036 Right, this is a thing that is cloning your voice 0:10:37.076 --> 0:10:39.636 so that you can make it sound like you're talking 0:10:39.756 --> 0:10:42.876 even if you're not talking. And the company is called Liarbird, 0:10:43.276 --> 0:10:47.276 and this is a somewhat fraud thing, right, Like, I 0:10:47.276 --> 0:10:49.156 feel like they're throwing it in my face that this 0:10:49.196 --> 0:10:53.636 is a sketchy product that they're developing. 0:10:54.436 --> 0:10:55.716 Did it cross your mind? 0:10:56.196 --> 0:10:59.276 Did it cross my mind? Is like the ethical quandary 0:10:59.276 --> 0:11:01.956 that we were getting into, or like the branding implications 0:11:01.996 --> 0:11:02.596 of the name. 0:11:03.636 --> 0:11:04.836 More the ethical quandary. 0:11:05.396 --> 0:11:10.236 Yeah, the ethical quandary absolutely entered our mind. And our 0:11:10.276 --> 0:11:12.756 point of view on that, and has been our point 0:11:12.796 --> 0:11:16.316 of view on these things in general, has been that 0:11:17.356 --> 0:11:22.196 we don't want to be like out there paving the 0:11:22.236 --> 0:11:25.756 way for any new paths to the apocalypse, so to speak. 0:11:27.756 --> 0:11:31.996 We actually think, like have always felt like not really 0:11:32.036 --> 0:11:36.236 sure how society was going to put the brakes on 0:11:36.236 --> 0:11:38.116 this sort of thing. We just knew that we didn't 0:11:38.116 --> 0:11:40.156 want to be part of it, and we tried to 0:11:40.156 --> 0:11:42.996 put guardrails in place on our product. That would make 0:11:43.036 --> 0:11:46.316 it easy to stay off the slippery slope. So in 0:11:46.356 --> 0:11:50.996 the case of Lyyerbird, which once we bought them, we 0:11:51.276 --> 0:11:53.596 integrated their technology and released it as something that we 0:11:53.636 --> 0:11:55.996 call overdub. It's a way that you can clone your voice. 0:11:56.636 --> 0:11:59.556 We require you to authenticate that it's actually you, and 0:11:59.596 --> 0:12:02.756 we only let you clone your own voice, and that's 0:12:02.796 --> 0:12:05.476 worked really well. We're now in a world where there's 0:12:05.476 --> 0:12:08.876 other people that have similar models and they're not putting 0:12:08.876 --> 0:12:11.756 those protections in place. And the use case that we've 0:12:11.756 --> 0:12:14.876 always been the most excited about is making it possible 0:12:14.876 --> 0:12:18.916 to edit your natural recordings, so going in and changing 0:12:18.956 --> 0:12:22.236 an individual word, and we've built some special stuff that 0:12:22.316 --> 0:12:24.676 will kind of listen to the audio on either sides 0:12:24.716 --> 0:12:27.396 and make sure that it blends in. From an intonation perspective, 0:12:27.996 --> 0:12:30.196 we started with the ability to delete stuff and move 0:12:30.236 --> 0:12:32.796 stuff around. Now you can just type and really make 0:12:32.836 --> 0:12:34.156 it feel like it's a word processor. 0:12:34.316 --> 0:12:39.316 Presumably the better you get, the better the technology you 0:12:39.436 --> 0:12:43.156 use to Clona Voice gets, the more words it can do. Right, 0:12:43.196 --> 0:12:46.196 I mean, every week, for what's your problem? I write 0:12:46.196 --> 0:12:50.676 a little introduction and then I read it. But presumably 0:12:50.716 --> 0:12:53.236 at some point overdub will be good enough that no 0:12:53.276 --> 0:12:55.556 one knows will know whether it's me reading it or 0:12:55.596 --> 0:12:56.836 I'm just typing it right. 0:12:57.316 --> 0:13:00.196 We have a new version of overdub that will release 0:13:01.116 --> 0:13:04.156 in the next couple of months, and it's the first 0:13:04.196 --> 0:13:08.236 time that I've heard my own voice doing a narration 0:13:08.356 --> 0:13:12.076 of something that made me say, like, this sounds so 0:13:12.236 --> 0:13:16.116 much like me in a way that it's not distracting 0:13:16.196 --> 0:13:18.156 or the AI does not get in the way. 0:13:18.556 --> 0:13:23.236 Can I try that new version now, like, not this minute, 0:13:23.276 --> 0:13:24.156 but like for the show? 0:13:24.516 --> 0:13:25.196 Yeah, for the show. 0:13:26.196 --> 0:13:27.956 I bet we could find a way to do it. 0:13:27.956 --> 0:13:30.396 It's just so you could hear it and stuff. 0:13:30.796 --> 0:13:33.116 There's a universe where I say, at this moment in 0:13:33.156 --> 0:13:37.236 the show, guess what today? That voice me reading the 0:13:37.236 --> 0:13:39.756 intro at the top of the show that was overdubbed. 0:13:39.756 --> 0:13:40.676 It wasn't really made. 0:13:40.916 --> 0:13:45.756 Yeah, we tried overdubb for the voice doing the intro 0:13:45.836 --> 0:13:47.196 at the top of the show. 0:13:47.516 --> 0:13:49.316 And we decided it wasn't quite. 0:13:49.076 --> 0:13:51.676 Good enough, but we decided it would work for this 0:13:51.796 --> 0:13:52.636 part of the show. 0:13:53.156 --> 0:13:57.036 What you're hearing right now, it's not really me. It's overdubbed. 0:13:58.116 --> 0:14:01.756 In a minute, what overdubb and chat GPT and generative 0:14:01.796 --> 0:14:04.116 AI will mean for descript and for the. 0:14:04.076 --> 0:14:06.396 World and also for me. 0:14:12.436 --> 0:14:16.316 Now back to the show, descript is expanding from podcasts 0:14:16.316 --> 0:14:18.716 to video, and it just took a big investment from 0:14:18.836 --> 0:14:22.636 open Ai, the company that makes chat GPT, and also 0:14:22.676 --> 0:14:26.516 this system called Dolly that uses AI to generate images. 0:14:26.916 --> 0:14:30.116 So Descript is clearly pointing toward a future where it's 0:14:30.156 --> 0:14:33.276 going to be software for creating AI generated or at 0:14:33.356 --> 0:14:34.316 least AI. 0:14:34.236 --> 0:14:35.956 Enhanced audio and video. 0:14:36.556 --> 0:14:39.356 And I asked Andrew, what does that future look like? 0:14:39.796 --> 0:14:42.596 How is generative AI going to work in descript? 0:14:43.236 --> 0:14:45.676 I don't think we know entirely yet. In a lot 0:14:45.676 --> 0:14:47.836 of ways, it feels to me like you're letting this 0:14:48.116 --> 0:14:51.876 alien into into your app. You're just giving it the 0:14:51.956 --> 0:14:56.716 keys and then the interfaces. How do you find how 0:14:56.756 --> 0:15:00.916 do you find a way to kind of give the 0:15:00.956 --> 0:15:04.556 aliens some buttons in tier UI, give them the ability 0:15:04.636 --> 0:15:06.796 to press the buttons, and then how do you talk 0:15:06.836 --> 0:15:07.476 to the alien? 0:15:07.796 --> 0:15:08.796 What do you mean? Like? 0:15:08.836 --> 0:15:13.396 That is a striking metaphor a little scarier right. It 0:15:13.476 --> 0:15:17.036 suggests a certain level of uncertainty and potential downside. It's 0:15:17.076 --> 0:15:18.716 not like, oh, this is great, this is going to 0:15:18.716 --> 0:15:20.556 solve a problem like, why do you say it's like 0:15:20.636 --> 0:15:21.716 letting an alien. 0:15:21.396 --> 0:15:24.116 In as opposed to letting a human in. 0:15:25.276 --> 0:15:27.836 It's a really interesting choice of words. Tell me more 0:15:27.836 --> 0:15:28.276 about it. 0:15:30.996 --> 0:15:35.436 So let's start by just saying, like, very specifically, what 0:15:35.476 --> 0:15:40.276 I mean. I think, when implemented, well, what this will 0:15:40.276 --> 0:15:43.596 feel like is as if you had a co editor 0:15:44.236 --> 0:15:47.196 in a document with you, in our case, in a 0:15:47.276 --> 0:15:52.436 video or a podcast that you're working on that is smart, 0:15:52.596 --> 0:15:54.956 knows how to do everything, definitely knows how to do 0:15:54.996 --> 0:15:59.316 the tedios busy work, and you can kind of kind 0:15:59.356 --> 0:16:03.156 of guide or direct through giving these tasks. You know, 0:16:03.716 --> 0:16:07.596 it's almost like it's the production assistant or something like that, 0:16:07.636 --> 0:16:10.716 and you're the director and you're able to just guide 0:16:10.756 --> 0:16:13.076 it and give it feedback on how it's doing and 0:16:13.076 --> 0:16:15.196 what it's doing well and what it's not doing well. 0:16:15.676 --> 0:16:18.076 There's a version of it where it's like we've gotten 0:16:18.156 --> 0:16:21.716 used to the graphical user interface, right, We've been trained 0:16:21.756 --> 0:16:24.796 since the Magintosh computer in the mid nineteen eighties that 0:16:24.876 --> 0:16:26.836 the way you interact with a computer is like there's 0:16:26.876 --> 0:16:28.996 little pictures and little folders and you point. 0:16:28.756 --> 0:16:30.676 And click one way or another, right, and. 0:16:31.276 --> 0:16:35.436 One possibility here is the new standard interface is chat. 0:16:35.476 --> 0:16:38.996 You just type in like whatever, please trim all the 0:16:39.116 --> 0:16:42.436 ums from this file, or even please turn this thirty 0:16:42.476 --> 0:16:44.836 minute interview into a twenty minute interview, and the way 0:16:44.836 --> 0:16:47.596 that makes it most interesting, right, and you just type 0:16:47.596 --> 0:16:48.516 that in and it happens. 0:16:48.556 --> 0:16:50.396 I mean that's a version of what I hear you 0:16:50.396 --> 0:16:50.916 saying there. 0:16:50.956 --> 0:16:53.316 I think some people believe that that chat or a 0:16:53.396 --> 0:16:57.796 text field will become the primary interface for making things. 0:16:58.476 --> 0:17:01.156 I think of it more as like it's the primary 0:17:01.196 --> 0:17:04.276 interface for interacting with the alien, and then you and 0:17:04.316 --> 0:17:07.396 the alien are still going to be working, like have 0:17:07.556 --> 0:17:11.356 other buttons that they can press. You still, sometimes you 0:17:11.436 --> 0:17:13.556 just want to take the thing in your hands and 0:17:13.556 --> 0:17:14.236 do it yourself. 0:17:14.796 --> 0:17:17.316 The alien metaphor, I mean there's a real like do 0:17:17.356 --> 0:17:21.316 we welcome our alien overlord's question? When you choose that metaphor, 0:17:21.396 --> 0:17:22.196 it makes. 0:17:21.996 --> 0:17:24.356 Me I mean, maybe it feels that way. 0:17:24.396 --> 0:17:26.316 It doesn't it it doesn't make me feel better. 0:17:26.396 --> 0:17:29.996 I'll say that I think it feels the way that 0:17:30.036 --> 0:17:33.916 an alien arrival would probably feel, where you know, maybe 0:17:33.916 --> 0:17:37.476 you shake its hand and immediately it has something in 0:17:37.516 --> 0:17:41.476 its skin that cures your cancer, and you feel hopeful, 0:17:42.996 --> 0:17:45.516 but you also want to know what they're up. 0:17:45.436 --> 0:17:48.876 To and yeah, and cure your cancer is definitely the 0:17:48.916 --> 0:17:49.636 happy version. 0:17:49.676 --> 0:17:51.476 Not usually in the alien movie. 0:17:51.276 --> 0:17:53.236 What happens, but I guess that could happen. 0:17:53.396 --> 0:17:55.356 Well, there's the good there's a good part, right, But 0:17:55.396 --> 0:17:57.636 you never really know, I think is the point. And 0:17:57.956 --> 0:18:00.516 I think we're all living in this kind of like 0:18:00.596 --> 0:18:03.916 pushing forward in this mystery, kind of kind of stuck 0:18:03.956 --> 0:18:05.876 between awe and terror. 0:18:06.276 --> 0:18:10.156 You sound more ambivalent than I might have thought. Why 0:18:10.196 --> 0:18:12.836 is that because you just took a giant investment from 0:18:12.876 --> 0:18:13.356 open AI. 0:18:14.876 --> 0:18:17.516 I think like at moments like this, you have a 0:18:17.596 --> 0:18:23.436 choice between either renunciation and just like stopping and out 0:18:23.476 --> 0:18:27.236 of from a place of fear. Which maybe that's right, 0:18:27.356 --> 0:18:32.836 you know, maybe fulfillment and happiness everything we have for 0:18:32.876 --> 0:18:36.796 that is is already here, and we should focus our 0:18:37.076 --> 0:18:39.756 energies on making peace with our inevitable death. 0:18:41.356 --> 0:18:42.676 In any case, we should do that. 0:18:42.836 --> 0:18:47.436 But go on the other way to think of it 0:18:47.516 --> 0:18:52.076 is to just forge ahead and realize that the potential 0:18:52.116 --> 0:18:54.356 of what's on the other end of this might make 0:18:54.476 --> 0:18:59.196 us feel in retrospect like we were just in the 0:18:59.236 --> 0:19:05.476 earliest possible innings of our of the human experiment. So 0:19:06.756 --> 0:19:09.396 you know, I feel like we're all going to die 0:19:09.436 --> 0:19:12.396 one way or another, might as well forge ahead. It's 0:19:12.396 --> 0:19:15.756 not ambivalence, but it's more just being clear eyed about 0:19:15.796 --> 0:19:18.716 the fact that not trying to pretend that there's parts 0:19:18.716 --> 0:19:20.156 of it that don't seem scary. 0:19:21.236 --> 0:19:23.996 I mean, one of the things that's really striking to 0:19:24.156 --> 0:19:29.756 me with AI, and that seems quite different from other 0:19:30.476 --> 0:19:34.636 technologies in the past, is the people who are working 0:19:34.676 --> 0:19:38.036 on it, the people who really understand it, seem more 0:19:38.076 --> 0:19:40.836 scared than everybody else. 0:19:41.516 --> 0:19:44.716 I'm not a first time founder. I went through the 0:19:44.796 --> 0:19:50.396 experience of being a young person building building group on 0:19:51.356 --> 0:19:54.396 telling myself a story about how it was going to 0:19:54.876 --> 0:19:59.156 revolutionize local commerce and all the good stuff, and it 0:19:59.276 --> 0:20:01.116 just didn't turn out that way. And I think we've 0:20:01.156 --> 0:20:06.156 seen a generation of tech companies that just like didn't 0:20:06.556 --> 0:20:10.756 turn out the way that the the super Rose colored 0:20:10.756 --> 0:20:15.356 Glasses mission statement would have suggested. And I think we're 0:20:15.396 --> 0:20:19.316 just trying to be we just have that experience, that 0:20:19.356 --> 0:20:24.196 recent experience at top of mind, and are trying to 0:20:24.316 --> 0:20:27.836 think about it in a way that has guardrails around 0:20:27.996 --> 0:20:30.356 around repeating that history and just make sure we're really 0:20:30.396 --> 0:20:33.476 proud of what we build. Does that make sense? 0:20:33.596 --> 0:20:34.236 It makes sense. 0:20:34.596 --> 0:20:36.196 Am I Am I going to regret saying all this? 0:20:37.156 --> 0:20:37.836 I don't think so. 0:20:37.996 --> 0:20:40.676 You haven't said anything like incriminating as far as I 0:20:40.676 --> 0:20:42.996 can tell. You know, I heard somebody saying the other day, like, 0:20:43.076 --> 0:20:46.076 it's an interesting question to ask somebody like, what was 0:20:46.116 --> 0:20:49.996 the first thing you asked GPT chet GPT to do? 0:20:50.596 --> 0:20:53.156 And the first thing I asked chet GPT to do 0:20:53.596 --> 0:20:57.236 was write an episode of Planet Money podcast I used 0:20:57.236 --> 0:20:59.236 to host, of which there are you know, a thousand 0:20:59.276 --> 0:21:01.956 transcripts on the internet. Write an episode of Planet Money 0:21:01.996 --> 0:21:04.276 about whether the FED is going to raise interest rates 0:21:04.276 --> 0:21:06.756 by twenty five basis points or leave them unchanged, right, 0:21:07.396 --> 0:21:10.516 And it wrote something that was pretty good, like not 0:21:10.596 --> 0:21:13.716 a whole show, it's not there now, but at the 0:21:13.796 --> 0:21:16.556 rate of current improvement, you could definitely imagine it writing 0:21:16.596 --> 0:21:20.716 that episode pretty well in whatever a year or two 0:21:20.836 --> 0:21:22.956 years or some amount of time when I will still 0:21:23.036 --> 0:21:26.716 want to be gainfully employed. And like I do wonder 0:21:26.756 --> 0:21:29.796 on this one, is there a day slash? How far 0:21:29.916 --> 0:21:30.716 are we from. 0:21:30.556 --> 0:21:34.596 The day when generative a I can just make a 0:21:34.636 --> 0:21:36.276