WEBVTT - An AI Debate with Marcus Hutchins 0:00:19.612 --> 0:00:21.572 All right. What I'm going to do today is we're 0:00:21.572 --> 0:00:24.652 going to have a debate between myself and Marcus Hutchins. 0:00:24.892 --> 0:00:26.932 If you don't know Marcus, he is probably the most 0:00:26.932 --> 0:00:30.932 famous malware researcher in the world. He's been on the 0:00:30.932 --> 0:00:35.652 front page of wired, and he has a dramatically different 0:00:35.652 --> 0:00:38.252 view of AI than I do. So I am in 0:00:38.292 --> 0:00:41.012 sort of the pro AI camp and believing that it 0:00:41.012 --> 0:00:46.412 is net positive and that it is extremely useful today, 0:00:46.652 --> 0:00:50.251 and also getting more useful and most importantly, that it's 0:00:50.251 --> 0:00:52.852 going to start affecting jobs in a very serious way. 0:00:53.292 --> 0:00:58.092 Marcus has called it basically, autocomplete does not believe that 0:00:58.092 --> 0:01:03.212 it's intelligence, and believes that the impact is dramatically overplayed, 0:01:03.212 --> 0:01:06.412 mostly by people trying to scam people and make money. 0:01:06.892 --> 0:01:11.132 So again, we are on pretty much opposite sides of this, 0:01:11.572 --> 0:01:15.252 but we are also both security people and, you know, 0:01:15.492 --> 0:01:19.422 fairly well known in the security space, and we've both 0:01:19.462 --> 0:01:23.661 done a lot and accomplished a lot. And this is 0:01:23.661 --> 0:01:26.581 really showing that you can have two people who are 0:01:26.581 --> 0:01:29.942 in kind of the same space, can have dramatically different opinions. 0:01:30.262 --> 0:01:33.101 And what turned out to be fantastic here, because Marcus 0:01:33.102 --> 0:01:36.062 and I have been friends for some amount of time now, 0:01:36.581 --> 0:01:39.222 and we had this debate previously in text and it 0:01:39.222 --> 0:01:41.381 was fine in text. I was a little worried that 0:01:41.862 --> 0:01:44.461 we're both a bit snarky online, and I was worried 0:01:44.462 --> 0:01:46.542 that we were going to be snarky to each other 0:01:46.542 --> 0:01:49.262 in this debate that turned out not to happen. It 0:01:49.262 --> 0:01:52.222 was a very positive experience, and I think we made 0:01:52.222 --> 0:01:55.102 our points. Well, not only that, but I think I 0:01:55.102 --> 0:01:57.181 made a number of points that he really enjoyed. He 0:01:57.182 --> 0:01:59.822 made a number of points that I really enjoyed. I 0:01:59.822 --> 0:02:01.981 feel like we both learned something from this, and we've 0:02:01.982 --> 0:02:04.822 already committed to doing a follow up at some point 0:02:04.822 --> 0:02:08.822 in the future, but this one is hopefully going to 0:02:08.822 --> 0:02:11.542 be useful to anybody who is kind of agnostic or 0:02:11.542 --> 0:02:15.512 really just kind of wants to hear strong points being 0:02:15.512 --> 0:02:18.672 made on both sides of the conversation and also in 0:02:18.672 --> 0:02:21.232 a civil way. So that's what we're going to do here. 0:02:21.512 --> 0:02:24.032 I do apologize a little bit for the audio quality. 0:02:24.072 --> 0:02:26.352 We actually did this on zoom, and we should have 0:02:26.352 --> 0:02:29.791 done it on Riverside because you can't actually do 4K 0:02:30.392 --> 0:02:34.112 in zoom yet. So anyway, the next one is going 0:02:34.112 --> 0:02:37.352 to be higher quality. But um, let's jump into it. 0:02:37.352 --> 0:02:40.792 This is a debate between myself and Marcus Hutchins on AI. 0:02:41.151 --> 0:02:43.592 I think it's wrong to call what is happening now 0:02:43.591 --> 0:02:48.792 with AI hype and to tell people not to worry. Um, 0:02:48.912 --> 0:02:51.312 because I think it is a significant threat to people's 0:02:51.312 --> 0:02:58.112 jobs and, um, overall, just like a massive thing, uh, in, 0:02:58.112 --> 0:02:59.872 in much more of a way than crypto was, I 0:02:59.871 --> 0:03:03.112 think crypto was a giant. Nothing. Um, I wasn't excited 0:03:03.112 --> 0:03:05.192 about crypto at the time. I didn't see why why 0:03:05.192 --> 0:03:08.232 it was like everyone was taking it so seriously. But 0:03:08.232 --> 0:03:11.552 I see this as very different. Um, and so what? 0:03:11.552 --> 0:03:14.722 I see people like yourself or someone like another person. 0:03:14.722 --> 0:03:17.362 I have a lot of respect for, uh, Christopher Hoff, 0:03:17.962 --> 0:03:20.841 who I've been friends with for a very long time, actually, 0:03:20.882 --> 0:03:24.522 for him, like 15 years or something. 20 years. And 0:03:24.522 --> 0:03:28.442 he's frequently putting out similar stuff to you, whereas like, look, 0:03:28.442 --> 0:03:30.522 what are you guys talking about? This thing is like, 0:03:30.562 --> 0:03:32.681 not anything. It's going to be just like crypto. It's 0:03:32.681 --> 0:03:37.282 going to be gone soon. It's massively overhyped. So, um, 0:03:37.322 --> 0:03:39.722 when I see you posting those things, I haven't responded 0:03:39.722 --> 0:03:42.162 to you and I'm sure you're, like, itching to respond 0:03:42.162 --> 0:03:44.282 to me as well. Uh, we haven't done that because 0:03:44.282 --> 0:03:46.321 we planned on doing this. And this is a better, 0:03:46.362 --> 0:03:49.362 I think, a better way to start that off. But, um, 0:03:49.362 --> 0:03:51.482 I would respond to Chris Hoff and he would respond 0:03:51.482 --> 0:03:53.842 back and it would get kind of like adversarial, snarky 0:03:53.842 --> 0:03:58.562 or whatever. Um, but I mean, all that to say 0:03:58.562 --> 0:04:03.122 that I, I think that position is steering people in a, 0:04:03.122 --> 0:04:05.522 in a bad direction. It's telling people not to worry. 0:04:05.802 --> 0:04:09.041 And I'm not saying like panic. I'm saying this thing 0:04:09.042 --> 0:04:10.442 is real and we have to get ready for it. 0:04:10.442 --> 0:04:12.932 I think it's wrong to call what is happening now 0:04:12.932 --> 0:04:18.132 with AI hype and to tell people not to worry. Um, 0:04:18.172 --> 0:04:19.492 because I think it is. 0:04:19.532 --> 0:04:23.972 But when you say it's staring people in the direction of, uh, 0:04:24.212 --> 0:04:26.612 I guess the, for a lack of a better term 0:04:26.612 --> 0:04:30.292 is harmful. What do you think the harm is there? 0:04:30.572 --> 0:04:33.732 So if I say now, this isn't my personal position, 0:04:33.731 --> 0:04:36.732 but let's say I say AI is completely useless. Like 0:04:36.972 --> 0:04:41.851 beyond useless. It does nothing of value whatsoever. What is 0:04:41.851 --> 0:04:44.652 inherently bad about people believing that. 0:04:46.372 --> 0:04:49.252 Uh, they will not, uh, think about their own skills 0:04:49.252 --> 0:04:53.051 and their own capabilities and their own value and how 0:04:53.052 --> 0:04:56.292 that is potentially going to be impacted by AI if 0:04:56.291 --> 0:04:58.731 they don't see AI as a threat because it's a 0:04:58.731 --> 0:05:01.812 nothing burger, and it's basically all hype and marketing and 0:05:01.812 --> 0:05:04.372 a bunch of rich people trying to get rich, uh, 0:05:04.412 --> 0:05:08.252 even more rich, then they will ignore it. So to 0:05:08.291 --> 0:05:11.822 answer your question. The threat, the risk of that is 0:05:11.862 --> 0:05:15.222 them ignoring something that is potentially harmful to their career. 0:05:16.862 --> 0:05:19.862 But I think you could make that argument for, uh, 0:05:19.902 --> 0:05:23.742 we use the the example of blockchain or crypto specifically. 0:05:24.142 --> 0:05:26.702 You could have made that same argument for crypto. We 0:05:26.702 --> 0:05:29.942 could have said, uh, this is going to be this 0:05:29.942 --> 0:05:33.541 future big, impactful technology, which is something many people did say. 0:05:33.541 --> 0:05:36.622 They were like, everything is going to be blockchain, crypto. 0:05:36.862 --> 0:05:39.062 There were people arguing that Doge was going to replace 0:05:39.062 --> 0:05:41.862 the US currency. And you could have made the argument 0:05:41.862 --> 0:05:45.222 back then that if you don't go out right now 0:05:45.222 --> 0:05:47.822 and spend hours and hours of your time learning blockchain, 0:05:48.142 --> 0:05:50.541 you're going to be unemployed. You're not going to be 0:05:50.582 --> 0:05:53.822 able to, uh, make it in the future. So we're 0:05:53.862 --> 0:05:57.061 sort of trying to come up with this reason for 0:05:57.342 --> 0:06:00.182 why AI is different from blockchain. Now, I'm not saying 0:06:00.182 --> 0:06:03.582 it's similar. I, I think most people will agree large 0:06:03.582 --> 0:06:08.462 language models have infinitely more usefulness than blockchain did. But 0:06:08.462 --> 0:06:12.262 then it still comes down to what are we thinking 0:06:12.262 --> 0:06:15.262 that where do we think that this technology is going 0:06:15.502 --> 0:06:18.262 that is going to have such a big impact that 0:06:18.262 --> 0:06:21.062 people need to be paying as much attention as possible? 0:06:21.101 --> 0:06:23.222 And like, what even is the level of attention people 0:06:23.222 --> 0:06:25.702 need to pay? Like, do I just need to know, okay, 0:06:25.742 --> 0:06:27.861 ChatGPT is a thing and here's what it can do. 0:06:28.101 --> 0:06:29.702 Or do I need to spend hours and hours of 0:06:29.702 --> 0:06:32.662 my day learning how to use ChatGPT in order to 0:06:32.702 --> 0:06:35.382 avert whatever problem we're going to say is going to 0:06:35.382 --> 0:06:37.022 be the thing in the future. 0:06:38.782 --> 0:06:44.301 Yeah, sure. I mean. Yeah. Well, first of all, I mean, 0:06:44.342 --> 0:06:47.422 ChatGPT is just one app from one group, but I 0:06:47.422 --> 0:06:49.182 think I think I see what you mean, just AI 0:06:49.182 --> 0:06:52.822 in general. Um, no, I mean, what I advocate is 0:06:52.822 --> 0:06:56.982 people focus more on first principles thinking and like, understanding 0:06:56.981 --> 0:07:00.822 their own opinions and forming their own opinions. Uh, because 0:07:00.822 --> 0:07:02.342 what I'm trying to get to is like, what comes 0:07:02.342 --> 0:07:04.582 after the job replacement, which is like more human to 0:07:04.622 --> 0:07:09.752 human interaction. So I think people need to be more, um, 0:07:09.792 --> 0:07:12.672 thoughtful internally about who they are and what they are 0:07:12.712 --> 0:07:16.752 and like, what they're about actually having opinions. I think 0:07:16.752 --> 0:07:19.272 too many people who are in danger of being replaced 0:07:19.272 --> 0:07:24.632 by this technology, um, simply had a job executing tasks 0:07:24.632 --> 0:07:28.352 that someone else gave them, and that didn't require that 0:07:28.352 --> 0:07:32.072 much intellect, but it required enough intelligence. And this is 0:07:32.072 --> 0:07:36.192 another point we can debate. Like, is AI actually intelligence? Um, 0:07:36.632 --> 0:07:40.432 but it has enough intelligence for all these hundreds of 0:07:40.432 --> 0:07:44.752 millions of knowledge worker jobs. It has enough intelligence required 0:07:44.952 --> 0:07:47.272 that you couldn't automate it. So which is why they 0:07:47.272 --> 0:07:51.232 have the job today, and that if the bar is 0:07:51.272 --> 0:07:54.152 getting to your question, if the bar for AI raises 0:07:54.192 --> 0:07:57.792 above that level. So we could do basic intelligence tasks, 0:07:58.112 --> 0:08:02.472 which is, you know, 90% or 99% of knowledge worker, uh, 0:08:02.562 --> 0:08:06.722 Capabilities is included in that. Then those jobs go away. 0:08:06.762 --> 0:08:09.522 Hundreds of millions. I don't know the actual numbers for 0:08:09.522 --> 0:08:13.522 knowledge worker total employment, but I would say that vast, 0:08:13.522 --> 0:08:16.842 vast majority of those jobs go away and they are 0:08:16.842 --> 0:08:19.602 left not doing anything. And it's not like they could 0:08:19.602 --> 0:08:24.642 just pivot to the next thing, like previous technology trends, because, um, those, 0:08:24.842 --> 0:08:28.122 those pivots that they will make likely are also taken 0:08:28.122 --> 0:08:29.202 over by that. 0:08:29.202 --> 0:08:34.282 I, I understand that line of thinking of, okay, it's 0:08:34.282 --> 0:08:38.042 not like the printing press or, uh, early machine learning 0:08:38.042 --> 0:08:40.762 where I can just go into something that is not 0:08:41.282 --> 0:08:44.562 covered by the current models. But I think the idea 0:08:44.562 --> 0:08:48.202 of this just sweeping replacement, where llms are just going 0:08:48.242 --> 0:08:52.002 to take out 90% of the workforce, I don't personally 0:08:52.002 --> 0:08:55.641 believe that that is rooted in reality. I think, uh, 0:08:55.642 --> 0:09:00.482 we're basically extrapolating with, uh, it's almost like a new car, right? 0:09:00.522 --> 0:09:03.651 If I, If I've just invented the car, it only 0:09:03.652 --> 0:09:07.252 goes 15mph. And then if I do some tweaks to it, 0:09:07.252 --> 0:09:10.132 I spend like a couple of months doing some tweaks. 0:09:10.131 --> 0:09:12.372 It goes a little bit faster. And now we could 0:09:12.372 --> 0:09:14.972 say that if I doubled the speed in three months 0:09:15.292 --> 0:09:17.732 than in six months, I can double the speed again. 0:09:17.892 --> 0:09:20.651 And then you can just keep extrapolating, extrapolating, extrapolating. So 0:09:20.692 --> 0:09:25.372 at some point this car will break the speed of light. Um, 0:09:25.412 --> 0:09:28.052 whereas I think that's what we've been seeing with Llms 0:09:28.052 --> 0:09:30.572 is the LLM technology was very new and it had 0:09:30.572 --> 0:09:33.292 a lot of issues in the beginning. And that gave 0:09:33.292 --> 0:09:35.771 the illusion of a lot of very fast progress. We 0:09:35.772 --> 0:09:39.131 weren't actually seeing these models rapidly getting like better and 0:09:39.131 --> 0:09:41.771 better and better. All that was happening is we made 0:09:41.772 --> 0:09:44.652 a new technology that was a little bit, uh, like finicky. 0:09:44.692 --> 0:09:47.532 It was it had problems. And we slowly ironed out 0:09:47.532 --> 0:09:49.972 some of those problems, which made it look like it 0:09:49.972 --> 0:09:53.252 was advancing at this rapid pace. But it seems now 0:09:53.292 --> 0:09:56.012 that it is sort of capped out. Um, I would 0:09:56.011 --> 0:09:59.212 say the leap from GPT two to GPT three and 0:09:59.302 --> 0:10:03.582 3.5 was pretty groundbreaking. But then as I look at 0:10:03.582 --> 0:10:06.782 the newer models, when we go from 3.5 to whatever 0:10:06.782 --> 0:10:10.542 they've called it now, like 001, I don't even know 0:10:10.542 --> 0:10:13.502 what the naming convention is anymore. Um, we haven't really 0:10:13.502 --> 0:10:17.902 seen any improvements. We're just seeing very slight increases in, oh, 0:10:17.942 --> 0:10:20.102 it's better at coding now. It's a little bit better 0:10:20.102 --> 0:10:22.542 at math. And what it looks like is happening is 0:10:22.582 --> 0:10:26.421 they've made the base large language model, and they've refined 0:10:26.422 --> 0:10:29.822 that almost as much as they can. And then they're 0:10:29.822 --> 0:10:32.702 building these sub models to then make up the slack 0:10:32.702 --> 0:10:35.862 where it can't do math or it can't do programming 0:10:36.102 --> 0:10:39.622 or whatever other areas it struggles. So I think what 0:10:39.622 --> 0:10:42.182 we're actually going to see is we're going to see 0:10:42.502 --> 0:10:45.022 large language models as a technology just cap out like 0:10:45.022 --> 0:10:46.942 they're just going to hit a wall. And then what 0:10:46.942 --> 0:10:49.382 we're going to see is them start building out sub 0:10:49.381 --> 0:10:54.102 models that make them better at a specific job. Now 0:10:54.142 --> 0:10:57.872 that would just be normal technological evolution like I as 0:10:57.912 --> 0:11:01.631 a malware researcher, I got into malware analysis when it 0:11:01.631 --> 0:11:05.232 was already machine learning automated. There was still enough space 0:11:05.232 --> 0:11:08.712 in my industry where I didn't need to become an 0:11:08.712 --> 0:11:10.712 expert in machine learning. I didn't even need to touch 0:11:10.712 --> 0:11:13.271 those models. I still to this day, have no idea 0:11:13.552 --> 0:11:17.872 what the virus classification machine learning models do, because there 0:11:17.872 --> 0:11:20.032 was still enough space for me to operate in that 0:11:20.032 --> 0:11:23.272 field as a human, and I don't see large language 0:11:23.272 --> 0:11:25.792 models as any different from that. I think what we're 0:11:25.792 --> 0:11:29.391 going to just see is people making sub models that make, say, 0:11:29.432 --> 0:11:32.152 malware analysis better. And it's like, okay, maybe I can't 0:11:32.152 --> 0:11:35.392 do manual virus classification anymore, but I can do this 0:11:35.392 --> 0:11:38.352 other thing and I think it's going to move slowly 0:11:38.352 --> 0:11:41.232 enough as they do that, that people can just pivot 0:11:41.511 --> 0:11:45.392 into not even different fields, but into different, different like 0:11:45.392 --> 0:11:47.912 roles in the same field where a lot of their 0:11:47.912 --> 0:11:51.472 skills are transferable. But rather than doing something that maybe 0:11:51.472 --> 0:11:54.802 is very pattern recognition and is very, uh, lm Them 0:11:54.842 --> 0:11:57.842 automatable they could move into something that takes the same 0:11:57.842 --> 0:12:00.042 skills but actually requires more critical thinking. 0:12:01.602 --> 0:12:04.482 Yeah, yeah. I mean, I, I definitely agree there's going 0:12:04.522 --> 0:12:08.642 to be pivots. Um, still possible. Unfortunately, I think those 0:12:08.642 --> 0:12:12.602 pivots are going to be mostly available to the really sharp, 0:12:13.042 --> 0:12:16.682 exceptional people who can pivot up to something much higher 0:12:16.682 --> 0:12:19.041 up in the stack, uh, which requires a lot more 0:12:19.042 --> 0:12:24.002 critical thinking that the AI can't do yet. Um, you 0:12:24.002 --> 0:12:26.521 had the analogy there of speed and approaching the speed 0:12:26.522 --> 0:12:29.282 of light, uh, a car approaching the speed of light. 0:12:29.761 --> 0:12:33.041 So here, here's what I think might be a fundamental difference, uh, 0:12:33.042 --> 0:12:36.602 in our opinions here, um, the speed of light that 0:12:36.602 --> 0:12:39.002 you're talking about, I agree it would be hard to say. 0:12:39.202 --> 0:12:42.962 Went from 30mph to 35mph. And next year we're going 0:12:42.962 --> 0:12:45.242 to be at the speed of light. That's a crazy statement. 0:12:46.522 --> 0:12:48.962 The difference is the speed of light that we're trying 0:12:48.962 --> 0:12:53.202 to get here, um, that I'm talking about is actually approaching. 0:12:53.572 --> 0:12:56.532 Cannot do the work of a knowledge worker. An average 0:12:56.932 --> 0:13:01.131 knowledge worker on earth. Okay. And so my argument to 0:13:01.172 --> 0:13:05.012 you is that that bar is not high at all. 0:13:05.772 --> 0:13:08.651 So it is basically and this is something we talked 0:13:08.652 --> 0:13:12.212 about in text. Um, right in the previous version of this, 0:13:12.212 --> 0:13:16.252 this discussion, it's like the average knowledge worker. They, they 0:13:16.292 --> 0:13:19.892 clock in and clock out. And I'm talking about mean, 0:13:19.932 --> 0:13:23.372 you know, like, uh, average just average across the globe. 0:13:23.372 --> 0:13:26.972 It's like, hey, got some new emails, respond to the email. Uh, 0:13:26.972 --> 0:13:29.292 the email says, hey, go, um, make a list of 0:13:29.292 --> 0:13:32.292 all the documents that are related to this project. Okay, cool. 0:13:32.332 --> 0:13:36.652 Here's your list. They send out the list. Um, someone says, hey, 0:13:36.892 --> 0:13:38.772 I need you to write a report on how our 0:13:38.772 --> 0:13:41.372 finances are doing. So what do they do? Is they 0:13:41.372 --> 0:13:46.052 pull a wiki, uh, article, they review this other document, 0:13:46.052 --> 0:13:49.012 they talk to Sarah in this other department that takes 0:13:49.011 --> 0:13:51.011 them 4 hours or 5 hours or two days or 0:13:51.011 --> 0:13:53.941 however long it takes, and they write a four page report, 0:13:54.302 --> 0:13:55.862 and then they take that report and they put it 0:13:55.862 --> 0:13:57.582 in an email and they send it out to the company. 0:13:57.782 --> 0:14:01.902 They are doing work. They're being paid 50 grand, 100 0:14:01.902 --> 0:14:05.782 grand for this work. And I'm claiming to you that 0:14:05.782 --> 0:14:09.462 this is a low bar, not for I in the future, 0:14:09.462 --> 0:14:12.302 but for I, we already had a year ago, in 0:14:12.302 --> 0:14:15.662 some cases two years ago. So my concern here is 0:14:15.662 --> 0:14:18.702 not about the technical definition of I. I don't give 0:14:18.742 --> 0:14:21.622 two shits about it. People are going to argue about, oh, 0:14:21.622 --> 0:14:25.342 AGI means this technically, specifically, I don't care. All I 0:14:25.382 --> 0:14:29.022 care about is impact to humans. So I'm claiming that 0:14:29.022 --> 0:14:31.862 the AI that we have a year ago already could 0:14:31.862 --> 0:14:35.542 do if it were properly orchestrated. A lot of the 0:14:35.542 --> 0:14:39.061 jobs that I just described. So customer service, um, responding 0:14:39.102 --> 0:14:43.102 to emails, setting up like a lunch meeting, just just 0:14:43.142 --> 0:14:45.582 basic tasks like that that could not be automated with 0:14:45.582 --> 0:14:48.542 regular tech, but can be automated with with this tech. 0:14:51.872 --> 0:14:54.352 Uh. Well, you see, I kind of disagree with that 0:14:54.352 --> 0:14:58.792 because I've unfortunately had to interface with LM based customer 0:14:58.792 --> 0:15:03.512 service agents. And they do have the the flaws that 0:15:03.512 --> 0:15:07.192 both of us are familiar with with LMS. They are 0:15:07.232 --> 0:15:11.472 not deterministic. If I give it the same question 50 times, 0:15:11.472 --> 0:15:14.392 it does not have the same response 50 times. Like 0:15:14.592 --> 0:15:19.272 depending on, uh, I don't know, the, the star alignment, 0:15:19.472 --> 0:15:22.312 how the electricity is going that day, who knows. It 0:15:22.312 --> 0:15:25.192 comes out with a different response to each, um, to 0:15:25.232 --> 0:15:28.952 each input. And that is fundamentally a problem because if 0:15:28.952 --> 0:15:32.232 you need a machine to do a deterministic task like 0:15:32.232 --> 0:15:36.192 finance reports or customer service, you cannot have a model 0:15:36.512 --> 0:15:39.512 that comes out with completely different answers every single time 0:15:39.512 --> 0:15:42.072 you ask the same question. And that is one of 0:15:42.072 --> 0:15:45.832 the major limitations that has been, uh, currently holding back 0:15:45.832 --> 0:15:49.442 large language models is that by their very nature they 0:15:49.442 --> 0:15:52.602 are not deterministic. So all of those tasks you are 0:15:52.602 --> 0:15:56.882 talking about are tasks which, uh, when you say knowledge task, you, 0:15:56.882 --> 0:15:59.842 you're kind of, I think, sort of talking about tasks 0:15:59.842 --> 0:16:03.322 that don't require critical thinking. They don't really require much intelligence. 0:16:03.522 --> 0:16:06.962 They're just doing like tasks that will be the same 0:16:06.962 --> 0:16:09.642 every time you do them. And if the large language 0:16:09.642 --> 0:16:13.402 model cannot even do tasks the same way multiple times, 0:16:13.682 --> 0:16:16.522 how can we then go and automate all of those jobs? 0:16:17.642 --> 0:16:20.602 Yeah, I think I think it's fairly easy. I mean, 0:16:20.882 --> 0:16:23.322 I've been doing this for a long time. Everyone I 0:16:23.322 --> 0:16:26.402 know who's building apps here, it's fairly easy to actually 0:16:26.402 --> 0:16:29.562 give somebody, uh, give an AI a list and say, 0:16:29.562 --> 0:16:31.842 you can only pick from these, like, this is the 0:16:31.842 --> 0:16:33.802 thing that was solved, you know, two and a half 0:16:33.842 --> 0:16:37.882 years ago, probably before that. Um, but you can make 0:16:37.922 --> 0:16:42.562 a system fairly deterministic. Um, and keep in mind, like 0:16:42.562 --> 0:16:46.682 customer service is already horrible. Everyone already hated customer service 0:16:46.682 --> 0:16:49.402 because it was so bad. Before I came out. Right. 0:16:49.442 --> 0:16:52.122 So it's not like customer service is the high bar, 0:16:52.242 --> 0:16:54.082 and we're seeing if we can get there. No, it's 0:16:54.082 --> 0:16:57.802 a horrible bar. Like, everyone hates it already. And oftentimes 0:16:58.082 --> 0:17:00.482 it has the exact same characteristics you were talking about, 0:17:00.482 --> 0:17:03.442 where it's like you're not getting good answers. The good 0:17:03.442 --> 0:17:07.202 answer is actually somewhere there in the knowledge base. But 0:17:07.202 --> 0:17:09.802 somehow this tier one or tier two person you're talking to, 0:17:09.842 --> 0:17:13.962 they didn't find the answer. Um, so a couple of, 0:17:14.002 --> 0:17:18.602 you know, objections to that. I don't think, um, current 0:17:18.602 --> 0:17:21.522 execution of a lot of these jobs is at a 0:17:21.522 --> 0:17:24.202 high level. And that's why I think it's possible to 0:17:24.242 --> 0:17:28.962 beat it. The other thing is, um, you're not actually 0:17:29.002 --> 0:17:32.002 when your boss says, hey, go make a write a 0:17:32.002 --> 0:17:35.322 report for this thing that is fundamentally a non-deterministic thing 0:17:35.362 --> 0:17:38.482 because the problem is always different that the human is given. 0:17:38.482 --> 0:17:41.802 So it definitely does require intelligence, right? You can't give 0:17:41.802 --> 0:17:43.842 that to a script because the script doesn't have like 0:17:43.882 --> 0:17:46.692 an if, then it doesn't have a deterministic thing to 0:17:46.732 --> 0:17:49.492 go look up the thing because the problem is always different. 0:17:49.492 --> 0:17:51.292 And this this is another thing that you and I 0:17:51.652 --> 0:17:54.892 talked about before, which I thought was super interesting, is like, 0:17:55.892 --> 0:18:00.172 what actually is intelligence and how much deviation from a 0:18:00.172 --> 0:18:03.611 standard slotted thing does it require to be called intelligence. 0:18:03.892 --> 0:18:06.612 So I would argue if someone is thrown a giant 0:18:06.612 --> 0:18:09.692 batch of emails and a giant batch of documents and 0:18:09.692 --> 0:18:12.292 they're like, hey, Chris. Hey, Sarah, um, go write a 0:18:12.292 --> 0:18:15.772 report on this. And I need it by today at 4:00. 0:18:17.372 --> 0:18:21.091 No one person or two people or no combination of 0:18:21.092 --> 0:18:23.692 humans is going to make the same report. They're going 0:18:23.692 --> 0:18:26.612 to make vastly different reports based on how they choose 0:18:26.612 --> 0:18:28.852 to go through the documents, based on how they choose 0:18:28.852 --> 0:18:30.652 to form the thing. How long is the report going 0:18:30.692 --> 0:18:32.852 to be? If you ask a thousand people to make 0:18:32.852 --> 0:18:35.292 a report, they're all going to be different, just like 0:18:35.292 --> 0:18:38.212 with an LLM. So and the other thing is just 0:18:38.212 --> 0:18:42.581 because that doesn't require like PhD level invention or creativity 0:18:43.182 --> 0:18:46.862 that still requires human intelligence. No, no tech prior to 0:18:46.902 --> 0:18:48.582 AI could do that job. 0:18:50.462 --> 0:18:51.982 So I'd actually would like to go back to the 0:18:51.982 --> 0:18:56.262 customer service example, because I think I would argue customer 0:18:56.262 --> 0:18:59.861 service is horrible for the exact same reason that the 0:18:59.862 --> 0:19:03.662 solution you proposed would be horrible. They are given these 0:19:03.662 --> 0:19:07.662 limited scripts of answers where if customer asks this, then 0:19:07.662 --> 0:19:10.742 you must answer that and they cannot go beyond their script. 0:19:10.942 --> 0:19:12.782 So if I come up with a unique problem that 0:19:12.782 --> 0:19:16.502 requires some kind of critical thinking, although the human on 0:19:16.502 --> 0:19:20.542 the phone has intelligence and is capable of critical thinking, 0:19:20.782 --> 0:19:23.182 they are not allowed to go outside of my script. 0:19:23.422 --> 0:19:25.942 And essentially you would run into the same problem with 0:19:25.942 --> 0:19:28.182 trying to do that with a large language model is 0:19:28.182 --> 0:19:30.302 you've just given it a list of canned answers it's 0:19:30.302 --> 0:19:32.381 not allowed to go outside of. And you're going to 0:19:32.382 --> 0:19:36.142 get the exact same results where there is no critical thinking. 0:19:36.302 --> 0:19:38.262 We can get the argument that I don't think I 0:19:38.302 --> 0:19:42.112 can think like full stop. But you're essentially just recreating 0:19:42.112 --> 0:19:46.352 the same awful system where it's just it's just parroting 0:19:46.352 --> 0:19:49.712 back canned responses. And I'm trying to ask my question, 0:19:49.872 --> 0:19:52.312 and it doesn't match up to some canned response that 0:19:52.311 --> 0:19:55.552 they have. So they're just repeating the same closest line 0:19:55.552 --> 0:19:57.872 that they have in their script. And then we both 0:19:57.872 --> 0:20:02.232 leave the interaction completely annoyed. And, um, I think that's 0:20:02.232 --> 0:20:04.152 exactly what is going to happen with large language models. 0:20:04.152 --> 0:20:07.671 And it's already happening. They're just feeding their dogs or 0:20:07.672 --> 0:20:11.872 their scripts into the LLM, and it's just making a 0:20:11.912 --> 0:20:15.752 worse version of a customer service representative, because not only 0:20:15.752 --> 0:20:18.631 does it still have the canned responses, but it doesn't 0:20:18.632 --> 0:20:20.672 have the critical thinking to go, oh, maybe I should 0:20:20.672 --> 0:20:21.952 escalate this to my boss. 0:20:23.071 --> 0:20:28.912 Yeah, no, it's an interesting point. I, I like that argument. Um, 0:20:29.472 --> 0:20:31.552 I would say the best argument I have against that 0:20:31.552 --> 0:20:35.911 is that, um, companies have been trying this customer service 0:20:35.952 --> 0:20:40.722 thing for decades, Like, uh, human customer service with the 0:20:40.722 --> 0:20:45.362 different levels. And they have tried like millions of different things. 0:20:45.482 --> 0:20:47.722 They've tried to give people autonomy. They've tried to not 0:20:47.722 --> 0:20:51.082 give people autonomy. The system they came up with is, look, 0:20:51.482 --> 0:20:55.121 follow this script, use this document. So I would argue 0:20:55.122 --> 0:20:56.881 to you that the reason that we see this in 0:20:56.882 --> 0:20:59.442 the industry is because it's the thing that works best. 0:20:59.882 --> 0:21:03.522 If what worked best was to just hire people who 0:21:03.522 --> 0:21:07.362 are somewhat smart and give them free rein, then that's 0:21:07.362 --> 0:21:09.722 what companies would be doing. Uh, and I don't think 0:21:09.722 --> 0:21:12.042 they're doing that because it doesn't work. Now, I would 0:21:12.042 --> 0:21:14.602 argue that as you move into tier three, tier four 0:21:14.602 --> 0:21:18.561 or whatever, whatever the structure is for these, um, like 0:21:18.602 --> 0:21:21.562 more senior customer service people, I feel like that thing 0:21:21.561 --> 0:21:24.762 you're describing is exactly what's happening. They are given free 0:21:24.762 --> 0:21:28.401 rein because all the known answers have already been tried, 0:21:28.882 --> 0:21:32.321 and now they're, um, now now they're going to have 0:21:32.321 --> 0:21:33.362 to use critical thinking. 0:21:36.122 --> 0:21:39.611 But I think a lot more jobs than you maybe 0:21:39.612 --> 0:21:43.372 estimate are they fall into that category of like, sure, 0:21:43.372 --> 0:21:46.012 a lot of it is automatable, but the reason they 0:21:46.012 --> 0:21:48.252 need a human in that role and not a script 0:21:48.252 --> 0:21:51.492 or a machine or whatever we want to call it, 0:21:51.772 --> 0:21:57.052 is because there is some aspect that requires critical thinking, um, uh, 0:21:57.052 --> 0:21:59.692 like the one I will push back against endlessly until 0:21:59.692 --> 0:22:02.932 my last breath is software engineers. I think there is 0:22:02.932 --> 0:22:06.812 no point in any of the AI's evolution, at which 0:22:06.811 --> 0:22:09.572 point you will be able to replace software engineers. And 0:22:09.571 --> 0:22:12.891 I think the disconnect that happens is a lot of 0:22:12.892 --> 0:22:16.892 the non software engineers or even the coders, the people 0:22:16.892 --> 0:22:20.532 who just don't fully understand software engineering don't understand the 0:22:20.532 --> 0:22:24.932 difference between writing code and designing software. So they're looking 0:22:24.932 --> 0:22:30.371 at ChatGPT and Cursor and Gemini or whatever, and they're looking, oh, 0:22:30.372 --> 0:22:31.851 I can put in a query and it spits out 0:22:31.852 --> 0:22:34.252 all of this code. And they're like, that is what 0:22:34.342 --> 0:22:38.422 coding is. So like that very easily could replace software engineering. 0:22:38.422 --> 0:22:41.422 But because they're not a software engineer, they don't understand 0:22:41.422 --> 0:22:44.381 the amount of critical thinking and design decisions that actually 0:22:44.382 --> 0:22:47.222 go into the software. I think, uh, a lot of 0:22:47.222 --> 0:22:50.902 software engineers will say that writing code is like 10% 0:22:50.902 --> 0:22:53.942 of software engineering. But then if you're not a software engineer, 0:22:53.942 --> 0:22:58.061 you're looking at that, that a small percentage that the AI, okay, 0:22:58.302 --> 0:23:01.262 sort of can do that and then ignoring the rest. 0:23:01.462 --> 0:23:04.342 And I think the same is true for every role. Um, 0:23:04.342 --> 0:23:07.582 until you actually have like spent a year in any 0:23:07.582 --> 0:23:11.262 given job, you don't really understand how much is automatable 0:23:11.262 --> 0:23:15.102 versus how much actually does require a human with critical thinking. 0:23:15.342 --> 0:23:19.502 So my personal belief is, I think for the large, 0:23:19.662 --> 0:23:23.782 like the overwhelming majority of all roles, there will never, 0:23:23.821 --> 0:23:26.702 ever be a point in which AI replaces that person. 0:23:27.102 --> 0:23:29.381 It might be able to make the person faster. You 0:23:29.382 --> 0:23:31.982 might be able to take one person and have them 0:23:32.022 --> 0:23:35.272 do work at a slightly more productive rate, but I 0:23:35.272 --> 0:23:37.832 don't think there is any point in which I actually 0:23:37.832 --> 0:23:42.032 starts replacing like, like large swaths of the job market. 0:23:43.512 --> 0:23:47.392 Yeah. Interesting. Um, so let me give you a counterpoint. 0:23:47.392 --> 0:23:50.552 And this this is anecdotal, of course, but I have 0:23:50.592 --> 0:23:54.272 a close friend who is a cardiologist, like actual MD 0:23:54.352 --> 0:23:56.912 actually sits in the, you know, the doctor's office every day. 0:23:57.272 --> 0:24:00.832 He also happens to be a bug bounty hunter and like, 0:24:00.872 --> 0:24:05.912 a programmer. So he he is, uh, you know, using 0:24:05.912 --> 0:24:09.792 whatever tools are approved at his particular job. And, um, 0:24:09.792 --> 0:24:12.672 he is now finding as of like this is this 0:24:12.672 --> 0:24:15.512 is six months ago. This is a year ago. Um, 0:24:15.952 --> 0:24:19.392 the analysis that this, this thing is doing is as 0:24:19.392 --> 0:24:23.152 good or better in some cases than the thing that 0:24:23.152 --> 0:24:26.191 he was going to write. So this is at the 0:24:26.192 --> 0:24:29.472 cardiologist level. Not not even like a nursing level. This 0:24:29.472 --> 0:24:32.322 is really, really advanced stuff. So it's it's taking all 0:24:32.321 --> 0:24:35.482 the inputs that came from like the patient and also 0:24:35.802 --> 0:24:39.522 from his notes. It's combining all that together with, of course, 0:24:39.522 --> 0:24:42.802 the knowledge of the model itself. And then, you know, 0:24:42.842 --> 0:24:47.682 giving out recommendations, recommending the actual drugs. And so my 0:24:47.722 --> 0:24:49.841 thing is like if this thing is actually doing the 0:24:49.842 --> 0:24:53.482 job and it's not just one person, right? A lot 0:24:53.522 --> 0:24:56.562 of doctors are actually saying that the output is matching 0:24:56.561 --> 0:25:00.282 their capabilities as well. We've already seen this with analyzing moles. 0:25:00.282 --> 0:25:03.762 We've seen this like in lots of different places. Um, 0:25:03.802 --> 0:25:05.922 and that brings me back to this one other point 0:25:05.922 --> 0:25:09.002 that you made earlier, which I thought was really interesting, um, 0:25:09.402 --> 0:25:11.561 how they're going to hit a wall. The Elon's hit 0:25:11.602 --> 0:25:14.562 a wall and they basically start building subsystems. This is 0:25:14.561 --> 0:25:18.162 exactly how I define AGI. I don't think AGI is 0:25:18.162 --> 0:25:22.242 going to be some breakthrough, generally intelligent thing. I think 0:25:22.242 --> 0:25:26.122 that is more of an AI breakthrough than we've actually seen. 0:25:26.162 --> 0:25:29.172 I don't think it's happened yet. What I'm arguing is 0:25:29.172 --> 0:25:32.372 going to replace the average knowledge worker is actually one 0:25:32.372 --> 0:25:37.812 of these, uh, Frankenstein systems where the LLM is like 0:25:37.811 --> 0:25:42.252 an orchestrator, and it's basically spinning up these individual things 0:25:42.252 --> 0:25:45.452 to go and accomplish individual tasks like write the email, 0:25:46.012 --> 0:25:50.492 organize the events, organize the conference, uh, summarize this or whatever. 0:25:50.612 --> 0:25:53.772 So it's going to have at its disposal dozens or 0:25:53.811 --> 0:25:57.611 hundreds of these little workers. But the problem is that 0:25:57.612 --> 0:25:59.692 you're going to buy this thing, let's call it a 0:25:59.852 --> 0:26:03.012 synth worker or whatever the stupid AI company is going 0:26:03.012 --> 0:26:06.532 to be called. You buy, you buy synth worker, synth 0:26:06.532 --> 0:26:09.292 worker shows up to work on Monday. It goes to 0:26:09.332 --> 0:26:12.452 the onboarding meeting. It talks to the manager. It talks 0:26:12.452 --> 0:26:15.892 to all its team members. It reads slack, it reads 0:26:15.892 --> 0:26:19.171 the wiki. And it starts working just like a regular employee. 0:26:19.532 --> 0:26:21.932 And it produces the work of a regular employee. And 0:26:21.932 --> 0:26:25.932 I would argue in a lot of cases, even better. Um, 0:26:26.492 --> 0:26:29.142 but if you look under the hood, it's not some 0:26:29.182 --> 0:26:32.862 all intelligent AI system. It's actually a whole bunch of tricks. 0:26:32.862 --> 0:26:35.341 It's like this big agent orchestrating a whole bunch of 0:26:35.342 --> 0:26:38.462 other ones. But functionally, what you end up with is 0:26:38.462 --> 0:26:42.262 this new AI named Julie actually does the work. And 0:26:42.262 --> 0:26:44.622 guess what? We didn't hire a human to do that job. 0:26:46.702 --> 0:26:48.982 Um, so I guess the first thing I'd say is 0:26:48.982 --> 0:26:52.821 that is not the definition of AGI that most people 0:26:52.821 --> 0:26:56.462 are going with. Um, but it is a fair point. Um, 0:26:56.502 --> 0:26:58.702 a lot of people have this idea of, uh, if 0:26:58.702 --> 0:27:01.422 we give the llms enough data, at some point, they 0:27:01.422 --> 0:27:04.061 start learning for themselves, and at some point they can 0:27:04.102 --> 0:27:06.782 do all of these things at the Submodels are doing, um, 0:27:06.821 --> 0:27:09.742 which I think most likely we both agree that that's 0:27:09.742 --> 0:27:13.062 that's horseshit. That's that's never happening. Um, but then the 0:27:13.061 --> 0:27:17.861 problem is when it comes to customizing these models, um, 0:27:17.902 --> 0:27:20.662 now you need a custom model for everything. You, uh, 0:27:20.662 --> 0:27:23.982 you need your model to identify tumors. And that is 0:27:23.982 --> 0:27:26.542 no different from where we've been like. That is what 0:27:26.542 --> 0:27:29.742 we've been doing for the past. Uh, I don't know. 0:27:29.782 --> 0:27:32.782 Probably longer than I've been alive, like 30 years, maybe more. 0:27:33.102 --> 0:27:36.502 And that hasn't replaced humans. Um, I would go back 0:27:36.502 --> 0:27:39.422 to my own personal example, which is machine learning based 0:27:39.422 --> 0:27:46.742 malware detection. Those machines have massively improved, uh, malware detection. Like, uh, 0:27:46.742 --> 0:27:48.702 back in the old days, you would have to open 0:27:48.702 --> 0:27:51.662 up a virus. You would have to manually determine that 0:27:51.662 --> 0:27:55.062 this is actually malicious code, and then enter a signature 0:27:55.061 --> 0:27:57.542 into a database. And can you imagine, like a human 0:27:57.542 --> 0:28:00.262 trying to do that with billions and billions of new 0:28:00.302 --> 0:28:04.142 files every single day? It's impossible. Um, so we built 0:28:04.142 --> 0:28:07.662 machine learning models to do that. But then the question is, well, 0:28:07.702 --> 0:28:10.341 how many malware analysts got laid off as a result? 0:28:10.821 --> 0:28:13.702 And the answer is there's actually more malware analysts today 0:28:13.902 --> 0:28:17.022 than there was in those days before the machine learning 0:28:17.022 --> 0:28:20.582 based virus classification. And the reason was the classification could 0:28:20.582 --> 0:28:23.022 only do so much, and you still needed people to 0:28:23.152 --> 0:28:25.952 check false positives. You still needed people to write reports, 0:28:26.071 --> 0:28:29.631 and there was always a place for humans and it 0:28:29.632 --> 0:28:32.552 wasn't even different skills like, uh, my job has been 0:28:32.552 --> 0:28:35.872 automated three times now, and in none of those three 0:28:35.872 --> 0:28:38.872 times have I had to, uh, even learn like, some 0:28:38.872 --> 0:28:41.472 massively different skill set. I've always been able to pivot, 0:28:41.752 --> 0:28:44.152 and I think that is all we're going to see 0:28:44.152 --> 0:28:48.432 with that is like, okay, maybe our cardiologist, I can 0:28:48.472 --> 0:28:51.232 automate a lot of that. So the cardiologists end up 0:28:51.232 --> 0:28:53.192 doing something else. I can't say what else they might 0:28:53.192 --> 0:28:54.792 end up doing because that's not my field and I 0:28:54.792 --> 0:28:57.432 don't know anything about it. But I do know that 0:28:57.432 --> 0:29:00.832 in my field, we have just been automating every single thing. 0:29:00.832 --> 0:29:03.952 We've been building sub models for malware classification. We've built 0:29:03.952 --> 0:29:07.512 some models for, uh, security alerts. We've built uh, sub 0:29:07.512 --> 0:29:10.911 models for analyzing code. Um, and we built up all 0:29:10.912 --> 0:29:13.232 of these sub models. But there's two problems there. The 0:29:13.232 --> 0:29:15.752 first is there is no evidence of it ever actually 0:29:15.752 --> 0:29:19.552 replacing humans. In fact, we've, uh, the industry has grown 0:29:19.552 --> 0:29:23.322 by a factor of ten since these machine learning models 0:29:23.322 --> 0:29:26.162 started coming out. And the second thing is they're very 0:29:26.162 --> 0:29:30.682 expensive to produce and maintain. So you're not actually getting 0:29:30.682 --> 0:29:32.882 the thing, which is the AI companies are trying to 0:29:32.882 --> 0:29:34.682 sell people on, which is, oh, you can get rid 0:29:34.682 --> 0:29:37.722 of employees, get rid of Greg, get rid of Simon, 0:29:37.722 --> 0:29:40.362 get rid of Sandra, and just keep all their salaries. 0:29:40.362 --> 0:29:43.562 Just pay $20 a month for this AI subscription. And 0:29:43.562 --> 0:29:46.242 now you're saving all of this, all of this money. 0:29:46.602 --> 0:29:49.202 But then when it comes to the actual who's maintaining 0:29:49.202 --> 0:29:53.082 these models, who's updating them for the latest, uh, whatever 0:29:53.082 --> 0:29:56.802 it is they're doing, because, um, like, obviously the technologies 0:29:56.802 --> 0:29:59.202 that they're working on are always changing. So someone has 0:29:59.202 --> 0:30:03.122 to maintain these models and they're phenomenally expensive to run. 0:30:03.402 --> 0:30:07.162 Like any kind of pattern matching model requires insane amounts 0:30:07.162 --> 0:30:10.802 of computational power. And what's actually just happening is you're 0:30:10.802 --> 0:30:13.842 taking those salaries, and I don't think you're even cutting them. 0:30:13.842 --> 0:30:17.482 You're actually spending more running these models or paying to 0:30:17.522 --> 0:30:20.732 have the models run than you would be having employees 0:30:20.732 --> 0:30:24.212 do it. It only makes sense when you're operating at scale, 0:30:24.772 --> 0:30:27.172 where employees physically like there aren't enough employees to do 0:30:27.172 --> 0:30:30.132 this thing. Like, imagine if you had to hire, uh, 0:30:30.172 --> 0:30:33.612 humans to manually categorize a billion malware samples. Like, it 0:30:33.612 --> 0:30:36.652 just wouldn't be possible. Uh, but when it actually comes to, hey, 0:30:36.652 --> 0:30:39.852 let's replace a human with this AI, uh, it usually 0:30:39.852 --> 0:30:42.732 ends up just working out more expensive. And the only 0:30:42.732 --> 0:30:46.892 reason we haven't started seeing that with Llms yet is 0:30:46.892 --> 0:30:50.972 because they're being subsidized by VCs. We don't see how much. Uh, 0:30:51.012 --> 0:30:52.932 I'm just going to say ChatGPT, because that's the one 0:30:52.932 --> 0:30:56.092 most people are familiar with. Um, we don't really see 0:30:56.132 --> 0:30:59.171 how much ChatGPT costs to actually operate. We don't know 0:30:59.172 --> 0:31:02.212 how much they're spending on acquiring the training data, training 0:31:02.212 --> 0:31:05.532 the model, how much computational power is being used, how 0:31:05.532 --> 0:31:08.372 much electricity is being used. And then when you think 0:31:08.372 --> 0:31:11.852 about an actual human being, it's just some cells. They 0:31:11.852 --> 0:31:14.252 need a little bit of water, a little bit of glucose, 0:31:14.532 --> 0:31:17.942 the amount of money it costs to operate a actual 0:31:17.942 --> 0:31:21.142 human being is almost zero. The only reason it's not 0:31:21.142 --> 0:31:24.661 zero is because capitalism, we've made houses so expensive, so 0:31:24.662 --> 0:31:26.862 now they've got to pay rent and we want to 0:31:26.902 --> 0:31:29.742 make a profit selling food. So now they've got to buy, uh, 0:31:29.782 --> 0:31:33.662 like this $0.01 banana is now a dollar. Um, so 0:31:33.662 --> 0:31:36.982 we basically just increased the cost of running humans to 0:31:37.022 --> 0:31:40.702 the point where these machines now look compatible. But then 0:31:40.702 --> 0:31:42.542 when you try and actually replace the humans, you're just 0:31:42.542 --> 0:31:44.382 going to run into the exact same problem, which is 0:31:44.382 --> 0:31:47.302 these machines are expensive. They may not need to pay rent. Actually, 0:31:47.342 --> 0:31:48.822 technically they do need to pay rent. You need a 0:31:48.822 --> 0:31:51.182 data centre, but they may not need to eat. They 0:31:51.182 --> 0:31:53.942 may not need water, they may not need plumbing. Uh, 0:31:53.942 --> 0:31:56.742 but they need a billion, trillion dollars worth of GPUs, 0:31:56.782 --> 0:32:00.142 a ton of electricity, a ton of cooling. And I 0:32:00.142 --> 0:32:03.262 actually don't think we are going to reach a point 0:32:03.262 --> 0:32:08.062 where we can make, uh, like computational based intelligence more 0:32:08.422 --> 0:32:12.102 cheaper than the equivalent human intelligence. I think right now 0:32:12.142 --> 0:32:16.032 we're in, uh, I call it sort of the Utopia era. 0:32:16.072 --> 0:32:17.912 Even though I don't feel like we're in Utopia, I 0:32:17.912 --> 0:32:21.112 think everything is terrible. I hate what's going on with AI. 0:32:21.472 --> 0:32:25.512 It makes me miserable. But it reminds me of, um, uh, 0:32:25.512 --> 0:32:27.712 in LA, we had this thing called, uh. Well, it's 0:32:27.752 --> 0:32:30.392 actually it's quite widespread now, but we had this program 0:32:30.392 --> 0:32:32.992 called the Bird Scooter, and it was actually the very 0:32:32.992 --> 0:32:35.832 first bird scooter was in LA, and I was there 0:32:35.832 --> 0:32:39.072 for the pilot program. It's like, okay, free public transport. 0:32:39.072 --> 0:32:41.112 You pay like maybe a dollar at most. You can 0:32:41.112 --> 0:32:43.712 go anywhere, you can leave the scooter anywhere you want. 0:32:43.992 --> 0:32:46.432 And I was like, this is incredible. Like, this is 0:32:46.432 --> 0:32:50.432 going to revolutionize transport. I can go from point A 0:32:50.472 --> 0:32:52.712 to point B, I don't need to find parking. It 0:32:52.752 --> 0:32:55.152 cost me a dollar. Like this is perfect. But what 0:32:55.152 --> 0:32:58.392 was actually happening on the back end is VCs were 0:32:58.432 --> 0:33:00.752 pouring billions and billions of dollars into this system to 0:33:00.792 --> 0:33:03.392 try and, like, build out this company. And what was 0:33:03.392 --> 0:33:05.712 happening is people were getting hit by cars. They were 0:33:05.712 --> 0:33:08.592 throwing the scooters into rivers and oceans, and they were 0:33:08.632 --> 0:33:12.232 actually losing more money, uh, to running this program than 0:33:12.232 --> 0:33:15.122 they were ever going to make. And I think we're 0:33:15.122 --> 0:33:17.522 seeing the same thing start to happen with AI is 0:33:17.522 --> 0:33:21.681 it's being funded by VCs, but it's there's not really 0:33:21.682 --> 0:33:24.802 yet any certain path where it's even going to be profitable. 0:33:25.002 --> 0:33:28.402 Like they are currently operating at a loss, and they're 0:33:28.402 --> 0:33:31.082 selling us on this idea that at some point you 0:33:31.082 --> 0:33:33.882 will be able to run like a genetic AI or 0:33:33.882 --> 0:33:37.282 whatever for less than the equivalent of an employee. But 0:33:37.322 --> 0:33:39.482 I actually don't think that's going to happen. I don't 0:33:39.482 --> 0:33:42.642 think we're going to reach a point where computers are 0:33:42.642 --> 0:33:46.242 cheap enough to run, that we can replace humans with these, uh, 0:33:46.402 --> 0:33:49.402 with these AI agents. And I think the more that 0:33:49.402 --> 0:33:52.322 we start, uh, going from this idea, which is never 0:33:52.322 --> 0:33:55.322 going to work of a single AI that is just 0:33:55.322 --> 0:33:58.282 able to do everything to building out sub models, I 0:33:58.282 --> 0:34:01.442 think that's just going to it's going to compound the cost. Uh, 0:34:01.442 --> 0:34:03.322 we're just going to get more and more and more 0:34:03.322 --> 0:34:05.962 cost to the point where eventually we'll come to the 0:34:06.002 --> 0:34:08.642 realization that actually, it's cheaper to just hire humans. 0:34:09.722 --> 0:34:14.292 Okay. I like this argument. I find that interesting. I 0:34:14.692 --> 0:34:18.132 don't believe it's true. My intuition is that it's not true. 0:34:18.292 --> 0:34:22.492 And the reason for that is just the the amount 0:34:22.492 --> 0:34:25.932 that the, um, the costs are falling. And the fact 0:34:25.932 --> 0:34:28.372 that I think we're so bad at what we're doing 0:34:28.372 --> 0:34:31.732 right now with artificial intelligence. I think there's so much 0:34:31.732 --> 0:34:33.972 slack in the rope that I think we're going to 0:34:33.972 --> 0:34:38.572 end up getting, you know, 99% or whatever, some, some 0:34:38.572 --> 0:34:41.492 dramatic amount of the cost that we're paying now just 0:34:41.492 --> 0:34:44.012 will keep falling out and falling out. But I would 0:34:44.012 --> 0:34:46.732 say that's an empirical question. And we'll just have to 0:34:46.772 --> 0:34:49.212 like see how that works out. Because I do agree 0:34:49.212 --> 0:34:52.692 with you that there is a the current state is 0:34:52.732 --> 0:34:55.332 like a state of confusion because you don't know how 0:34:55.332 --> 0:34:58.292 much it's being propped up. So so I like your argument, 0:34:58.532 --> 0:35:01.612 and I definitely think the there will be an effect 0:35:01.612 --> 0:35:07.612 from what you're saying. Um, I, um, I want to 0:35:07.612 --> 0:35:09.942 I want to touch on this, this whole concept of 0:35:10.222 --> 0:35:16.142 intelligence itself in like, because I feel like, um, we 0:35:16.142 --> 0:35:18.702 were trying to define it before, uh, when we had 0:35:18.702 --> 0:35:22.142 the previous conversation, we were trying to define like what, uh, 0:35:22.182 --> 0:35:25.302 what exactly do we mean by it? So I think 0:35:25.542 --> 0:35:29.182 the way that I defined it before was, um, or 0:35:29.182 --> 0:35:30.702 at least the way that I wanted to define it now, 0:35:30.702 --> 0:35:33.782 which which I think is roughly similar, the ability to 0:35:33.822 --> 0:35:38.982 solve everyday human problems using knowledge about the world. So 0:35:39.022 --> 0:35:41.022 I think that I try to hit it from a 0:35:41.022 --> 0:35:44.062 bunch of angles, and I think that accounts for why 0:35:44.102 --> 0:35:48.982 an average knowledge worker can't be automated. Um, and I 0:35:49.022 --> 0:35:51.782 put every day in there because it's basic stuff like, 0:35:51.982 --> 0:35:54.902 should I break up with my boyfriend? Um, and keep 0:35:54.902 --> 0:35:57.302 in mind, just just the case of, like, how useful 0:35:57.302 --> 0:36:01.062 this stuff is. I think the current numbers are OpenAI 0:36:01.182 --> 0:36:05.582 or ChatGPT has like a billion users per day. So 0:36:05.582 --> 0:36:09.992 I feel like for everyday problems, It's if we're using 0:36:09.992 --> 0:36:12.552 that as a as a metric doesn't that to me 0:36:12.552 --> 0:36:15.512 indicates that it's very useful. Um, but but I guess 0:36:15.512 --> 0:36:17.752 that's a red herring because you're, you're already acknowledged that 0:36:17.752 --> 0:36:21.552 it's useful. Um, but I've got some examples of, like, 0:36:22.152 --> 0:36:25.432 things that ChatGPT is really good at. So write an 0:36:25.432 --> 0:36:27.712 essay about a book that you read. It could already 0:36:27.712 --> 0:36:30.192 write pretty good essays. Oh, and by the way, this 0:36:30.192 --> 0:36:33.032 is unrelated to the debate, but we actually blew by 0:36:33.032 --> 0:36:36.112 the Turing test because most people can't tell the difference 0:36:36.112 --> 0:36:39.592 between AI and humans at this point, because it's already. 0:36:39.992 --> 0:36:41.992 I just thought it was an interesting piece of trivia 0:36:41.992 --> 0:36:44.272 that that used to be the gold standard for AI, 0:36:44.272 --> 0:36:47.712 and now no one even cares that we passed it. Um, managing. 0:36:47.792 --> 0:36:52.672 That one was actually passed the with the very early GPT. Um, 0:36:52.672 --> 0:36:53.312 I thought I. 0:36:53.312 --> 0:36:53.792 Think it might. 0:36:53.832 --> 0:36:58.272 Have. Yeah. Because I mean, so as, like a, like 0:36:58.552 --> 0:37:01.192 putting my, uh, like my philosophy hat on instead of 0:37:01.192 --> 0:37:04.592 my scientist hat on. I have never, ever seen the 0:37:04.592 --> 0:37:09.162 Turing test as a reliable metric for anything. Because essentially, 0:37:09.162 --> 0:37:11.642 for anyone who doesn't know what that is, it is 0:37:11.642 --> 0:37:14.762 just you put a machine and you put a person, uh, 0:37:15.162 --> 0:37:19.322 like basically concealed from the person they're talking to. And 0:37:19.322 --> 0:37:22.922 the question is, can that person distinguish between a machine 0:37:23.682 --> 0:37:26.762 or a human conversation? But the bar for that is 0:37:26.762 --> 0:37:30.342 very low, because your average person will spend like 30 0:37:30.342 --> 0:37:33.762 hours arguing with a troll bot on Twitter, not realizing 0:37:33.762 --> 0:37:36.681 that that is a Python script. Um, so I've always 0:37:36.682 --> 0:37:39.562 felt that Turing test is more a test of humans 0:37:39.562 --> 0:37:44.042 lack of ability to distinguish a real human, rather than, uh, 0:37:44.042 --> 0:37:47.322 AI being successful at what it does. Now, granted, I 0:37:47.322 --> 0:37:51.482 will concede that AI is, or at least Llms are very, 0:37:51.482 --> 0:37:55.562 very good at imitating human like conversation. I will give 0:37:55.562 --> 0:37:59.322 them that. But there is a huge difference between imitating 0:37:59.322 --> 0:38:03.202 human like conversation and human intelligence, which is why I 0:38:03.202 --> 0:38:06.812 believe that people aren't, uh, that they're not seeing it 0:38:06.812 --> 0:38:09.892 passing the Turing test as this amazing feat because it's 0:38:09.892 --> 0:38:13.452 not showing that the AI can mimic human intelligence. It's 0:38:13.452 --> 0:38:15.812 showing that it can mimic human conversation, which is a 0:38:15.812 --> 0:38:16.812 very different thing. 0:38:17.412 --> 0:38:20.732 Yeah, that makes sense. I think we agree there. Um, 0:38:20.892 --> 0:38:23.972 so what about this definition that I'm using? Um, and 0:38:23.972 --> 0:38:26.372 I can't remember the exact definitions that we used before. 0:38:26.372 --> 0:38:28.572 We had a slight disagreement there. But what do you 0:38:28.572 --> 0:38:31.452 think about this? The ability to solve everyday human problems 0:38:31.452 --> 0:38:36.812 using knowledge about the world. So it's kind of general. Right. 0:38:37.012 --> 0:38:39.012 And it's general in the sense that a knowledge worker 0:38:39.012 --> 0:38:41.492 could do it. It's general in the sense that AGI 0:38:41.532 --> 0:38:43.652 would be able to do it. AC obviously would be 0:38:43.652 --> 0:38:46.892 able to do it. But I've got examples here managing 0:38:46.892 --> 0:38:51.132 someone's daily schedule, writing performance review for an employee, reading 0:38:51.132 --> 0:38:54.532 and answering emails, writing code for an application at a company. Um, 0:38:54.532 --> 0:38:59.252 by the way, I completely agreed with, um, your your 0:38:59.252 --> 0:39:03.892 differentiation between writing code and doing software engineering or software architecture. 0:39:03.892 --> 0:39:06.132 Completely agree with you there. I think that's a big 0:39:06.132 --> 0:39:09.132 thing that coding is missing out on. Um, writing a 0:39:09.132 --> 0:39:12.692 requirements document based on conversations, writing a work status update 0:39:12.692 --> 0:39:16.172 for your boss plan, a four day trip to Switzerland. 0:39:16.172 --> 0:39:19.572 So these are my examples of like everyday human problems 0:39:19.852 --> 0:39:24.532 that are nowhere near free, like automation could possibly even 0:39:24.572 --> 0:39:28.412 try to do versus, um, with AI, these are kind 0:39:28.412 --> 0:39:30.532 of like trivial. And I would argue that these are 0:39:30.532 --> 0:39:33.212 the types of things that are going to make up 0:39:33.692 --> 0:39:36.412 a human replacing AI in the workforce. 0:39:38.292 --> 0:39:43.532 I'm not sure I agree because, um, uh, the the 0:39:43.612 --> 0:39:46.732 point on, uh, I mean, it goes back to your 0:39:46.732 --> 0:39:50.972 general knowledge worker point, the question of can it replace 0:39:51.012 --> 0:39:54.612 a average worker? Um, and we've talked about that already. Uh, 0:39:54.612 --> 0:39:57.292 but I don't think that is the definition of intelligence. 0:39:57.292 --> 0:40:00.582 For me, the definition of intelligence is not the ability 0:40:00.582 --> 0:40:04.742 to perform a task that wasn't performable by previous automation. 0:40:04.942 --> 0:40:08.782 It is the ability to work with incomplete information. Like 0:40:08.782 --> 0:40:12.502 if I were to, I don't know. Let's say I 0:40:12.542 --> 0:40:14.702 build you a puzzle and I take away some of 0:40:14.702 --> 0:40:17.302 the pieces and let's say I take away five of 0:40:17.302 --> 0:40:22.182 the pieces. Um, and I give you the puzzle with the, uh, 0:40:22.182 --> 0:40:25.702 without those pieces that I took away, you could probably 0:40:25.982 --> 0:40:28.942 at some level redraw the rest of the puzzle. Right? 0:40:28.982 --> 0:40:32.502 You could use your, uh, critical thinking and your intelligence 0:40:32.502 --> 0:40:35.382 to see. Oh, okay. I don't have all the pieces here, 0:40:35.382 --> 0:40:37.622 but I can see this is clearly a photo of 0:40:37.662 --> 0:40:41.182 a waterfall. Right? It's that ability to work with. Not 0:40:41.222 --> 0:40:44.182 what we know, but what we don't know. That is 0:40:44.182 --> 0:40:48.742 my definition of intelligence. Um, because I think I think 0:40:48.782 --> 0:40:51.342 it gets a little confusing because at least in the 0:40:51.342 --> 0:40:54.942 early schooling era, they're not really teaching you intelligence. They're 0:40:54.942 --> 0:40:57.502 teaching you knowledge. Uh, the teacher will just give you 0:40:57.502 --> 0:40:59.832 a fact, and you memorize that fact, then they'll give 0:40:59.832 --> 0:41:02.112 you a test and you repeat back the fact. And 0:41:02.112 --> 0:41:03.592 that is the metric that a lot of people are 0:41:03.592 --> 0:41:07.432 using for large language models. Yeah, they're doing like bullshit, like, oh, 0:41:07.472 --> 0:41:11.112 can it do the SATs? Can it do the bar exam? Like, 0:41:11.152 --> 0:41:13.032 of course it can do the bar exam. It's a 0:41:13.032 --> 0:41:17.472 database of all the answers. It can just regurgitate the answers. Now, 0:41:17.512 --> 0:41:21.032 my definition of intelligence would be okay. What happens when 0:41:21.032 --> 0:41:22.992 we put things that, like, we ask it to do 0:41:22.992 --> 0:41:26.592 things that aren't in its database, which is admittedly very 0:41:26.592 --> 0:41:29.632 hard for large language models because, well, they have all 0:41:29.632 --> 0:41:32.072 of human knowledge pretty much in there. They have like 0:41:32.072 --> 0:41:35.872 every book, every movie, every web page. So it's very 0:41:35.872 --> 0:41:39.552 hard to find a task where the AI cannot just 0:41:39.552 --> 0:41:42.512 use basic pattern matching and be. It isn't in their 0:41:42.512 --> 0:41:45.952 data set. And this is where I say AI is 0:41:45.952 --> 0:41:50.992 not in any like any definition of the word intelligent. Um, 0:41:50.992 --> 0:41:54.672 and the way I would assess this is by product. Um, 0:41:54.672 --> 0:41:57.162 if we ask the AI to do something that it 0:41:57.162 --> 0:41:59.362 already has in its database. It's very easily going to 0:41:59.362 --> 0:42:01.362 be able to do it because it already knows the answer. 0:42:01.682 --> 0:42:05.362 But then we have a lot of problems, like we 0:42:05.362 --> 0:42:08.802 as humans have a lot of unanswered questions. We have 0:42:08.802 --> 0:42:11.722 mathematical problems that have not yet been solved. We have 0:42:11.882 --> 0:42:16.202 incomplete scientific theories. Now, uh, I like to go to 0:42:16.202 --> 0:42:20.842 Einstein for this example, because Einstein lived before computers. Uh, 0:42:20.842 --> 0:42:23.002 in his day, he could read as many books as 0:42:23.002 --> 0:42:26.122 he could read. He couldn't Google things. He couldn't control 0:42:26.122 --> 0:42:28.362 F through a book. He would have to read a 0:42:28.362 --> 0:42:31.002 book or a scientific paper. And back then I think 0:42:31.002 --> 0:42:34.082 they were published in like, actual physical journals. So he 0:42:34.082 --> 0:42:37.002 would have had to gone and got a scientific journal 0:42:37.202 --> 0:42:39.562 and read through it to learn a little bit about 0:42:39.562 --> 0:42:43.922 a thing. And with as little like, I'm not going 0:42:43.962 --> 0:42:47.202 to say Einstein wasn't knowledgeable, but with as little knowledge 0:42:47.202 --> 0:42:50.362 as he had access to, he was able to create 0:42:50.402 --> 0:42:56.052 theories like special and general relativity, which revolutionized physics. Now, 0:42:57.092 --> 0:42:59.412 in physics classes today we work with a lot of 0:42:59.412 --> 0:43:03.412 his theories. And your average like, not very bright college 0:43:03.412 --> 0:43:07.412 student can can work with Einstein's theories because we already 0:43:07.412 --> 0:43:10.172 have the complete picture. He's already come up with them. 0:43:10.172 --> 0:43:12.212 He's already given us the answers and we can now 0:43:12.252 --> 0:43:15.492 work with them. But what made Einstein so amazing is 0:43:15.492 --> 0:43:17.692 he did not have the answers. He came up with 0:43:17.692 --> 0:43:23.292 them initially. So, um, kind of my point there would be, okay, 0:43:23.292 --> 0:43:26.652 so Einstein could come up with this amazing theory with very, 0:43:26.652 --> 0:43:30.892 very limited access to information. We have these machines that have, 0:43:31.612 --> 0:43:34.572 I don't even know how much trillions and trillions and 0:43:34.572 --> 0:43:37.732 trillions and trillions of data points like every book, every 0:43:37.732 --> 0:43:43.572 scientific paper, absolutely everything is in their knowledge database. But like, 0:43:43.612 --> 0:43:45.772 what have they come up with? If I ask it 0:43:45.772 --> 0:43:49.092 for a theory that unifies classical physics and quantum physics, 0:43:49.292 --> 0:43:51.772 it just shits out some existing theories that someone else 0:43:51.772 --> 0:43:55.742 came up with, and if it had any intelligence at all, 0:43:55.902 --> 0:43:58.902 with the sheer amount of knowledge it has access to 0:43:59.062 --> 0:44:01.702 and the sheer amount of computing power it has access to, 0:44:01.942 --> 0:44:03.822 I would expect it would be able to complete at 0:44:03.862 --> 0:44:06.662 least one of those theories. Like the fact that it 0:44:06.662 --> 0:44:09.742 just has so much knowledge, yet seems to have done 0:44:09.742 --> 0:44:14.502 nothing novel whatsoever. And in my opinion says that not 0:44:14.502 --> 0:44:17.822 only does it lack intelligence, it isn't intelligent at all. 0:44:18.622 --> 0:44:21.142 Yeah. So. So this is fascinating. This is where I 0:44:21.142 --> 0:44:23.862 wanted to get to. So I think a couple of 0:44:23.862 --> 0:44:28.502 times you've made this like horrible, horrible error. And this 0:44:28.502 --> 0:44:32.222 is why. This is why it's actually a risk to 0:44:32.262 --> 0:44:35.342 regular people. So when we were talking about computer science 0:44:35.342 --> 0:44:39.102 and programming, you used yourself as an example of being 0:44:39.102 --> 0:44:42.502 able to pivot. You have been on the front page 0:44:42.502 --> 0:44:49.462 of wired. You. You're an exceptional person. Um, malware analysts 0:44:49.822 --> 0:44:54.592 are also exceptional people within computer science and within security. 0:44:54.592 --> 0:44:58.472 So very few people in security can do malware analysis. 0:44:58.472 --> 0:45:01.752 So you're already exceptional and then exceptional within that group. 0:45:02.072 --> 0:45:06.952 And then you just defined intelligence giving the example of Einstein. 0:45:06.952 --> 0:45:10.792 Einstein is one of the smartest people, arguably the smartest 0:45:10.792 --> 0:45:16.032 person I would argue Newton. Uh, but anyway, um, one 0:45:16.032 --> 0:45:19.672 of the smartest people in the entire world that's ever existed. 0:45:20.352 --> 0:45:22.792 If you put the bar there, I agree with you. 0:45:22.792 --> 0:45:25.552 And I also want to just grant you the overall 0:45:25.552 --> 0:45:28.472 point of like, where are the novel discoveries? Here's my 0:45:28.472 --> 0:45:31.911 problem with this, Marcus. Like, that is not the bar 0:45:31.912 --> 0:45:34.352 that matters. The bar that matters is what's going to 0:45:34.352 --> 0:45:38.512 change society. The bar that matters is taking normal people's 0:45:38.872 --> 0:45:42.872 level of intelligence and getting above that bar. So you have, 0:45:42.912 --> 0:45:46.952 you know, John Smith, you know, working at some job. Again, 0:45:46.952 --> 0:45:49.002 I go down the list writing an essay about the 0:45:49.002 --> 0:45:52.762 book that you read, reading and answering emails, organizing a conference. 0:45:52.962 --> 0:45:55.962 Hundreds of millions of people are being paid a full 0:45:55.962 --> 0:46:01.042 living salary to do these jobs. Not very well. That 0:46:01.042 --> 0:46:04.442 level of intelligence, which I think you and I can agree, 0:46:04.482 --> 0:46:07.362 like it's not that great. And they are not Einstein. 0:46:07.362 --> 0:46:10.202 They are not writing malware. They are. The standard here 0:46:10.202 --> 0:46:13.082 is extremely low. The amount of critical thinking needed for 0:46:13.122 --> 0:46:15.762 that is extremely low. So what I'm talking about is 0:46:15.762 --> 0:46:20.082 an AI that can replace that level of intelligence. And 0:46:21.122 --> 0:46:24.762 I just think the the definition of inventing net new 0:46:24.762 --> 0:46:31.242 things is an unbelievably high like bar for intelligence. To 0:46:31.282 --> 0:46:34.522 me it has to be can you be hit with 0:46:34.522 --> 0:46:38.322 regular everyday problems and can you solve them using your 0:46:38.362 --> 0:46:41.642 knowledge of how the world works? Because another thing, just 0:46:41.642 --> 0:46:43.722 one last thing to say here. Every single one of 0:46:43.722 --> 0:46:48.492 those mundane tasks that I just gave, They're not actually 0:46:48.492 --> 0:46:52.572 just knowledge lookup. They actually require critical thinking for every 0:46:52.572 --> 0:46:54.852 single one of them because the email is not the same. 0:46:55.172 --> 0:46:57.252 The report that they're going to write is not the same. 0:46:57.972 --> 0:47:00.652 Each each one of those, even though it's relatively simple 0:47:00.652 --> 0:47:05.252 for a human, it's impossible for pre-human technology or pre 0:47:05.292 --> 0:47:09.012 AI technology that wasn't human. So it's in my mind 0:47:09.012 --> 0:47:12.012 it's 100% intelligence because it can't be automated. 0:47:13.572 --> 0:47:17.172 I think that's an entirely different argument. Well, there's always 0:47:17.172 --> 0:47:20.572 2 or 3 separate arguments going on here. There's first like, 0:47:20.612 --> 0:47:23.692 is it enough to replace the average worker and which 0:47:23.692 --> 0:47:26.052 you're sort of extrapolating and saying that is a definition 0:47:26.052 --> 0:47:28.972 of intelligence. But I would argue the average worker is 0:47:28.972 --> 0:47:32.972 not using their intelligence. They're working primarily with knowledge. Like 0:47:32.972 --> 0:47:35.652 if we were to take your average like John Smith 0:47:35.692 --> 0:47:39.732 office worker, and apply them to a task that was 0:47:39.772 --> 0:47:43.852 novel and did require intelligence, they would probably be able 0:47:43.852 --> 0:47:46.302 to do that because they have intelligence that is going 0:47:46.342 --> 0:47:49.182 to waste in that job, like they're being made to 0:47:49.222 --> 0:47:53.662 do essentially busy work, which is primarily knowledge based, not 0:47:53.662 --> 0:47:58.862 intelligence based. Now, I think what um, the kind of, uh, 0:47:59.102 --> 0:48:04.502 the miscommunication or the misunderstanding here is trying to attribute, uh, 0:48:04.902 --> 0:48:08.942 what requires intelligence for a human as intelligence for a 0:48:08.942 --> 0:48:11.902 large language model, because these large language models are doing 0:48:11.902 --> 0:48:16.022 pattern matching. While, sure, every email isn't the same, it's 0:48:16.022 --> 0:48:19.342 not thinking about the differences between emails. It simply just 0:48:19.342 --> 0:48:22.902 has a big database of here's all the different conversations 0:48:22.902 --> 0:48:26.422 and every conversation that has pretty much ever occurred. Like, 0:48:26.622 --> 0:48:29.582 almost every single conversation I've ever had in my life 0:48:29.822 --> 0:48:32.302 is a slight variation of a conversation that has already 0:48:32.302 --> 0:48:35.182 been had by someone else. Like there is nothing novel 0:48:35.222 --> 0:48:37.742 going on. So I would argue that that is not 0:48:37.742 --> 0:48:40.662 a sign of, uh, AI intelligence. It is a sign 0:48:40.662 --> 0:48:45.192 that you can, uh, you can almost emulate a a 0:48:45.352 --> 0:48:49.672 not fully applied intelligence with a very advanced pattern matching algorithm. 0:48:49.912 --> 0:48:50.552 And the reason? 0:48:50.592 --> 0:48:53.432 Can I jump in? Can I jump in real quick? Yeah. 0:48:53.472 --> 0:48:56.112 Sorry to interrupt. Um, I want to give you an 0:48:56.112 --> 0:49:01.472 example here. Couples therapy. So a couples therapist studies for 0:49:01.472 --> 0:49:03.632 whatever they have, a master's degree or a PhD, whatever 0:49:03.632 --> 0:49:06.392 they have. And a couple comes to them and says, 0:49:06.392 --> 0:49:08.472 you know, we're we're about to break up our marriage 0:49:08.472 --> 0:49:11.872 because of so and so problem. And they, they help 0:49:11.872 --> 0:49:14.272 them for 2 or 3 years listening to all their 0:49:14.272 --> 0:49:19.312 different problems. And then they're coaching them through all the psychology, 0:49:19.312 --> 0:49:23.232 the sociology, all the trauma stuff, whatever, whatever they're talking 0:49:23.232 --> 0:49:28.392 to them about cognitive behavior, behavioral therapy, all this stuff, Marcus, 0:49:28.392 --> 0:49:32.752 all that stuff is just knowledge. It's just knowledge. And 0:49:32.752 --> 0:49:36.312 guess what? That conversation with this couple is kind of 0:49:36.352 --> 0:49:39.272 just like the other conversation with the other couple. Like, 0:49:39.272 --> 0:49:42.632 there's not really anything net new. Can we really argue 0:49:43.152 --> 0:49:47.192 that what that marriage therapist is doing is not isn't 0:49:47.472 --> 0:49:50.552 requiring intelligence? Of course it does. Of course it does. 0:49:50.592 --> 0:49:53.552 And what I'm arguing is all this knowledge work. It 0:49:53.552 --> 0:49:56.272 does to a lesser degree, of course, and to a 0:49:56.272 --> 0:49:59.911 lesser degree than my friend who's a cardiologist, because those 0:49:59.912 --> 0:50:02.512 are like really difficult things. But when you were trying 0:50:02.512 --> 0:50:05.272 to plan a four day trip to Switzerland with a 0:50:05.272 --> 0:50:08.272 different type of family, and one has different food needs, 0:50:08.392 --> 0:50:11.112 and the weather is also different in Switzerland, it is 0:50:11.112 --> 0:50:13.752 a new it is a net new problem each time, 0:50:13.752 --> 0:50:17.712 even though the problem specifics look different than previous versions. 0:50:19.992 --> 0:50:22.232 I would say it's a net new problem to a 0:50:22.232 --> 0:50:27.032 specific human, but not humanity as a whole, because, um, but. 0:50:27.032 --> 0:50:30.472 That doesn't matter. Because that's that's what that's what we 0:50:30.512 --> 0:50:33.712 solve every day. It doesn't matter what's new to to 0:50:33.752 --> 0:50:37.112 humanity overall, day to day, we're being hit with regular 0:50:37.112 --> 0:50:40.842 everyday problems. And that that's the work that we have 0:50:40.842 --> 0:50:43.602 to do. We we can't leverage all of humanity for that. 0:50:43.602 --> 0:50:44.922 We have to solve it ourselves. 0:50:45.962 --> 0:50:48.682 Well, the I guess the problem is, are you trying 0:50:48.682 --> 0:50:50.922 to argue that AI is intelligent, or are you trying 0:50:50.922 --> 0:50:54.522 to argue that it could do the average or like 0:50:54.562 --> 0:50:57.602 any of the jobs that you've given? Because my argument. 0:50:57.602 --> 0:50:58.162 I'm arguing. 0:50:58.162 --> 0:51:02.002 Both. Okay. So I would agree with you that. Yeah. Um, 0:51:02.002 --> 0:51:05.082 a lot of therapy is just pattern matching. Uh, there 0:51:05.082 --> 0:51:08.042 are very, uh, I would use like an example like 0:51:08.042 --> 0:51:12.362 attachment theory. There's four different types of attachment style. Um, 0:51:12.402 --> 0:51:15.722 and everyone fits into one of those four groups. Um, 0:51:15.842 --> 0:51:19.082 and there's, uh, some very clearly defined rules of what 0:51:19.122 --> 0:51:22.642 tends to cause someone to become a certain attachment style. 0:51:23.002 --> 0:51:26.522 But every single situation is going to be slightly different. Um, 0:51:26.522 --> 0:51:30.122 but all of those situations map to a single rule 0:51:30.242 --> 0:51:33.642 that then maps to your specific attachment style. So essentially 0:51:33.642 --> 0:51:35.882 what a therapist is doing there is they're listening to 0:51:35.882 --> 0:51:38.612 you and they're trying to pick out like, okay, what 0:51:38.612 --> 0:51:41.572 are what are their traumas? Like how did their parents 0:51:41.572 --> 0:51:44.732 teach them? And then they're mapping that just to a rule. Now, 0:51:44.732 --> 0:51:47.332 as a human, that requires a lot more intelligence than 0:51:47.332 --> 0:51:49.732 it would require for a large language model, because the 0:51:49.732 --> 0:51:52.812 large language model has way more examples to go on. 0:51:52.852 --> 0:51:57.092 So what you're basically taking is um, uh, almost it's 0:51:57.092 --> 0:52:00.892 like a, a kind of hand in hand relationship. You can, 0:52:00.892 --> 0:52:05.172 to an extent, replace intelligence with knowledge and knowledge with intelligence. 0:52:05.372 --> 0:52:09.252 And with the large language model, you have infinitely more knowledge, 0:52:09.372 --> 0:52:12.972 which means it needs infinitely less intelligence to replace, uh, 0:52:13.212 --> 0:52:16.092 whatever job you want to say it's going to replace. 0:52:16.092 --> 0:52:19.532 But my argument is that's not a sign of AI intelligence. 0:52:19.532 --> 0:52:22.692 That is a sign of the AI's knowledge. I think 0:52:22.732 --> 0:52:26.652 in order to, uh, even demonstrate a small amount of intelligence, 0:52:26.972 --> 0:52:29.292 the bar should be a lot higher than for a human. 0:52:29.292 --> 0:52:32.372 And that's the reason why I went with the Einstein example. Because, sure, 0:52:32.692 --> 0:52:34.812 a lot of people consider Einstein to be the smartest 0:52:34.812 --> 0:52:38.262 human on earth. But he was working with a very, 0:52:38.262 --> 0:52:41.582 very small amount of information, like a very small amount 0:52:41.582 --> 0:52:43.822 of knowledge and a very small amount of data compared 0:52:43.822 --> 0:52:47.062 to an AI. So knowledge sort of acts as this 0:52:47.102 --> 0:52:51.462 sort of amplifier for intelligence. So if Einstein, with his 0:52:51.462 --> 0:52:55.622 amazing intelligence but very limited access to data, could accomplish 0:52:55.622 --> 0:53:00.022 such amazing feats, why can something with trillions and trillions 0:53:00.022 --> 0:53:04.022 and trillions of data points and books not really accomplish 0:53:04.022 --> 0:53:07.422 more than your average Joe, who's not even fully applying 0:53:07.422 --> 0:53:10.502 all the intelligence they have. They're just doing data entry 0:53:10.622 --> 0:53:14.342 or pattern matching. So my argument there is I'm not 0:53:14.342 --> 0:53:17.702 saying that current generation large language models might not be 0:53:17.742 --> 0:53:20.662 able to replace certain roles. I'm saying it is not 0:53:20.662 --> 0:53:24.062 evidence that they are intelligent. It is intelligent. It's evidence 0:53:24.062 --> 0:53:28.022 that their level of knowledge can act in place of intelligence. 0:53:28.022 --> 0:53:31.502 And this becomes important because there is a ceiling, there 0:53:31.502 --> 0:53:33.952 is a cap where you can't just shove in more 0:53:33.952 --> 0:53:35.512 and more knowledge and it just gets more and more 0:53:35.512 --> 0:53:38.552 intelligent or sorry, it's not getting intelligent. It's able to 0:53:38.592 --> 0:53:41.352 emulate intelligence more and more. At some point you hit 0:53:41.352 --> 0:53:44.272 a ceiling and we've already hit that ceiling because we 0:53:44.272 --> 0:53:47.752 have these machines that have all of this, like all 0:53:47.792 --> 0:53:50.952 of science in their database, and they can't do shit. 0:53:50.992 --> 0:53:54.312 Like they cannot complete a single scientific theory. 0:53:54.552 --> 0:53:56.512 Yeah. No, no, I love this. First of all, I 0:53:56.512 --> 0:53:59.112 just want to say I love this line that you 0:53:59.112 --> 0:54:03.272 were on. This is actually not being talked about anywhere. Um, well, 0:54:03.472 --> 0:54:05.392 I want to say in all the places that I'm 0:54:05.392 --> 0:54:07.672 looking and I'm looking a lot of different places, I 0:54:07.712 --> 0:54:11.472 love this point that you just made about how more 0:54:11.472 --> 0:54:15.392 and more knowledge is a cheat code to intelligence. It 0:54:15.432 --> 0:54:19.752 requires less intelligence. I don't think it matters. And and 0:54:19.752 --> 0:54:23.472 here's why. So. So first of all, there's, um, there's 0:54:23.472 --> 0:54:25.592 a bunch of post training stuff that gets done. And 0:54:25.592 --> 0:54:27.672 I'm not an expert on post training, but there's a 0:54:27.672 --> 0:54:31.192 bunch of stuff that you do with these models afterwards, 0:54:31.402 --> 0:54:36.122 and I believe that we are making massive inroads on that, 0:54:36.602 --> 0:54:42.082 what I call tricks. Um, basically tricking AI as a 0:54:42.082 --> 0:54:46.402 system overall to understand how to make these jumps. In 0:54:46.402 --> 0:54:49.802 other words, I think it is possible to teach AI 0:54:50.442 --> 0:54:55.762 how to go from Newtonian physics to Einsteinian theories, and 0:54:55.762 --> 0:54:58.122 I think it's possible to teach it, to generalize, to 0:54:58.162 --> 0:55:00.322 be able to do that. And there's already some evidence 0:55:00.322 --> 0:55:04.442 of this working. I had a thing in the show recently. Um, basically, 0:55:04.442 --> 0:55:10.082 researchers have been working on this one problem, um, with bacteria, bacteriophages, 0:55:10.402 --> 0:55:13.322 which are basically viruses that propagate and try to take 0:55:13.322 --> 0:55:16.562 over bacteria. And so, um, they've been struggling with this 0:55:16.562 --> 0:55:20.042 for like over a decade or maybe two decades. The 0:55:20.042 --> 0:55:24.722 absolute pinnacle researchers in this particular thing, they gave it 0:55:24.722 --> 0:55:27.762 to this new, uh, Google model, which is specifically designed, 0:55:27.922 --> 0:55:30.252 designed to do exactly what you're talking about, Marcus. To 0:55:30.292 --> 0:55:33.252 actually find net new things. And it came back and said, 0:55:33.252 --> 0:55:36.411 here's my hypothesis for why this is happening. They looked 0:55:36.412 --> 0:55:39.572 at it and they said, Holy crap. That is the answer. 0:55:39.652 --> 0:55:42.052 It's a net new answer. They went and tested it. 0:55:42.092 --> 0:55:46.212 It was 100% confirmed. And there's a bunch of more 0:55:46.252 --> 0:55:49.292 examples of this. There's companies that are doing this basically 0:55:49.452 --> 0:55:54.612 harvesting like research that dormant research, raw research and coming 0:55:54.612 --> 0:55:57.332 up with net new hypotheses. So I think that's just 0:55:57.332 --> 0:55:59.812 a matter of like, we just haven't done the work 0:55:59.972 --> 0:56:06.092 of the scaffolding of teaching the AI how to make jumps, right. 0:56:06.132 --> 0:56:08.292 Because like you said, it's been so easy to just 0:56:08.292 --> 0:56:10.692 pull from the knowledge. But I think this is a 0:56:10.692 --> 0:56:12.852 training step that we could do. And this is an 0:56:12.852 --> 0:56:16.172 empirical thing. Like it'll either work or it doesn't. And 0:56:16.172 --> 0:56:20.092 I can send you over the stuff. But the bigger 0:56:20.092 --> 0:56:24.812 point here for me is that, um, this is why 0:56:24.812 --> 0:56:27.132 this matters to me is because of humans. So if 0:56:27.132 --> 0:56:29.142 we go back to the point of like, don't worry 0:56:29.142 --> 0:56:32.382 about I if I could, you know, summarize your thing 0:56:32.422 --> 0:56:35.022 about that saying there's always going to be a place 0:56:35.022 --> 0:56:37.502 for humans. This thing is not a big deal. Um, 0:56:37.502 --> 0:56:41.662 I think you've even called it like, um, auto autocorrect. 0:56:41.702 --> 0:56:45.342 No autocomplete. Right? You call it autocomplete? I'm like, no, 0:56:45.342 --> 0:56:49.662 it is not autocomplete. It's actually doing this work. My 0:56:49.662 --> 0:56:54.902 point to you, Marcus, is my friends cardiology practice is 0:56:54.902 --> 0:57:00.942 nothing but pattern matching. Okay. Most high level work, um, 0:57:00.942 --> 0:57:05.302 is just pattern matching. This, um, couples therapist. They are 0:57:05.302 --> 0:57:07.902 also doing pattern matching. And if you look at the 0:57:07.902 --> 0:57:11.022 average job, they are also doing pattern matching like okay, 0:57:11.062 --> 0:57:14.342 it's we're taking emails. We're writing a report. You can 0:57:14.342 --> 0:57:18.782 reduce all this work that hundreds of millions of people 0:57:18.782 --> 0:57:20.782 are doing. You could reduce that to pattern matching if 0:57:20.782 --> 0:57:26.072 you want. But we're talking about hundreds of millions of jobs, Right? 0:57:26.072 --> 0:57:29.472 So if I could do that work, then it still 0:57:29.472 --> 0:57:31.992 has the impact that I'm concerned about, whatever we want 0:57:31.992 --> 0:57:32.592 to call it. 0:57:34.752 --> 0:57:38.752 I don't disagree, but then the question is like, why 0:57:38.752 --> 0:57:41.712 hasn't it? Are we are we not there yet? Um. 0:57:41.752 --> 0:57:42.112 Are we? 0:57:42.152 --> 0:57:44.912 Because it's just now getting started. It's just now getting started. 0:57:44.952 --> 0:57:47.592 Like the tech is just now getting to the companies. 0:57:47.592 --> 0:57:49.632 These companies don't even know what AI is. They don't 0:57:49.632 --> 0:57:52.272 know what it isn't. It takes time to spin it up. 0:57:52.472 --> 0:57:55.792 Like we see the experiment starting. We see thousands of 0:57:55.792 --> 0:57:58.512 companies trying to adopt it, but they're trying to figure 0:57:58.512 --> 0:58:00.152 out what it is at the same time that they're 0:58:00.152 --> 0:58:02.112 trying to adopt it. So it's just a matter of 0:58:02.152 --> 0:58:04.682 it's just a matter of one, two, three, 4 or 0:58:04.682 --> 0:58:08.312 5 years in my opinion. And it's already happening. We 0:58:08.312 --> 0:58:10.632 already have evidence that it's happening. It's just a matter 0:58:10.632 --> 0:58:11.072 of time. 0:58:12.472 --> 0:58:16.872 I'm not sure I agree because like companies have, as 0:58:16.872 --> 0:58:18.912 you said, they've been in a rush to roll out AI. 0:58:19.312 --> 0:58:21.832 But then the question is like, if it is so 0:58:21.832 --> 0:58:25.122 close to being able to replace the average knowledge worker. 0:58:25.402 --> 0:58:29.602 Why aren't we seeing any massive replacement? Why aren't we 0:58:29.642 --> 0:58:33.642 seeing any change in the productivity of companies? Because, like, 0:58:33.642 --> 0:58:37.922 we have objective metrics for those things. Um, and one 0:58:37.962 --> 0:58:40.562 sort of tangent I'd want to go on here is 0:58:40.562 --> 0:58:43.322 a lot of industries aren't finite. Like if you think 0:58:43.322 --> 0:58:45.962 of something like, say, farming, there's only so much food 0:58:45.962 --> 0:58:48.162 we need. Uh, if we were able to automate all 0:58:48.202 --> 0:58:50.802 the farming, sure, farmers could go away. Uh, we don't. 0:58:50.882 --> 0:58:54.602 There's no value to. Hey, let's re allocate the farmers 0:58:54.602 --> 0:58:56.762 to figuring out how to make even more food. Like, 0:58:56.762 --> 0:58:59.642 once we've got enough food to feed everyone, we're good. Um, 0:58:59.642 --> 0:59:02.642 but then you have industries which is also overlap with 0:59:02.642 --> 0:59:06.322 the industries that are most invested in AI and are 0:59:06.322 --> 0:59:10.322 most trying to replace workers with AI like tech. Now, 0:59:10.362 --> 0:59:14.082 tech is not a finite industry. There is infinite software 0:59:14.082 --> 0:59:17.162 you can write. There are infinite, uh, feature improvements. There's 0:59:17.162 --> 0:59:21.162 infinite patches now, uh, like, uh, I guess the example 0:59:21.162 --> 0:59:24.332 I give is Microsoft is not a search engine company, 0:59:24.332 --> 0:59:25.852 but they felt the need to come out with a 0:59:25.852 --> 0:59:28.572 search engine. Google is not an operating system company, but 0:59:28.572 --> 0:59:30.852 they felt the need to come up with an operating system. 0:59:31.132 --> 0:59:34.132 And why? Because the more products you make, the more 0:59:34.132 --> 0:59:36.612 market share you can capture, and the more you can 0:59:36.612 --> 0:59:40.972 compete in, uh, in the global market. So from a 0:59:40.972 --> 0:59:44.772 tech company perspective, uh, let's say I can right now 0:59:44.772 --> 0:59:48.452 replace all of my software engineers with AI. Um, the 0:59:48.452 --> 0:59:50.332 illogical thing to do would be to lay off all 0:59:50.332 --> 0:59:53.412 my engineers and do the same thing that I'm already doing, 0:59:53.412 --> 0:59:56.252 but with AI, because all of the companies that are 0:59:56.252 --> 0:59:58.812 smart are just going to do more. They're going to 0:59:58.852 --> 1:00:01.452 just start breaking into all the other industries that all 1:00:01.492 --> 1:00:04.852 their competitors are in, and they're going to dominate every space. 1:00:05.012 --> 1:00:08.692 So the logical path, uh, if I could replace employees, 1:00:08.852 --> 1:00:11.892 would not be laying off your employees. It would be 1:00:11.892 --> 1:00:14.412 to just expand. It's to have your employees do more 1:00:14.412 --> 1:00:16.892 and more things, do more and more products, and capture 1:00:16.892 --> 1:00:20.292 more and more market share. Now, if that was happening, 1:00:20.572 --> 1:00:23.172 we would see an increase in GDP, because GDP is 1:00:23.172 --> 1:00:26.532 the total value of all of the goods and products 1:00:26.532 --> 1:00:30.372 produced within the country. Yet we're not seeing any any 1:00:30.372 --> 1:00:33.132 change in GDP. We're not seeing any change in productivity. 1:00:33.572 --> 1:00:35.612 All of these companies who are claiming that they're going 1:00:35.612 --> 1:00:39.012 to replace all their engineers with AI, we've not seen 1:00:39.012 --> 1:00:44.892 any change in anything. Um, so firstly, I wouldn't especially 1:00:44.892 --> 1:00:47.652 in tech, I would not be worried about employees being 1:00:47.652 --> 1:00:52.372 just replaced, uh, because, um, well, firstly, AI doesn't fully 1:00:52.372 --> 1:00:57.532 replace employee. It accelerates their productivity. So, uh, doing the 1:00:57.532 --> 1:01:00.972 same amount of productivity but with less employees is far 1:01:00.972 --> 1:01:04.972 less desirable than keeping your employees and just doing more. Um, 1:01:04.972 --> 1:01:07.532 but we've not really seen either. We haven't seen employees 1:01:07.532 --> 1:01:10.172 being replaced with AI, and we haven't seen this massive 1:01:10.172 --> 1:01:13.852 explosion in productivity from tech companies. In fact, they're doing 1:01:13.892 --> 1:01:18.222 the opposite. They're laying off their employees to focus on AI. 1:01:18.582 --> 1:01:21.502 They're like, oh, we don't even have enough money to do, uh, 1:01:21.542 --> 1:01:24.702 whatever it was. Like whatever products Google has been, they're 1:01:24.742 --> 1:01:27.502 like binning half of their products and services to go 1:01:27.502 --> 1:01:31.222 and focus on AI. Whereas if I was actually capable 1:01:31.222 --> 1:01:33.382 of what they say it's capable of, they would be 1:01:33.382 --> 1:01:37.862 going in the opposite direction. They'd be making a gaming console, 1:01:37.862 --> 1:01:41.022 they'd be making a desktop operating system better. They would 1:01:41.022 --> 1:01:44.222 be cornering like every single market. And they're not. They're 1:01:44.222 --> 1:01:45.382 actually cutting down. 1:01:46.262 --> 1:01:50.182 Yeah, yeah, yeah. Interesting point, I hear you. Um, I 1:01:50.182 --> 1:01:53.142 think the reason it's not affecting GDP yet is because 1:01:53.222 --> 1:01:56.062 it's it's not rolled out. I mean, this is just 1:01:56.062 --> 1:02:00.422 starting like last year. We weren't even talking about agents yet, 1:02:00.422 --> 1:02:02.342 which is kind of like the the way we try 1:02:02.382 --> 1:02:06.062 to get into, um, all this automation. Um, and now 1:02:06.062 --> 1:02:08.462 agents are just now starting to get serious. My, my 1:02:08.502 --> 1:02:12.662 estimate for this has always been, uh, before 29. So 1:02:12.702 --> 1:02:14.302 I my I think this is going to be like 1:02:14.302 --> 1:02:20.032 a 20, 27 thing when this AGI, um, worker replacement 1:02:20.232 --> 1:02:23.152 technology is good enough to actually start replacing workers and 1:02:23.152 --> 1:02:24.632 it could be way sooner than that. It could be 1:02:24.632 --> 1:02:28.312 this year. It could be whenever, um, that it's not 1:02:28.312 --> 1:02:30.952 until that thing gets rolled out and it starts getting 1:02:30.952 --> 1:02:35.232 implemented by the thousands, by many, many different companies that 1:02:35.232 --> 1:02:38.112 we're actually going to see a GDP bump. So I 1:02:38.152 --> 1:02:41.632 would anticipate that being in like 28, 29, 30 and 1:02:41.632 --> 1:02:45.272 into the 30s, because that's just a slow like massive 1:02:45.272 --> 1:02:50.992 ramp up. Um, and the other thing, um, I can't 1:02:50.992 --> 1:02:52.792 remember your other point. What was the other point that 1:02:52.792 --> 1:02:54.872 you made about, uh, other than the GDP. 1:02:56.552 --> 1:02:59.552 Um, that it would make sense to, uh, not lay 1:02:59.552 --> 1:03:02.112 off employees, but to increase productivity? 1:03:02.392 --> 1:03:05.752 Yeah, yeah. So the difference there, the reason that doesn't 1:03:05.792 --> 1:03:10.392 work is because, um, an employee might cost like, let's 1:03:10.392 --> 1:03:14.722 just call it $200,000 with benefits, but for Depending on 1:03:14.722 --> 1:03:18.402 the level, that could be 3 or $400,000. Well, their 1:03:18.402 --> 1:03:22.562 entire contract with the AI company might be $200,000 if 1:03:22.562 --> 1:03:28.642 they can spin up 40,000 of these employees for $200,000 1:03:28.642 --> 1:03:33.762 instead of the 500 human employees. Then I think they 1:03:33.762 --> 1:03:37.642 would start with getting rid of the the previous ones. Obviously, 1:03:37.642 --> 1:03:39.602 they should do a slow thing, but I think the 1:03:39.602 --> 1:03:41.762 natural thing they're going to do, just based on what 1:03:41.762 --> 1:03:45.082 the CFO says is, yeah, we have these 500 people 1:03:45.442 --> 1:03:49.282 making 400 grand or 300 grand. Yeah, we need to 1:03:49.322 --> 1:03:52.722 phase them out. Keep the top 1%, keep the top 10%, 1:03:52.882 --> 1:03:54.682 and they're going to be the ones, you know, helping 1:03:54.682 --> 1:03:58.842 us move into other areas. But all. Net new actual, um, 1:03:58.842 --> 1:04:01.522 software engineers are going to be this other thing which 1:04:01.522 --> 1:04:04.282 we're already paying for the subscription, which. So it's it's 1:04:04.322 --> 1:04:08.282 like no marginal or. Yeah, the marginal cost is virtually 1:04:08.282 --> 1:04:11.282 zero to add new software engineers. 1:04:12.572 --> 1:04:15.452 But I think that rests on the assumption that the, uh, 1:04:15.532 --> 1:04:19.412 the AI agent can wholly replace an engineer, which is, 1:04:19.412 --> 1:04:21.612 as I said, is something I don't believe will ever, 1:04:21.612 --> 1:04:25.492 ever happen. I think what we will see is maybe 1:04:25.492 --> 1:04:27.772 I'm not even convinced of this yet, but I think 1:04:27.772 --> 1:04:30.732 maybe we will see these AI models get to the 1:04:30.732 --> 1:04:33.732 point where they can actually accelerate the productivity of a 1:04:33.732 --> 1:04:36.132 software engineer. But at the end of the day, there 1:04:36.132 --> 1:04:38.292 is always going to be a human on the end. 1:04:38.412 --> 1:04:41.892 It's just an abstraction. It's like, uh, how when we 1:04:41.892 --> 1:04:44.132 used to write things in machine code, it was super 1:04:44.132 --> 1:04:46.932 slow scrolling ones and zeros on a punch card. And 1:04:46.932 --> 1:04:48.852 then we made assembly language, and that made things a 1:04:48.852 --> 1:04:51.412 bit more productive. And then we made a C, and 1:04:51.412 --> 1:04:54.252 then we made Python. And now you can write whole 1:04:54.252 --> 1:04:57.572 software suites with these like point and click applications. But 1:04:57.612 --> 1:04:59.412 at the end of the day there is always a 1:04:59.412 --> 1:05:02.172 human on the end of that, it is just an 1:05:02.612 --> 1:05:06.052 like an abstraction and an acceleration of a human employee. 1:05:06.292 --> 1:05:09.052 Whereas I think the assumption you're working from here is 1:05:09.052 --> 1:05:11.142 that we get to a point where we can just 1:05:11.142 --> 1:05:14.222 wholly replace the human, which, uh, for the same reason 1:05:14.222 --> 1:05:16.822 I said AI is not intelligent. I don't think we 1:05:16.822 --> 1:05:20.462 can do, because it will always lack that ability to 1:05:20.502 --> 1:05:23.422 fill in the gaps, to come up with answers when 1:05:23.422 --> 1:05:25.982 not all the pieces are there. Like, I don't know 1:05:26.062 --> 1:05:28.942 how much you've done software engineering, but a lot of 1:05:28.942 --> 1:05:31.702 it is like the client doesn't even know what they want. Like, 1:05:31.702 --> 1:05:34.382 how do you prompt ChatGPT to build code when the 1:05:34.382 --> 1:05:37.222 client isn't even sure what it is they want to build? Um, 1:05:37.222 --> 1:05:39.582 and that's always been a large part of software engineering. 1:05:39.622 --> 1:05:43.622 It's not the writing code, which, uh, admittedly it's iffy, 1:05:43.622 --> 1:05:46.542 but you can to an extent do with large language models. 1:05:46.702 --> 1:05:49.742 It's the making the design decisions. It's like, what kind 1:05:49.742 --> 1:05:52.022 of server infrastructure do you want? How do you want 1:05:52.022 --> 1:05:54.262 the applications to talk to each other? And those are 1:05:54.262 --> 1:05:57.462 not decisions that an agent can make. Their decisions that 1:05:57.462 --> 1:05:59.862 need to be made by a human and the human 1:05:59.862 --> 1:06:02.222 making the decisions doesn't even know what decisions they want 1:06:02.262 --> 1:06:02.782 to make. 1:06:03.182 --> 1:06:07.182 Yeah, but so so check this out. Um, I want 1:06:07.222 --> 1:06:10.152 to use your previous points kind of to counter that argument, 1:06:11.032 --> 1:06:15.352 making those server choices and making those customer choices, those 1:06:15.352 --> 1:06:19.752 conversations are exactly the same as the marriage therapists, okay. 1:06:20.192 --> 1:06:25.992 They're they're basing that on good fundamental principles of building 1:06:25.992 --> 1:06:31.152 good applications with servers and applications and network connectivity and 1:06:31.192 --> 1:06:34.152 authentication and security and privacy and all these different things. 1:06:34.152 --> 1:06:36.712 These are fundamental rules. These are like written down and 1:06:36.712 --> 1:06:39.992 you can debate them or whatever. But fundamentally, AI is 1:06:39.992 --> 1:06:41.912 going to be doing the same exact thing that they're 1:06:41.912 --> 1:06:47.392 doing for like writing emails and sending emails and summarizing documents. Fundamentally, 1:06:47.392 --> 1:06:51.552 we're not making a Einsteinian jump here. We're talking about 1:06:51.552 --> 1:06:55.192 teaching an AI. What a fundamental building of an application 1:06:55.192 --> 1:06:58.792 infrastructure looks like, and what good software engineering looks like. 1:06:58.912 --> 1:07:01.752 That is a knowledge base. So the fact that we 1:07:01.792 --> 1:07:03.752 haven't done that yet and we're still stuck in this 1:07:03.752 --> 1:07:06.922 vibe coding land, that's because vibe coding started about 48 1:07:06.922 --> 1:07:11.442 seconds ago in I time, right? It started just now. 1:07:11.762 --> 1:07:15.322 So the vibe coding or the the eyes are not 1:07:15.322 --> 1:07:18.442 good at doing the software engineering piece yet. But it is. 1:07:18.482 --> 1:07:22.482 It is the same. It's not any harder than the cardiologist, 1:07:22.882 --> 1:07:25.322 which is also a knowledge base, or the marriage therapist, 1:07:25.322 --> 1:07:27.602 which is also a knowledge base. It's just a matter 1:07:27.602 --> 1:07:30.642 of time before this stuff becomes more capable. Now in 1:07:30.642 --> 1:07:32.962 terms of like, well, there'll always be a human in 1:07:32.962 --> 1:07:36.282 the loop. I mean, I think the higher level you go, 1:07:36.322 --> 1:07:41.522 the more advanced you go. Like, I think there's going 1:07:41.522 --> 1:07:44.082 to be humans in the loop over agent farms that 1:07:44.082 --> 1:07:46.722 are doing things, and the human's going to be applying 1:07:47.162 --> 1:07:49.722 what I believe to be the most irreplaceable thing, which 1:07:49.722 --> 1:07:54.762 is like taste and judgment and like preference, uh, because 1:07:54.802 --> 1:07:57.642 eyes are not alive. They don't have their own opinions. 1:07:58.082 --> 1:07:59.682 And so I think that's going to be the kind 1:07:59.682 --> 1:08:01.882 of like spiritual shaping these things are going to be 1:08:01.882 --> 1:08:04.892 putting into these agent farms. But in terms of execution, 1:08:04.892 --> 1:08:07.412 of writing the code and making sure it's on a 1:08:07.412 --> 1:08:10.652 good infrastructure that's secure. That, to me is all knowledge 1:08:10.652 --> 1:08:11.332 based stuff. 1:08:12.892 --> 1:08:15.732 I would actually push back and say that I completely 1:08:15.732 --> 1:08:19.972 disagree because unlike with the past examples you've given me, 1:08:20.212 --> 1:08:25.092 tech moves insanely fast. Like we have, uh, there's probably 1:08:25.092 --> 1:08:28.532 been 15 new database, uh, frameworks that have come out 1:08:28.532 --> 1:08:31.892 since I started programming. And back in the day, it 1:08:31.892 --> 1:08:34.732 would be buying everything into an SQL database. And that 1:08:34.732 --> 1:08:37.772 was horrible for the most the majority of applications. And 1:08:37.772 --> 1:08:39.412 then we started coming out with all these new types 1:08:39.412 --> 1:08:42.572 of databases. We came out with NoSQL and all of 1:08:42.572 --> 1:08:47.852 these unstructured, um, like query based. Um, uh, I don't 1:08:47.852 --> 1:08:49.252 even want, I just want to call it like a 1:08:49.252 --> 1:08:52.772 data bin. And then we had like AWS buckets, and 1:08:52.812 --> 1:08:57.171 then we're constantly, uh, inventing new technologies. And the problem 1:08:57.172 --> 1:09:00.372 there is twofold. The first is there is no set 1:09:00.372 --> 1:09:03.782 of rules because we're inventing new technologies at a rate 1:09:03.782 --> 1:09:07.502 in which there it's it's basically it is opinions, it's decisions. 1:09:07.742 --> 1:09:10.142 And there's no hard and fast rule of if this 1:09:10.142 --> 1:09:14.822 then that for database or server design. And also uh, 1:09:15.262 --> 1:09:18.902 unlike with, say, the marriage therapist, almost everything that has 1:09:18.902 --> 1:09:21.702 ever happened to you has happened to someone else. Like, 1:09:22.182 --> 1:09:24.262 like a lot of people would like to believe that 1:09:24.302 --> 1:09:27.502 their experience is, like, unique and no one else has 1:09:27.502 --> 1:09:30.662 lived their life and no one else has their problems. 1:09:30.822 --> 1:09:34.502 But in reality, your average human is like a hundred. 1:09:34.542 --> 1:09:36.422 Other people have been through the same stuff that they 1:09:36.462 --> 1:09:39.822 have been through. Whereas with technology we're trying to build new, 1:09:39.982 --> 1:09:43.062 new applications. We're trying to build stuff that hasn't already 1:09:43.062 --> 1:09:47.142 been built. So if we would say just, uh, confined 1:09:47.142 --> 1:09:50.782 within only the software that already exists, sure. You could 1:09:50.782 --> 1:09:53.182 just use rules, but then what would be the value 1:09:53.182 --> 1:09:55.302 of that? Like what is the value of, hey, we 1:09:55.302 --> 1:09:57.782 can just build software we already have like, oh, I 1:09:57.782 --> 1:10:00.062 can build a HTTP server, why don't I just download 1:10:00.062 --> 1:10:04.272 Apache or nginx. Um, so that is a problem. And 1:10:04.272 --> 1:10:06.992 then the second thing is, uh, in order for the 1:10:06.992 --> 1:10:10.392 large language models to actually make, uh, even emulate making 1:10:10.392 --> 1:10:13.752 those decisions, they need knowledge of those frameworks, which means 1:10:13.752 --> 1:10:16.232 they have to be trained on that data. So someone 1:10:16.232 --> 1:10:19.232 has to go and they have to create massive amounts 1:10:19.232 --> 1:10:21.671 of data to feed into the large language model, because 1:10:21.672 --> 1:10:24.592 it's almost like, uh, I'd call it like a lossy 1:10:24.592 --> 1:10:27.912 compression algorithm. Uh, you can't just put like a really 1:10:27.912 --> 1:10:31.392 good paper on database design into a large language model. 1:10:31.392 --> 1:10:33.032 And now your large language model is really good at 1:10:33.032 --> 1:10:36.552 database design. You need like, thousands and millions of data 1:10:36.552 --> 1:10:39.312 points in order for it to even be mediocre at 1:10:39.352 --> 1:10:41.832 that decision making. So now we're going to have a 1:10:41.832 --> 1:10:44.552 lag where like a new technology will come out. And 1:10:44.552 --> 1:10:47.631 as a human with like actual intelligence, but a lack 1:10:47.632 --> 1:10:49.592 of knowledge, because this technology is new and I'm not 1:10:49.592 --> 1:10:52.392 an expert in it, I can read the documentation and 1:10:52.392 --> 1:10:55.392 I can understand it. I can fill in the missing pieces. 1:10:55.392 --> 1:10:57.592 I can make decisions about how that might pertain to 1:10:57.592 --> 1:11:01.072 my software. but the large language model. Now, it doesn't 1:11:01.072 --> 1:11:03.472 have any intelligence to begin with, and it doesn't have 1:11:03.472 --> 1:11:05.152 the knowledge because it's not being put out there on 1:11:05.152 --> 1:11:07.512 the internet yet for it to be trained on. So 1:11:07.512 --> 1:11:10.552 now we just have this massive lag where an actual 1:11:10.552 --> 1:11:13.512 human developer is going to be better than the AI 1:11:13.552 --> 1:11:15.552 because they're going to know more and they're going to 1:11:15.552 --> 1:11:18.312 be able to work with newer technologies better. So I 1:11:18.312 --> 1:11:21.472 would really push back on that, especially with software engineering, 1:11:21.672 --> 1:11:24.152 because a lot of fields are very static. It's like 1:11:24.152 --> 1:11:29.032 things haven't changed in like decades or millennia. Whereas technology 1:11:29.072 --> 1:11:31.112 like in the time we've been recording this video, there's 1:11:31.112 --> 1:11:33.192 probably like 15 new technologies. I'm going to have to 1:11:33.192 --> 1:11:35.312 go and learn. So I would. 1:11:35.352 --> 1:11:39.192 Disagree. I hear you there. I hear you there. Um, 1:11:39.192 --> 1:11:43.952 let me respond to that. Um, so essentially, who do 1:11:43.952 --> 1:11:45.872 you think is going to be better at learning a 1:11:45.872 --> 1:11:49.792 new tech? Uh, so so I think you're incorrect about 1:11:50.352 --> 1:11:52.751 you need to go and get, um, you know, thousands 1:11:52.752 --> 1:11:56.112 or millions of examples, uh, what people are currently doing, uh, 1:11:56.232 --> 1:11:58.522 in state of the art of building this stuff now, 1:11:58.802 --> 1:12:00.482 and I'm pretty sure you and I could just test 1:12:00.482 --> 1:12:04.002 this offline afterwards. Um, we could potentially make, like, a 1:12:04.002 --> 1:12:07.842 fake language, a fake new programming language, also using AI, 1:12:08.162 --> 1:12:11.762 and then say, rewrite this working application in this new 1:12:11.762 --> 1:12:15.602 application in this new, um, programming language, and you give 1:12:15.602 --> 1:12:18.282 it a full programming language spec and it could just 1:12:18.282 --> 1:12:22.522 port automatically over. So I think you're making my point 1:12:22.522 --> 1:12:26.041 for me. So if you think about the concept of 1:12:26.082 --> 1:12:31.802 creating a. Net new application, um, going back to human psychology, um, 1:12:31.922 --> 1:12:36.082 and going back to philosophical principles, think about this. There 1:12:36.122 --> 1:12:38.322 are only so many net new things that you can make. 1:12:38.362 --> 1:12:40.082 And they're going to be oriented. They're going to have 1:12:40.082 --> 1:12:43.842 the shape, the pothole shape of human problems. So like 1:12:43.882 --> 1:12:46.002 you're not going to make something up that is net 1:12:46.042 --> 1:12:48.602 new to a computer that it that the AI has 1:12:48.602 --> 1:12:50.802 never heard of before. You're going to make things like, 1:12:50.842 --> 1:12:52.322 oh it's going to be a game. Oh it's going 1:12:52.322 --> 1:12:54.282 to be an application where I submit this and get 1:12:54.282 --> 1:12:57.171 this back. And when you show it a new programming language, 1:12:57.172 --> 1:12:59.572 it's not going to be like, Holy crap, I've never 1:12:59.572 --> 1:13:01.532 seen that before. You've blown my mind. It's going to 1:13:01.532 --> 1:13:04.812 be like, are you kidding me? I know all programming languages, 1:13:04.812 --> 1:13:07.412 and I see what you've done here with the spec. Okay, cool. Yeah, 1:13:07.412 --> 1:13:11.252 I could use that. So if there are 25 new 1:13:11.252 --> 1:13:15.572 programming paradigms and programming languages and programming specs that come 1:13:15.572 --> 1:13:18.412 out while we're doing this, the advantage of these new 1:13:18.572 --> 1:13:23.291 software engineers, uh, things, they're they're parsing those all the time. 1:13:23.292 --> 1:13:26.012 That's just part of the agent infrastructure to constantly be 1:13:26.012 --> 1:13:29.492 re ingesting these new languages. And here's the crazy part. 1:13:29.532 --> 1:13:32.652 If it hadn't done that, then when you say just 1:13:32.692 --> 1:13:35.211 I want you to program in, uh, you know, booga 1:13:35.212 --> 1:13:37.252 booga or whatever it is, it's like, I don't know 1:13:37.252 --> 1:13:39.532 what that is. Hold on. Give me a second, okay? 1:13:39.532 --> 1:13:42.652 I just consumed everything anyone's ever said about it because 1:13:42.652 --> 1:13:44.972 I live crawled the internet about it. I read all 1:13:44.972 --> 1:13:48.171 the forums, I read the entire programming spec. I brought 1:13:48.172 --> 1:13:50.252 that in. That's now part of my knowledge. Would you 1:13:50.252 --> 1:13:52.052 like me to rewrite this in? Booga booga. 1:13:53.022 --> 1:13:56.302 I don't, uh, so I don't agree for multiple reasons. 1:13:56.302 --> 1:14:01.982 The first is that I think we're talking more about, um, uh, 1:14:02.382 --> 1:14:05.022 I don't know what actually is the correct term for it, 1:14:05.022 --> 1:14:07.662 but you have two functionalities to a large language model. 1:14:07.662 --> 1:14:10.982 You have the actual training database where they've ingested the data, 1:14:10.982 --> 1:14:14.622 they've run it through their training system, they've done their, uh, 1:14:14.622 --> 1:14:17.902 human like reinforcement learning on the data. And that's how 1:14:17.902 --> 1:14:20.822 you get the AI being able to do something. Now, 1:14:20.822 --> 1:14:22.622 on top of that, it can search the internet, it 1:14:22.622 --> 1:14:25.182 can grab some new data. And then based on its 1:14:25.182 --> 1:14:30.262 existing knowledge base, it can to an extent parse that data. 1:14:30.262 --> 1:14:32.542 But I would argue that firstly, like even something as 1:14:32.542 --> 1:14:35.501 simple as a new programming language would not be something 1:14:35.502 --> 1:14:38.062 that it could just go out, fetch the spec, and 1:14:38.062 --> 1:14:41.142 then using its existing database, be able to write in 1:14:41.182 --> 1:14:43.142 that language. I think it would have to be not 1:14:43.142 --> 1:14:45.822 just retrained, but also it would have to go through 1:14:45.822 --> 1:14:49.582 the human, uh, reinforcement learning process. And then my second 1:14:49.582 --> 1:14:52.552 point is I've actually done this, um, not even with 1:14:52.552 --> 1:14:55.352 a new programming language, but, uh, there's a niche programming 1:14:55.352 --> 1:14:58.152 language that I have to use quite a lot. And 1:14:58.152 --> 1:15:01.392 for the life of me, I cannot get any large 1:15:01.392 --> 1:15:04.432 language model to write functional code in that language, because 1:15:04.432 --> 1:15:07.352 there's just not enough data points about it. I can 1:15:07.352 --> 1:15:09.352 go on to a forum, I can grab a script 1:15:09.352 --> 1:15:12.352 and it'll work. But whenever I ask the large language 1:15:12.352 --> 1:15:15.392 model to make me something in that language, it just 1:15:15.392 --> 1:15:17.631 falls flat on its face like it cannot do it. 1:15:17.912 --> 1:15:20.912 So like based on my experience and we're talking something 1:15:20.912 --> 1:15:24.022 as simple as a programming language, which can be 1 1:15:24.022 --> 1:15:27.112 to 1 mapped to an existing programming language, when we 1:15:27.112 --> 1:15:30.352 talk about like brand new database frameworks, it's not like, oh, 1:15:30.352 --> 1:15:33.672 this is just SQL, but the words are different. Like, 1:15:33.672 --> 1:15:37.072 this is an entirely new framework with entirely new, uh, 1:15:37.072 --> 1:15:40.232 design decisions and things to comprehend. Um, and if I 1:15:40.272 --> 1:15:43.312 can't even get these large language models to write me 1:15:43.312 --> 1:15:45.512 a language where it literally is a 1 to 1 1:15:45.512 --> 1:15:48.712 map of existing languages, I can't imagine it being able 1:15:48.712 --> 1:15:51.722 to just on the fly, consume, uh, say like a 1:15:51.722 --> 1:15:54.802 documentation or a press release about a new technology and 1:15:54.802 --> 1:15:57.282 then be able to immediately work with that technology. 1:15:58.122 --> 1:16:00.962 Yeah. I mean, I would agree it's not possible. Now, um, 1:16:01.202 --> 1:16:03.722 I've actually tried this as well. I mean, there's also 1:16:03.722 --> 1:16:06.642 just languages that a given model is better with, and 1:16:06.642 --> 1:16:09.722 everyone kind of knows that. So it's not as simple 1:16:09.722 --> 1:16:12.282 as just go grab the spec for the language. And 1:16:12.282 --> 1:16:14.722 now it's suddenly better because of the previous point that 1:16:14.722 --> 1:16:17.881 you made. But this is the type of thing that 1:16:17.882 --> 1:16:22.002 billions of dollars are being spent on, like, uh, cursor and, 1:16:22.042 --> 1:16:25.001 you know, um, well, I guess all the pinnacle models 1:16:25.002 --> 1:16:28.282 are obviously working on this, but the kind of the 1:16:28.282 --> 1:16:31.362 biggest challenge that everyone's trying to solve right now is 1:16:31.362 --> 1:16:34.762 how to get that context into the current working model. 1:16:35.082 --> 1:16:39.082 And that is accelerating at like a crazy amount. I 1:16:39.082 --> 1:16:42.002 don't think it requires retraining. Um, I think fine tuning 1:16:42.002 --> 1:16:44.122 is largely kind of blown up and just not a 1:16:44.122 --> 1:16:50.171 great thing. Um, I expect that to be very possible, uh, 1:16:50.292 --> 1:16:53.732 very soon. And I'm not saying, like, um, an entire 1:16:53.732 --> 1:16:57.532 new database structure, entirely new programming language. Um, but it 1:16:57.532 --> 1:17:00.572 takes time for a human to learn a programming language 1:17:00.572 --> 1:17:02.212 as well. It's not like the human just go reads 1:17:02.212 --> 1:17:04.092 the spec and now they can suddenly do it. That's 1:17:04.092 --> 1:17:06.251 a thing that also takes a human years to learn. 1:17:06.492 --> 1:17:09.892 I think AIS are going to be way faster, um, 1:17:09.932 --> 1:17:12.652 at doing that. And I think that'll speed up, um, 1:17:12.652 --> 1:17:15.692 pretty dramatically over the next. I don't know who knows 1:17:15.692 --> 1:17:17.092 how fast 1 to 3 years. 1:17:19.692 --> 1:17:22.012 Yeah. So that's kind of where I think we our 1:17:22.012 --> 1:17:25.732 core disagreement is, which is basically, I think in order 1:17:25.732 --> 1:17:29.572 to be able to do that, you need actual functional intelligence, 1:17:29.812 --> 1:17:32.211 which I do not believe a machine can ever have. 1:17:32.452 --> 1:17:35.692 So I think we can only keep building these models 1:17:36.012 --> 1:17:40.492 that they emulate intelligence on specific tasks. But I actually 1:17:40.492 --> 1:17:42.812 think it's going to take longer to build those models 1:17:42.812 --> 1:17:45.292 for these new technologies than it is for a bunch 1:17:45.292 --> 1:17:48.342 of humans to just learn about the tech. So I 1:17:48.342 --> 1:17:50.062 think we're actually just going to we're going to hit 1:17:50.062 --> 1:17:53.582 a wall where a lot of people are thinking that, um, 1:17:53.622 --> 1:17:55.342 these models are just get more and more advanced, more 1:17:55.342 --> 1:17:57.742 and more intelligent, and at some point they will reach 1:17:57.742 --> 1:18:00.942 human level intelligence. Whereas my personal belief is that we 1:18:00.942 --> 1:18:03.542 have already capped out and we're now at the stage 1:18:03.542 --> 1:18:05.542 of we will just build some models that can do 1:18:05.542 --> 1:18:09.622 this specific task. So now you're, uh, essentially now your 1:18:09.622 --> 1:18:12.902 question is, can I learn a new framework or a 1:18:12.902 --> 1:18:16.461 new language or a new technology, then understand it enough 1:18:16.462 --> 1:18:19.902 to build a submodel, build the Submodel, make sure it 1:18:19.902 --> 1:18:23.542 actually works functionally faster than someone could, just learn it 1:18:23.542 --> 1:18:26.542 and write code in it. And I, I don't see 1:18:26.542 --> 1:18:30.142 any way in which without getting to a model that 1:18:30.142 --> 1:18:34.462 has actual human level intelligence, not knowledge, actual intelligence, we 1:18:34.462 --> 1:18:36.421 would ever get to the point where it can learn 1:18:36.422 --> 1:18:38.102 that quickly without retraining. 1:18:39.462 --> 1:18:41.742 Well, keep in mind it only has to do it once, right? 1:18:41.742 --> 1:18:44.152 Because then it just becomes like an MCP that somebody 1:18:44.152 --> 1:18:46.631 can call so the whole world can use it, right? 1:18:46.672 --> 1:18:49.872 So I think that's just the scale of the the 1:18:49.872 --> 1:18:52.791 scale of the benefit of doing it this way, um, 1:18:52.832 --> 1:18:55.392 is extraordinary. The other thing is like, this is not 1:18:55.392 --> 1:18:58.312 even talk about like actual ASI, which I consider to 1:18:58.312 --> 1:19:01.912 be like true general intelligence, which I'm still a bit 1:19:01.952 --> 1:19:04.472 agnostic on. I would say I'm like 90% that it 1:19:04.472 --> 1:19:07.072 gets there eventually, but I have no idea when. But 1:19:07.072 --> 1:19:11.192 more importantly, like, I tend to like, listen, when somebody like, uh, 1:19:11.512 --> 1:19:14.992 Karpathy or, uh, Ilya or people like this, people who 1:19:14.992 --> 1:19:17.671 are like knee deep into this and are the true 1:19:17.672 --> 1:19:21.192 experts on what the frontier looks like. Um, plus a 1:19:21.192 --> 1:19:22.992 lot of the people that I follow, actually, at the 1:19:22.992 --> 1:19:26.232 labs who are not like the marketing people, not the CEO, 1:19:26.512 --> 1:19:30.192 but the actual researchers talking about how fast they're making 1:19:30.192 --> 1:19:34.392 these jumps. Um, I think this whole thing of, like, 1:19:34.432 --> 1:19:37.592 learn a new framework for programming is going to be 1:19:37.592 --> 1:19:43.642 relatively small in the difficulty, um, category. Um, but I 1:19:43.682 --> 1:19:46.322 do hear your main point. I would say that this 1:19:46.322 --> 1:19:48.562 whole thing we've been talking about here is not even 1:19:48.562 --> 1:19:51.802 my main point. My main point is that what I 1:19:51.842 --> 1:19:55.762 care about is the impact of humans, and that if, 1:19:56.362 --> 1:20:01.402 you know, cardiologists and marriage therapists and really advanced, like, 1:20:01.442 --> 1:20:05.642 highly trained people are largely doing rote knowledge and having 1:20:05.682 --> 1:20:10.842 an an interaction about rote knowledge that could be defined, 1:20:10.882 --> 1:20:14.882 that could be applied to software engineering. Okay, what if 1:20:14.882 --> 1:20:18.322 we just didn't start using brand new tech all the 1:20:18.322 --> 1:20:21.082 time to build new tech? So let's say, for example, 1:20:21.082 --> 1:20:24.882 we actually locked in on TypeScript and a certain back 1:20:24.922 --> 1:20:27.642 end technology and a certain, you know, database and a 1:20:27.642 --> 1:20:30.881 certain server structure. And then the AI got really, really 1:20:30.882 --> 1:20:34.042 good at that. And then the whole goal became, let's 1:20:34.042 --> 1:20:38.162 maximize GDP for the planet or whatever it is that 1:20:38.802 --> 1:20:43.732 that would potentially make a software engineer extremely an AI 1:20:43.772 --> 1:20:47.492 based software engineer way more productive and way lower marginal 1:20:47.492 --> 1:20:50.372 cost than hiring a human right. There's no rule that 1:20:50.372 --> 1:20:53.092 says we must adopt new technologies all the time as 1:20:53.092 --> 1:20:55.492 they come out the following day, even though that is 1:20:55.532 --> 1:20:58.972 like the way humans do it. But I just don't 1:20:58.972 --> 1:21:02.892 think we can look at the top tier of someone 1:21:02.892 --> 1:21:06.972 like a Marcus, someone like myself, someone like a malware engineer, 1:21:07.212 --> 1:21:10.652 someone like, you know, a cardiologist or whatever, and say, 1:21:10.692 --> 1:21:14.092 you know, there's parts of their job that they're going 1:21:14.092 --> 1:21:16.372 to be able to pivot because they're exceptional. There's parts 1:21:16.372 --> 1:21:18.412 of their job and I might not be able to do. 1:21:18.572 --> 1:21:22.092 I'm worried about the other 99% who are not doing 1:21:22.092 --> 1:21:26.612 anything extraordinarily exceptional. All those examples that I gave of, 1:21:26.652 --> 1:21:29.452 like regular tasks, that's what most people are being paid 1:21:29.452 --> 1:21:33.171 money to do. And that's what I can already do. 1:21:33.212 --> 1:21:37.171 It just hasn't been like systematized and like brought into 1:21:37.172 --> 1:21:40.132 companies where it's actually running as an engine, where they 1:21:40.132 --> 1:21:44.492 can start laying everyone off and hiring more of those things. 1:21:44.812 --> 1:21:47.372 So if you could replace doctors like this, if you 1:21:47.372 --> 1:21:52.772 could replace, you know, um, marriage counselors, then it's going 1:21:52.812 --> 1:21:54.692 to head knowledge work. It's going to hit it really bad. 1:21:56.492 --> 1:22:00.331 Um, I mean, I guess it's a big if, but 1:22:00.332 --> 1:22:04.292 if it could, then. Sure. Yeah. Um, but then there's 1:22:04.292 --> 1:22:06.852 there's two, two points I would make. The first is, 1:22:06.852 --> 1:22:09.211 does it all happen at once? Does it happen so 1:22:09.212 --> 1:22:12.452 fast that all of these people are just out of work? Um, 1:22:12.772 --> 1:22:15.852 and two, is that, like, do we not find a 1:22:15.852 --> 1:22:19.212 way to deal with that? Um, I would argue that 1:22:19.212 --> 1:22:21.932 there's always going to be critical thinking work in any 1:22:21.932 --> 1:22:25.812 role that cannot be automated with large language models. So 1:22:25.812 --> 1:22:28.132 there is always going to be a place for someone 1:22:28.252 --> 1:22:32.372 who has the knowledge of that domain and the critical thinking. Um, 1:22:32.372 --> 1:22:35.012 I see this a lot in security, actually. Uh, we, 1:22:35.012 --> 1:22:37.822 we try and secure medical systems. It's like, I don't 1:22:37.822 --> 1:22:40.422 know shit about medical systems. I can be like, oh, 1:22:40.462 --> 1:22:42.222 let's fire a wall it off from the internet. And 1:22:42.222 --> 1:22:44.142 it's like, oh no, that person is dead now. It 1:22:44.142 --> 1:22:47.102 needed to talk to the cloud to set their heartbeat 1:22:47.102 --> 1:22:50.142 pace or some shit. Um, so then we have medical, uh, 1:22:50.142 --> 1:22:53.182 professionals come in and they understand these technologies, they understand 1:22:53.182 --> 1:22:55.982 the patient and they understand the needs of the patient, 1:22:55.982 --> 1:22:58.342 and they, uh, they communicate them to us, and then 1:22:58.342 --> 1:23:01.182 we do the security. And I think there's always going 1:23:01.182 --> 1:23:04.142 to be an equivalent of that for anything you can 1:23:04.142 --> 1:23:06.062 do with AI. There is always going to be a 1:23:06.062 --> 1:23:09.022 need for someone who has the same knowledge as the AI, 1:23:09.422 --> 1:23:13.222 but also the, uh, the critical thinking, the intelligence to 1:23:13.262 --> 1:23:17.142 actually work out the problems. So it might be we just, uh, 1:23:17.382 --> 1:23:22.222 everyone becomes project managers, like, instead of having, uh, humans doing, like, 1:23:22.262 --> 1:23:28.262 data entry and busywork, everyone becomes a product, uh, project manager. 1:23:28.262 --> 1:23:31.182 And the AI is now the low level employee. Now, personally, 1:23:31.182 --> 1:23:33.822 I would love if that was possible. If we could 1:23:34.072 --> 1:23:38.272 take the entire working class and move everyone up a 1:23:38.272 --> 1:23:41.791 level and have all those people make, uh, good salaries 1:23:41.832 --> 1:23:44.792 like project manager salaries, and then have the AI do 1:23:44.792 --> 1:23:46.711 all the grunt work that no one wants to do. 1:23:46.992 --> 1:23:49.112 I think that would be amazing, but I also don't 1:23:49.112 --> 1:23:51.792 think it's realistic. I don't think we're actually going to 1:23:51.792 --> 1:23:54.671 get to the point where it can even do enough 1:23:54.672 --> 1:23:56.072 to get us to that level. 1:23:57.072 --> 1:23:59.312 Don't you think the AIS are going to be okay 1:23:59.312 --> 1:24:01.712 if the AI could do cardiologist work, don't you think 1:24:01.752 --> 1:24:07.792 it could be a better project manager? Nope. Isn't project management? 1:24:07.792 --> 1:24:10.072 Isn't all the things that you just described? That is 1:24:10.072 --> 1:24:14.392 knowledge that that thing of like, well, we didn't know 1:24:14.392 --> 1:24:16.912 how to secure it because it's a medical thing. That 1:24:16.912 --> 1:24:20.312 medical thing is knowledge that is captured somewhere. Like, this 1:24:20.312 --> 1:24:23.232 is all just a matter of orchestration. I mean, you 1:24:23.232 --> 1:24:25.152 could feed all this to a project manager and it 1:24:25.152 --> 1:24:28.072 would know, hey, Marcus, don't, like, set up that firewall 1:24:28.072 --> 1:24:32.482 rule because so-and-so is going to die like this. This is, um. 1:24:32.522 --> 1:24:34.482 This is not net new stuff. This goes back to 1:24:34.482 --> 1:24:37.322 your previous point of like. Well, now you're just calling 1:24:37.322 --> 1:24:39.442 up knowledge. Yeah, that's that's the whole game. The whole 1:24:39.442 --> 1:24:40.762 game is calling up knowledge. 1:24:41.482 --> 1:24:44.122 Well I meant that example, not as an example of 1:24:44.122 --> 1:24:46.961 something I couldn't do, but as an example of when 1:24:46.962 --> 1:24:49.482 we need to bring in people with one skill set 1:24:49.522 --> 1:24:52.522 to another. Um, and the same would be true for 1:24:52.682 --> 1:24:55.722 the skill set essentially just being intelligence, like critical thinking, 1:24:55.962 --> 1:24:59.122 the the parts that are missing from AI, there would 1:24:59.122 --> 1:25:03.082 exist that in every field. And you can't take someone 1:25:03.082 --> 1:25:05.602 who has, uh, like what I was trying to get 1:25:05.602 --> 1:25:07.602 at is you can't take someone who has the intelligence 1:25:07.602 --> 1:25:10.082 and the knowledge, sorry, the intelligence and the critical thinking, 1:25:10.082 --> 1:25:12.762 but not the knowledge, and then give them the knowledge 1:25:12.762 --> 1:25:15.522 of the AI and make it whole. Uh, because that 1:25:15.522 --> 1:25:17.842 person now has no ability to fact check the AI. 1:25:18.242 --> 1:25:21.802 They have no ability to work with that information because 1:25:21.802 --> 1:25:23.282 they don't know if it's true or not. They just 1:25:23.282 --> 1:25:25.762 know what this model is telling them. Which is why 1:25:25.762 --> 1:25:29.802 I say, like with the medical professionals, I don't know 1:25:29.802 --> 1:25:33.852 about how medical devices work. I could ask ChatGPT. It's 1:25:33.852 --> 1:25:35.452 going to give me a horrible answer and I'm going 1:25:35.492 --> 1:25:38.612 to end up killing someone. So we need someone who understands, 1:25:38.612 --> 1:25:40.812 who has both the knowledge in the medical field and 1:25:40.812 --> 1:25:44.972 the critical thinking and intelligence. So my argument is that, uh, 1:25:45.092 --> 1:25:47.692 these large language models will always be able to replace 1:25:47.692 --> 1:25:50.331 the knowledge. They will always be able to, to an extent, 1:25:50.332 --> 1:25:53.692 substitute knowledge, but they can never do the critical thinking. 1:25:53.692 --> 1:25:56.532 So you always need a human for that. And it 1:25:56.532 --> 1:25:58.292 can't just be any human. It's going to have to 1:25:58.292 --> 1:26:01.532 be someone who also has some of the knowledge to 1:26:01.732 --> 1:26:03.412 actually work with the model. 1:26:04.372 --> 1:26:07.412 Yeah. So here's what I think the fundamental issue is 1:26:07.412 --> 1:26:11.172 that you gave an example of like when AI is 1:26:11.172 --> 1:26:14.892 using its knowledge and its vast knowledge, it's not actually 1:26:14.892 --> 1:26:18.171 being intelligent. Well guess what? When you bring over Sarah 1:26:18.172 --> 1:26:22.092 Meier to be the medical expert to help us not 1:26:22.092 --> 1:26:25.692 do the wrong firewall rule, she is not being Einstein right? 1:26:25.692 --> 1:26:28.902 Then she's just calling on her knowledge. She is also 1:26:28.902 --> 1:26:34.142 not invoking her creativity and her massive critical thinking. She's 1:26:34.142 --> 1:26:38.342 calling a database inside her own brain. That database can 1:26:38.382 --> 1:26:42.022 be in the context that is given when this AI 1:26:42.062 --> 1:26:44.662 tries to write the firewall rule. When the AI goes 1:26:44.662 --> 1:26:47.622 to consider writing the firewall rule, it will pull from 1:26:47.622 --> 1:26:51.102 its knowledge, which includes Sarah Meyer's knowledge, which is you 1:26:51.102 --> 1:26:54.742 don't put firewall rules for egress traffic inside of medical, 1:26:55.102 --> 1:26:58.622 you know, um, high criticality systems because you might block 1:26:58.662 --> 1:27:02.542 the heartbeat monitor that that was not critical thinking that 1:27:02.542 --> 1:27:06.462 Sarah did she she called on her knowledge. So the 1:27:06.462 --> 1:27:10.421 same exact thing that how how is that how is that? 1:27:11.262 --> 1:27:14.742 Because essentially, when you're trying to bridge these two different fields, 1:27:14.742 --> 1:27:17.501 like we're trying to secure a medical system that's taking 1:27:17.502 --> 1:27:20.902 security and medical systems, there is a gap there. There 1:27:20.902 --> 1:27:23.302 is a gap where let's say we have Sarah, who 1:27:23.302 --> 1:27:25.942 knows like a lot about medical systems. And then there's 1:27:25.942 --> 1:27:28.912 me who knows a lot about cybersecurity. Well, there needs 1:27:28.912 --> 1:27:30.631 to be communication. There needs to be thought, there needs 1:27:30.632 --> 1:27:33.272 to be problem solving. And that is where the intelligence 1:27:33.272 --> 1:27:35.592 comes in. If we just sit there with our collective 1:27:35.592 --> 1:27:38.152 knowledge and we sit in a room together and we're like, 1:27:38.152 --> 1:27:41.872 I'm knowledgeable and you're knowledgeable, nothing's going to happen. We 1:27:41.872 --> 1:27:43.872 have to discuss and we have to reason and we 1:27:43.872 --> 1:27:46.552 have to think. And that is the part that large 1:27:46.552 --> 1:27:49.872 language models cannot do. They cannot reason. They cannot think. 1:27:50.152 --> 1:27:53.512 They just have the knowledge. Now you need the intelligence 1:27:53.512 --> 1:27:56.872 to apply the knowledge. And that is where we come in. 1:27:57.192 --> 1:28:00.232 So so yeah, I love this example by the way. 1:28:00.232 --> 1:28:03.592 This is great. So imagine you're in the room with Sarah. 1:28:04.272 --> 1:28:09.112 And again we're going back to the fact that intelligence 1:28:09.272 --> 1:28:12.192 is needed when you don't have the knowledge. Right. So 1:28:12.192 --> 1:28:14.832 so what's this? You're sitting in the, in the room 1:28:14.832 --> 1:28:17.512 with all the knowledge about cybersecurity. She's sitting in the 1:28:17.512 --> 1:28:21.032 room with all the knowledge about medical systems, and you're 1:28:21.032 --> 1:28:25.282 having a conversation. It turns out you're just exchanging knowledge, 1:28:26.042 --> 1:28:31.041 because the fact of what should be done is already known. Right? 1:28:31.282 --> 1:28:36.242 I disagree. Yeah. Like you're making a situation where, like, 1:28:36.282 --> 1:28:38.961 whatever it is we're trying to solve, there is already 1:28:38.962 --> 1:28:42.162 a known solution. Um, whereas if that was the case, 1:28:42.162 --> 1:28:44.482 then obviously security would be solved, right? Like, if there 1:28:44.482 --> 1:28:48.842 was a known solution to every problem, we wouldn't have jobs. Um, 1:28:48.842 --> 1:28:52.122 I actually don't know if you work in cyber security still. Um, but. 1:28:52.722 --> 1:28:55.162 So so check this out. That, that that kind of 1:28:55.202 --> 1:28:58.482 is the case, though, isn't it? Marcus doesn't. Has anyone 1:28:58.482 --> 1:29:00.282 come up with an answer, like, have you coming up 1:29:00.282 --> 1:29:03.642 with an answer for a customer where? Um, it hadn't 1:29:03.642 --> 1:29:06.682 been talked about, like basically doing that solution hadn't been 1:29:06.682 --> 1:29:09.602 talked about a million times throughout the history of cybersecurity. 1:29:11.282 --> 1:29:14.642 Yes, I personally have, um, but a lot of the 1:29:14.642 --> 1:29:17.922 time it isn't just like, it's kind of hard to 1:29:17.922 --> 1:29:21.202 explain in a way that the non-security viewers will will 1:29:21.322 --> 1:29:26.172 easily grasp. but a lot of security is known solutions. 1:29:26.732 --> 1:29:30.932 They're just not good solutions. Like, let's say, password theft. 1:29:30.932 --> 1:29:33.532 Your password gets stolen. What's the everyone's solution to that 1:29:33.532 --> 1:29:37.492 two factor authentication? You send an SMS to their phone. Um, 1:29:37.932 --> 1:29:41.892 why doesn't everything have two factor authentication? Well, because there's 1:29:41.892 --> 1:29:45.532 problems with that. Okay. Let's say we choose phone based 1:29:45.532 --> 1:29:48.332 two factor authentication. What if I lose my phone? 1:29:49.052 --> 1:29:49.532 Sure. 1:29:49.572 --> 1:29:51.252 Now I have to go down to it and I 1:29:51.252 --> 1:29:53.852 have to get my password reset. And I'm losing productivity, 1:29:54.132 --> 1:29:57.052 and it has to somehow verify my identity and know 1:29:57.052 --> 1:29:59.452 that I'm not a scammer pretending to be me. So 1:29:59.452 --> 1:30:02.292 there's a huge productivity loss there. And then what happens 1:30:02.292 --> 1:30:04.132 when I pick up my phone to to get my 1:30:04.172 --> 1:30:07.732 toufar code? And I see there's a notification from my girlfriend. 1:30:07.972 --> 1:30:09.972 So I start talking to my girlfriend and we've just 1:30:09.972 --> 1:30:12.732 now lost a bunch of productivity. And then you amplify 1:30:12.732 --> 1:30:17.572 that over an entire office space and to whatever the 1:30:17.572 --> 1:30:21.422 cost of password theft was. Um, being hacked was you 1:30:21.462 --> 1:30:24.982 probably lost more in productivity loss. So while we have 1:30:24.982 --> 1:30:29.502 existing solutions, we actually don't have good solutions. Everything is 1:30:29.502 --> 1:30:33.381 just trying to get the best possible, uh, the best 1:30:33.382 --> 1:30:37.862 possible idea for our specific scenario, which is really what 1:30:37.902 --> 1:30:40.822 cybersecurity is about. And that is the critical thinking. That's 1:30:40.822 --> 1:30:43.662 the intelligence. And I can just go, are your passwords 1:30:43.662 --> 1:30:47.622 are being stolen? Use Toofar. Okay. What? Toofar. Yeah. How 1:30:47.622 --> 1:30:49.062 do I handle password reset? 1:30:49.502 --> 1:30:52.342 But that's not what an AI is doing. Okay. So 1:30:52.382 --> 1:30:56.022 modern AIS like because I do this with risk assessment 1:30:56.182 --> 1:30:58.381 all the time, I do. Yeah. To answer your question, 1:30:58.382 --> 1:31:01.702 I'm doing this all the time for actual customers when 1:31:01.702 --> 1:31:04.702 you give the thing proper context. So if you give 1:31:04.702 --> 1:31:06.862 it the thing that look toofar doesn't work in this 1:31:06.862 --> 1:31:09.742 situation because of this, um, you don't even have to 1:31:09.742 --> 1:31:12.342 give it that. You could just say, this situation exists. 1:31:12.382 --> 1:31:15.622 This situation exists. We have this complexity over here. We 1:31:15.622 --> 1:31:17.822 have this complexity over here. What do you think the 1:31:17.822 --> 1:31:21.032 security control should be. It will say things like, well, 1:31:21.592 --> 1:31:24.392 based on your current state of your cloud infrastructure and 1:31:24.392 --> 1:31:26.232 based on the fact that you do business in France 1:31:26.232 --> 1:31:28.192 and based on the fact that tufa won't work for 1:31:28.192 --> 1:31:32.711 this and you can't actually do SMS because, uh, whatever, 1:31:32.712 --> 1:31:36.312 it's it's too easy to do, um, spoofing of, you know, 1:31:36.592 --> 1:31:38.792 cell phones. So we're going to rule that one out, 1:31:38.792 --> 1:31:42.472 especially in this scenario because of this jurisdiction whatever. And 1:31:42.472 --> 1:31:44.752 you start giving it the more context you give it, 1:31:44.752 --> 1:31:48.192 the more it's going to navigate all those special situations 1:31:48.192 --> 1:31:50.631 just the way that that you would or that I would. 1:31:51.072 --> 1:31:52.792 My argument to you is that when you look at 1:31:52.792 --> 1:31:56.112 an average pen test report and again, we can't use 1:31:56.152 --> 1:31:59.232 me or you or some other person who's been doing 1:31:59.232 --> 1:32:01.952 this forever as the example, because, yeah, maybe we can 1:32:01.952 --> 1:32:04.192 come up with a novel thing that's never been thought 1:32:04.192 --> 1:32:06.712 of before. But if you go take all the pen 1:32:06.712 --> 1:32:09.712 test reports produced last week and we go step by 1:32:09.712 --> 1:32:12.631 step through them, what kind of novelty is in there? 1:32:12.632 --> 1:32:15.712 What kind of new solutions are being proposed? Aren't they 1:32:15.712 --> 1:32:19.482 mostly saying, um, this thing was wide open. You've got 1:32:19.482 --> 1:32:22.002 a config problem here. Uh, that config should not be 1:32:22.002 --> 1:32:23.881 in that way. Oh, by the way, you should review 1:32:23.922 --> 1:32:26.602 your changes before you actually publish them. Oh, by the way, 1:32:26.602 --> 1:32:29.842 you have, you know, your credentials are open. Uh, like, 1:32:29.882 --> 1:32:32.442 it's going to be very common stuff. Like you've got 1:32:32.482 --> 1:32:34.362 to patch this stuff, you've got to patch. 1:32:35.602 --> 1:32:37.042 But who's giving it the context? 1:32:38.922 --> 1:32:41.362 The point is people are being paid to do this work. 1:32:41.402 --> 1:32:43.282 They don't have to give any context. They can just 1:32:43.282 --> 1:32:45.522 give the list of vulnerabilities. They can give the list 1:32:45.522 --> 1:32:47.402 of vulnerabilities. They can say, I was able to get 1:32:47.402 --> 1:32:49.842 in through this. Um, you have to patch that system 1:32:49.842 --> 1:32:52.442 because it's critical. And the people read it and they're like, 1:32:52.442 --> 1:32:54.882 oh yeah, yeah, it's a critical system. This is what 1:32:54.882 --> 1:32:57.921 my team has been telling me for 14 years. Everything 1:32:57.922 --> 1:32:59.881 in this report, my team has already told me a 1:32:59.882 --> 1:33:02.762 million times, I just needed to get this report because 1:33:02.762 --> 1:33:05.802 now I can justify the security spend. Like how much 1:33:05.802 --> 1:33:11.522 of novel, quote unquote novel things that very highly paid 1:33:11.522 --> 1:33:15.922 people are being paid for is actually exactly the thing 1:33:15.922 --> 1:33:19.682 that you described before, which is it's very much known. 1:33:19.722 --> 1:33:22.962 Things just applied in a specific way for a specific customer. 1:33:23.242 --> 1:33:26.522 It's not always the same. It's non-deterministic because it's a 1:33:26.522 --> 1:33:30.842 particular tester doing it. But the output looks remarkably similar, 1:33:31.362 --> 1:33:34.002 almost identical to the output of another pentester doing the 1:33:34.002 --> 1:33:34.802 same exact thing. 1:33:36.162 --> 1:33:39.362 Yeah, but now we're switching from the reports to the 1:33:39.362 --> 1:33:42.802 decision making. Because, Ali, you said about the context, like, okay, 1:33:42.842 --> 1:33:45.042 the company is operating in this jurisdiction, blah blah, blah, blah, 1:33:45.042 --> 1:33:48.002 blah blah blah, but who is giving it that context? Like, 1:33:48.042 --> 1:33:51.162 surely there has to be a person who is gathering 1:33:51.162 --> 1:33:53.482 and giving the model that context. 1:33:53.682 --> 1:33:58.082 Sure, sure. The agent, an agent is crawling all the docs. 1:33:58.442 --> 1:34:01.362 It's having a chat, interviews with people. It's doing voice 1:34:01.362 --> 1:34:04.562 calls and interviewing people. Like these are this is all 1:34:04.562 --> 1:34:09.002 just like common common stakes for like something an AI 1:34:09.042 --> 1:34:11.122 could do. It could do an interview, it could talk 1:34:11.122 --> 1:34:14.412 to you and extract information. It could read an entire wiki. 1:34:14.452 --> 1:34:17.251 It could read every doctrine on, uh, you know, Google 1:34:17.252 --> 1:34:20.492 Docs and Confluence. Like, it can go and gather that context. 1:34:21.292 --> 1:34:23.092 You see, I don't think it can. Like, I don't 1:34:23.092 --> 1:34:26.892 think it has the capability or the intelligence to gather 1:34:26.892 --> 1:34:30.172 that kind of context. I think maybe there is an 1:34:30.172 --> 1:34:34.012 argument that in the future it could, uh, but given 1:34:34.012 --> 1:34:37.852 any current generation model I have used, we're not even 1:34:37.852 --> 1:34:39.852 close to being able to gather the context. We can't 1:34:39.852 --> 1:34:43.052 even answer the questions. Like I can, I've had multiple 1:34:43.052 --> 1:34:45.532 times where I've asked the AI stuff, and because it's 1:34:45.532 --> 1:34:48.212 been trained on all of the past knowledge, which, as 1:34:48.212 --> 1:34:50.892 I mentioned, is constantly changing, it'll be like, oh, we 1:34:50.892 --> 1:34:54.131 should do password rotation. And then like we've disproven password 1:34:54.132 --> 1:34:58.131 rotation like 20 years ago. Like, why are you suggesting this? Um, 1:34:58.492 --> 1:35:02.412 and so we're right now, we're still running into the 1:35:02.412 --> 1:35:05.612 problem of, oh, the best practices have changed in the 1:35:05.612 --> 1:35:08.251 last five years. Maybe it's even the last one year 1:35:08.492 --> 1:35:11.662 and the model hasn't even updated to to account for that. 1:35:11.662 --> 1:35:14.542 And now you're saying we could have the model that 1:35:14.542 --> 1:35:17.462 can barely even keep up with the current state of 1:35:17.462 --> 1:35:21.302 the world? Now, actually, like interviewing people and like profiling 1:35:21.302 --> 1:35:25.182 my company and like the local laws and the documentations. 1:35:25.182 --> 1:35:28.982 And I think it may be in ten, 20 years 1:35:28.982 --> 1:35:32.462 is something that I could do. Uh, but there is 1:35:32.462 --> 1:35:36.142 no current generation model I've seen that even comes close 1:35:36.422 --> 1:35:37.662 to being able to do that. 1:35:37.702 --> 1:35:40.422 But but it's not a model, though, Marcus. It's like. 1:35:40.422 --> 1:35:45.222 It's not like it's not like, um, O3 or some 1:35:45.222 --> 1:35:48.942 model has has capabilities like that. It's more like we 1:35:48.982 --> 1:35:52.542 hired this, I don't know. Securus. This Securus company comes 1:35:52.542 --> 1:35:54.782 out and it's like our model Securus is the smartest 1:35:54.782 --> 1:35:58.302 model ever, and it will do a security assessment for 1:35:58.302 --> 1:36:01.422 you when you give the thing to Securus. It's actually 1:36:01.422 --> 1:36:03.862 smoke and mirrors. It's actually a whole bunch of agents 1:36:03.862 --> 1:36:07.382 sitting behind it. So it's really this really smart orchestrator 1:36:07.382 --> 1:36:11.032 in the front. It has a million different gatherers that 1:36:11.032 --> 1:36:13.872 go and answer questions that haven't been talked about. So 1:36:13.872 --> 1:36:17.352 you handed us this thing that's like, hey, we're building this, uh, 1:36:17.712 --> 1:36:21.672 you know, whatever it is, um, this special new application. 1:36:21.832 --> 1:36:24.312 And the top level agent is like. I have no 1:36:24.312 --> 1:36:27.032 idea what that is. Let me go crawl all the documentation. 1:36:27.032 --> 1:36:31.432 So it's all smoke and mirrors. Orchestration of multiple things 1:36:31.432 --> 1:36:34.352 going together. It's like, hey, what about this application? Well, 1:36:34.392 --> 1:36:37.752 it's being developed by so-and-so team. Okay, who's on that team? Okay, 1:36:37.752 --> 1:36:39.672 here's the list of people. Okay. Set up meetings and 1:36:39.672 --> 1:36:42.352 go talk to them or send them this form. Have 1:36:42.352 --> 1:36:43.832 them fill it out and bring it back to us, 1:36:43.832 --> 1:36:48.072 and we'll parse that. So it's nothing about those individual 1:36:48.072 --> 1:36:52.152 steps are difficult. What's difficult is just combining all those 1:36:52.152 --> 1:36:55.392 steps up into this overall product, so that it looks 1:36:55.392 --> 1:36:57.832 like with a customer, when you're talking to it, it's 1:36:57.832 --> 1:37:00.712 figuring it all out when behind the scenes, it's like 1:37:00.712 --> 1:37:03.912 you were saying before, it's a whole bunch of subtasks 1:37:03.912 --> 1:37:07.312 that are being doled out to this thing. And ultimately, 1:37:08.202 --> 1:37:11.842 that's just orchestration. That's just like all these little pieces 1:37:11.842 --> 1:37:15.762 being combined to produce the illusion of an overall intelligence. 1:37:16.322 --> 1:37:20.082 And my argument to you is that that capability is 1:37:20.082 --> 1:37:23.402 going to be functionally the same as the capability of 1:37:23.402 --> 1:37:25.282 you hired someone else to do it. You hire an 1:37:25.282 --> 1:37:27.282 actual person to do it. Because at the end of 1:37:27.282 --> 1:37:28.722 the day, it's going to be like the Turing test. 1:37:28.722 --> 1:37:30.722 You're going to look at two different pen test reports, 1:37:31.402 --> 1:37:33.881 and one was just you or me doing this assessment 1:37:33.882 --> 1:37:36.362 for three weeks where we had to do the interviews, 1:37:36.362 --> 1:37:40.282 and this other one launched off 780 agents and came back. 1:37:40.282 --> 1:37:42.922 And because the cost of inference is so short, you know, 1:37:42.962 --> 1:37:45.802 the whole thing cost them $13, whereas they would have 1:37:45.802 --> 1:37:47.442 paid you or me 90 grand to do it. 1:37:49.282 --> 1:37:51.602 I think that only works if you can build the 1:37:51.602 --> 1:37:54.762 system with the system, because if you're making all of 1:37:54.762 --> 1:37:56.921 these sub models to handle all of these, like niche 1:37:56.962 --> 1:38:02.122 edge cases and these like specific domains of expertise, you 1:38:02.122 --> 1:38:05.242 need people in those domains of expertise to build those systems. 1:38:05.532 --> 1:38:07.732 So now essentially all you have is rather than a 1:38:07.732 --> 1:38:11.052 bunch of pen test companies, you have the pen test company, 1:38:11.212 --> 1:38:14.052 which is every pen tester now working on building some 1:38:14.052 --> 1:38:17.412 model that can, uh, roughly emulate what a human pen 1:38:17.412 --> 1:38:20.372 tester would do. Um, but that I would not see 1:38:20.412 --> 1:38:23.212 as a job replacement. I would see as a job shift. Uh, 1:38:23.212 --> 1:38:27.131 you would essentially just end up with this horrible late 1:38:27.132 --> 1:38:32.532 stage capitalist bullshit where everyone works for a single AI company. Um, 1:38:32.532 --> 1:38:34.252 and no one has been replaced. We still need the 1:38:34.252 --> 1:38:36.932 same number of people just there, building models to automate 1:38:36.932 --> 1:38:40.972 things instead of doing the things. Um, but I don't 1:38:40.972 --> 1:38:45.132 see it as such a productivity boost that we would 1:38:45.172 --> 1:38:48.172 actually see like massive amounts of employees being lost. We 1:38:48.172 --> 1:38:51.612 would just see them going over to these, uh, these AGI, 1:38:51.652 --> 1:38:54.052 I guess we would call it companies to make these 1:38:54.052 --> 1:38:57.372 fake Agis by like, cobbling together a bunch of, uh, 1:38:57.492 --> 1:39:00.972 agents and sub models. But the amount of humans that 1:39:00.972 --> 1:39:04.022 you would need to build and maintain such a system 1:39:04.022 --> 1:39:07.622 would be like, phenomenal. It would be, uh, it would 1:39:07.622 --> 1:39:10.182 be like the biggest company on earth. So I wouldn't 1:39:10.182 --> 1:39:13.102 see that as people are going to be losing their jobs. 1:39:13.102 --> 1:39:14.982 I would see it as problematic because now you just 1:39:14.982 --> 1:39:18.942 have a monopoly that just it does everything. Um, and 1:39:18.942 --> 1:39:22.062 that's going to be horrible for working conditions. Um, but 1:39:22.062 --> 1:39:24.062 I wouldn't see that as a risk of, okay, now 1:39:24.062 --> 1:39:26.902 pen testers are obsolete. There's no more pen testers. Uh, okay. 1:39:26.942 --> 1:39:29.902 Medical professionals obsolete. There's no more medical professionals. It would 1:39:29.902 --> 1:39:33.102 just be. Everyone would just be like a medical professional 1:39:33.102 --> 1:39:34.862 working for an AI company. 1:39:35.982 --> 1:39:39.822 Yeah, but aren't these all just modules, though? So you 1:39:39.822 --> 1:39:42.142 have a module for going to collect more information when 1:39:42.142 --> 1:39:45.422 you don't know something, you have a module for summarizing 1:39:45.422 --> 1:39:48.142 that information and bringing it into the context. You have 1:39:48.142 --> 1:39:53.621 a module for um, you know. Net learning some new thing, right? 1:39:54.222 --> 1:39:56.702 I mean, let's go back to the cardiologist. Think of 1:39:56.702 --> 1:40:00.662 how much schooling that cardiologist had to have, right. And 1:40:00.662 --> 1:40:04.671 how different the different scenarios are. I would argue it's 1:40:04.672 --> 1:40:10.352 not that much more complex than, um, or less complex 1:40:10.352 --> 1:40:12.792 than pen testing. So I think what you have a 1:40:12.792 --> 1:40:15.872 module for understanding pen testing. Really well, a lot of 1:40:15.872 --> 1:40:19.512 these concepts are general enough. They are abstracted enough that 1:40:19.512 --> 1:40:24.072 when an AI understands them deeply enough, it is then 1:40:24.232 --> 1:40:27.752 using your analogy here or your concept, which I really like, 1:40:28.072 --> 1:40:31.632 it becomes knowledge. It becomes knowledge as opposed to intelligence. 1:40:32.032 --> 1:40:36.831 I really like this distinction. So you when it understands. 1:40:36.832 --> 1:40:39.552 And by the way, I would argue that we've already 1:40:39.712 --> 1:40:43.152 reached this with security. Um, it's already really, really good 1:40:43.152 --> 1:40:47.112 at cybersecurity. We have. Let me give you an example here. Okay. 1:40:47.112 --> 1:40:50.832 Let me give you an example of an extremely dynamic situation. Um, 1:40:51.552 --> 1:40:57.512 right now on hacker one, a completely automated, completely automated 1:40:58.032 --> 1:41:01.802 AI system has the As the number one spot. That 1:41:01.802 --> 1:41:06.282 means it's dealing with all sorts of random situations it's 1:41:06.282 --> 1:41:10.041 never seen before. It's doing its own research. It's finding 1:41:10.042 --> 1:41:13.242 new information on new technologies that's never seen before. It's 1:41:13.242 --> 1:41:16.882 launching all the attacks. It's actually showing proof and evidence 1:41:17.482 --> 1:41:20.362 that it can exploit it. It's actually submitting the reports 1:41:20.762 --> 1:41:25.642 and getting back the points with humans hand off. So 1:41:25.682 --> 1:41:29.722 I mean, if that's not the level of complexity that 1:41:29.722 --> 1:41:33.042 could replace a knowledge worker and let alone do a 1:41:33.042 --> 1:41:37.282 pen test, I mean, we already have all the evidence here, 1:41:37.322 --> 1:41:41.082 like the models are already good at security. And if 1:41:41.082 --> 1:41:44.122 you could just gather more knowledge to put into the context, 1:41:44.522 --> 1:41:45.442 that's all you need. 1:41:47.482 --> 1:41:49.322 But I think you have to ask yourself, like how 1:41:49.322 --> 1:41:53.122 many people were working on that model? Like, is this 1:41:53.122 --> 1:41:55.722 now a model that is just on autopilot forever that 1:41:55.722 --> 1:41:59.172 requires no human interaction whatsoever. And it just sits. And 1:41:59.172 --> 1:42:02.412 it does pen tests. Or does someone keep having to 1:42:02.452 --> 1:42:06.332 update systems behind the scenes because, uh, it kind of 1:42:06.372 --> 1:42:07.732 reminds me of that. I don't remember the name of 1:42:07.732 --> 1:42:10.732 the specific company, but there was this one company that was, uh, 1:42:10.732 --> 1:42:13.172 they were selling Vibe coding, and they went bankrupt after 1:42:13.172 --> 1:42:16.291 it was found that there was no I it was 1:42:16.292 --> 1:42:18.412 just a bunch of people in a, in an impoverished 1:42:18.412 --> 1:42:21.132 country writing the code for them. And essentially, this is 1:42:21.132 --> 1:42:23.572 all you're doing is an abstraction of this is you're 1:42:23.572 --> 1:42:26.252 having you don't know how many people are behind the 1:42:26.252 --> 1:42:30.131 scenes working the AI. Um, and I think that's kind 1:42:30.132 --> 1:42:32.892 of where the metrics get to get a bit skewed, 1:42:32.892 --> 1:42:35.772 because when we say someone was replaced with AI, we 1:42:35.812 --> 1:42:38.852 don't count how many jobs were gained by the creation 1:42:38.852 --> 1:42:42.092 of that AI. We're just like, oh, Jeff has been 1:42:42.092 --> 1:42:45.852 replaced with Pentester AI. Um, that means one person has 1:42:45.852 --> 1:42:48.171 now lost their job to AI, and what you don't 1:42:48.172 --> 1:42:50.892 see is on the back end. There's 500 people making this. 1:42:50.892 --> 1:42:53.932 I actually do anything at all. Um, so I think 1:42:53.932 --> 1:42:57.222 that's kind of the that's my issue with this argument 1:42:57.222 --> 1:43:02.541 of like the job replacement, that people being unemployed because, um, 1:43:02.582 --> 1:43:05.542 at least from what I'm seeing, the amount of work 1:43:05.542 --> 1:43:08.782 that goes into actually maintaining and running these systems is 1:43:08.782 --> 1:43:12.262 equivalent to doing it manually still. So I don't see 1:43:12.702 --> 1:43:14.702 any world in which we're going to see something as 1:43:14.702 --> 1:43:18.262 crazy as like 90% job loss. Um, and then that 1:43:18.262 --> 1:43:22.782 also assumes that everything is capped, right? That maybe pen 1:43:22.782 --> 1:43:24.902 testing is capped, like maybe there is only so much 1:43:24.902 --> 1:43:28.102 security we can do. But then there's other industries where 1:43:28.102 --> 1:43:30.622 it's like, oh, we gained some productivity. We can do 1:43:30.622 --> 1:43:33.662 more now. Um, so I would think that we would 1:43:33.662 --> 1:43:37.742 just see, uh, assuming we can get an AI model 1:43:37.982 --> 1:43:39.542 that just for the sake of argument, let's say it 1:43:39.542 --> 1:43:42.262 doesn't require any people. Like somehow we built an AI 1:43:42.302 --> 1:43:45.822 that just pilots itself, doesn't need any programmers, doesn't need 1:43:45.822 --> 1:43:48.302 any tweaks. It just works on its own. Would we 1:43:48.342 --> 1:43:51.102 not just take those systems and then focus into doing 1:43:51.102 --> 1:43:55.702 something more like, um, maybe we go in mined materials 1:43:55.702 --> 1:43:59.262 from asteroids. Maybe, uh, we start sending probes to, uh, 1:43:59.262 --> 1:44:01.662 to other planets within the galaxy. Like, there's always more 1:44:01.662 --> 1:44:04.542 work to do. Um, and the idea of, like, job 1:44:04.542 --> 1:44:07.502 replacement and job loss kind of hinges on this idea 1:44:07.662 --> 1:44:11.662 that there is finite work, there's finite productivity, um, and 1:44:11.662 --> 1:44:14.462 that a, the productivity isn't just being transferred to the 1:44:14.462 --> 1:44:16.862 AI side, and those people are just doing the same 1:44:16.862 --> 1:44:19.142 thing they were already doing. But for an AI company 1:44:19.502 --> 1:44:21.782 and that like there is just a ceiling and we 1:44:21.862 --> 1:44:23.902 hit this ceiling, it's like, okay, we don't need people anymore. 1:44:23.942 --> 1:44:26.622 Like we've done all the things we need to do. Um, 1:44:26.622 --> 1:44:28.421 and I just I don't see either. 1:44:29.422 --> 1:44:32.942 Yeah. So, so, so I think let's say it costs. 1:44:32.982 --> 1:44:35.382 Let's say it took, um, I actually don't know how 1:44:35.382 --> 1:44:37.422 big this company is that actually did the thing that 1:44:37.422 --> 1:44:41.022 I'm describing. I don't know how much, you know, human 1:44:41.022 --> 1:44:43.902 effort went into building the agents, but the thing running is, 1:44:43.942 --> 1:44:47.222 is not involving humans. It's actually the agents running that 1:44:47.222 --> 1:44:49.342 are doing the thing. Now, you have a good point 1:44:49.342 --> 1:44:51.942 of like, well, those agents start to rot, and we 1:44:51.942 --> 1:44:54.512 have to maintain the agent. Sure, but let's say that's 1:44:54.512 --> 1:44:57.752 only ten people or 100 people. The point is that 1:44:57.752 --> 1:45:02.472 is now replaceable infrastructure that the entire planet can use. 1:45:02.912 --> 1:45:05.711 So the planet can go and now run this, this 1:45:05.712 --> 1:45:09.832 company's application and pointed at targets, and it could find 1:45:09.832 --> 1:45:15.632 vulnerabilities and submit reports. Now that, um, the scaling of 1:45:15.672 --> 1:45:19.232 that infrastructure there to run that I don't think compares 1:45:19.232 --> 1:45:23.992 at all to, let's say, um, a million tests, let's 1:45:23.992 --> 1:45:26.552 say a million pen tests, a million bug bounties or 1:45:26.552 --> 1:45:29.232 pen tests or whatever. You're going to stick this thing 1:45:29.232 --> 1:45:33.752 on scaling up to how many humans are going to 1:45:33.792 --> 1:45:36.552 run that thing and how much they're going to cost, 1:45:37.072 --> 1:45:41.512 versus a million instances of this AI, especially given the 1:45:41.552 --> 1:45:43.832 I mean, look at the inference costs of AI and 1:45:43.832 --> 1:45:45.831 how they come down over like the last two years. 1:45:45.872 --> 1:45:50.402 I think we're paying like 0.01% or something of the 1:45:50.402 --> 1:45:53.121 inference cost of what we were paying to two years ago. 1:45:53.162 --> 1:45:57.082 It's something extraordinarily like that. So there's no reason to 1:45:57.122 --> 1:45:59.362 believe that that's not going to keep falling. Who knows 1:45:59.362 --> 1:46:02.282 what the bottom level is? But the marginal cost for 1:46:02.282 --> 1:46:06.162 doing a pen test using this automated system is going 1:46:06.162 --> 1:46:09.602 to be way lower than adding yet another tester. Not 1:46:09.602 --> 1:46:12.762 only that, but the testers, they're humans, they move on. 1:46:12.762 --> 1:46:15.642 They stop being testers and move on to they become 1:46:15.642 --> 1:46:19.402 managers or whatever. And so I think the scalability there's 1:46:19.442 --> 1:46:20.682 like no comparison. 1:46:24.322 --> 1:46:28.042 Um, I could like, given your assumptions, I could maybe 1:46:28.082 --> 1:46:31.722 see that, um, I don't know about, uh, obviously the 1:46:31.722 --> 1:46:33.922 first thing I would steer away from is the whole 1:46:33.922 --> 1:46:37.002 prior performance, future results thing. Like, if the costs are 1:46:37.002 --> 1:46:38.842 falling that we can't just expect, they're going to keep 1:46:38.842 --> 1:46:40.202 falling forever. Sure, sure. 1:46:40.322 --> 1:46:42.881 Um, that's why I said I know where the bottom is. 1:46:43.602 --> 1:46:45.682 In fact, we may actually see the opposite happen as 1:46:45.682 --> 1:46:49.652 we start to, uh, as, like, resources start to get constrained, 1:46:49.652 --> 1:46:51.652 we might actually see the costs start going back up 1:46:51.652 --> 1:46:54.812 because that's true. Silicon is pretty hard to get. Um, 1:46:54.812 --> 1:46:58.131 energy is um, we're kind of hitting our cap on 1:46:58.132 --> 1:46:59.852 energy right now. So we're having to build out more 1:46:59.852 --> 1:47:02.732 power plants, and those will end up getting factored into 1:47:02.732 --> 1:47:06.612 the costs of running these models. Um, but with something 1:47:06.612 --> 1:47:11.092 like that, say Pentesting, I could see in that specific example. Yeah, 1:47:11.092 --> 1:47:14.492 I think we could probably automate it to make it cheaper. Um, 1:47:14.652 --> 1:47:18.211 that's not because I think there's something fundamentally wrong with, uh, 1:47:18.692 --> 1:47:22.211 with people or the AI is fundamentally good. It's because 1:47:22.212 --> 1:47:24.932 the industry is kind of a scam. We go and 1:47:24.932 --> 1:47:28.852 we bill like $100,000 for some dude to, like, spend 1:47:28.852 --> 1:47:33.372 15 minutes running Nessus or something. Um, so I think 1:47:33.652 --> 1:47:36.852 in a lot of those spaces we could see, uh, 1:47:37.692 --> 1:47:41.092 I guess full AI automation, but a lot of that 1:47:41.092 --> 1:47:44.532 stuff is really just it was always automatable like, you 1:47:44.532 --> 1:47:46.742 could have scripted a lot of that. It's just that 1:47:46.742 --> 1:47:50.381 no one was doing it. The real work, which is 1:47:50.382 --> 1:47:54.542 not being AI automated, is the stuff that is actually dynamic. Um, 1:47:54.582 --> 1:47:56.662 because like one of the things with bug bounties is 1:47:56.662 --> 1:48:00.862 it is very, very, uh replicatable. Like, you can reproduce it. Um, 1:48:00.902 --> 1:48:04.662 there's only a very few classes of vulnerabilities. And if 1:48:04.662 --> 1:48:07.662 you go for something like low hanging fruit, like cross-site scripting, 1:48:07.902 --> 1:48:09.662 you could write a Python script to go and find 1:48:09.662 --> 1:48:11.982 a million of those. And a lot of the older 1:48:11.982 --> 1:48:15.102 pen testers did, sorry, bug bounty testers did do that. 1:48:15.342 --> 1:48:18.542 They were also automated, just not with AI. So that 1:48:18.542 --> 1:48:21.342 is a, uh, like that is something that has always 1:48:21.342 --> 1:48:24.222 been automatable. Like, I don't think AI has really made 1:48:24.222 --> 1:48:27.902 any amazing breakthroughs there. Um, it might have maybe increased 1:48:27.902 --> 1:48:31.662 the scope of vulnerabilities that are automatable to a, I 1:48:31.662 --> 1:48:33.982 would say a very small degree, but I don't think 1:48:33.982 --> 1:48:36.142 we're going to start seeing like large language models who 1:48:36.142 --> 1:48:40.182 are like writing novel exploits. Um, I think we're just 1:48:40.182 --> 1:48:43.581 going to see stuff that like, um, was already automated, 1:48:43.582 --> 1:48:49.032 like fuzzing, Like fuzzing. I guess you could call automation like, it's, uh, 1:48:49.072 --> 1:48:51.392 you basically just the equivalent of shaking a tree and 1:48:51.392 --> 1:48:54.592 seeing what falls out. Uh, you could put slap AI 1:48:54.592 --> 1:48:56.592 on that and say, oh, look, we have AI, automated 1:48:56.592 --> 1:48:59.912 pen testing, but a lot of those things were already 1:48:59.912 --> 1:49:06.472 automated or automatable. Um, so I think I'd caution like the, uh, 1:49:06.472 --> 1:49:09.671 going from, hey, look, this thing is doing very well 1:49:09.792 --> 1:49:15.552 in an automated sense to cybersecurity is already like, automatable. Um, 1:49:15.592 --> 1:49:19.631 because I, I personally stay away from pen testing. Um, 1:49:19.632 --> 1:49:21.832 but I would say that's a very, very small portion 1:49:21.832 --> 1:49:25.592 of cybersecurity. And most of the parts of cybersecurity, uh, 1:49:25.832 --> 1:49:29.712 that like really do need doing isn't automatable. Like, it's 1:49:29.712 --> 1:49:32.152 not stuff we could just set, uh, pen test bot 1:49:32.192 --> 1:49:35.952 on and it'll it'll fix the system. It's human problems. It's, uh, 1:49:35.992 --> 1:49:40.112 psychological problems. Um, it's like actual complex problems that require 1:49:40.112 --> 1:49:43.242 critical thinking versus. Let's scan this code and look for 1:49:43.242 --> 1:49:48.682 XSS or integer overflows. Um, I think the what seems 1:49:48.682 --> 1:49:51.921 so significant there is that we just weren't doing it. 1:49:51.962 --> 1:49:54.362 Like we have had the capability to automate a lot 1:49:54.362 --> 1:49:56.961 of this for a lot of time, but we've just 1:49:56.962 --> 1:49:58.002 decided not to. 1:49:58.122 --> 1:50:01.642 Um, doesn't that doesn't that apply to hundreds of millions 1:50:01.642 --> 1:50:02.242 of jobs? 1:50:04.162 --> 1:50:08.242 Um, I don't know. That's the thing is, the issue 1:50:08.242 --> 1:50:12.242 is kind of that the people aren't doing the jobs. Um, 1:50:12.442 --> 1:50:16.682 like one thing I've noticed, uh, with vulnerability research is 1:50:16.682 --> 1:50:20.122 a lot of times what will happen is a company will, uh, 1:50:20.122 --> 1:50:22.242 they'll fix a vulnerability in their code base, like a 1:50:22.242 --> 1:50:25.042 very specific type of vulnerability, and they won't go and 1:50:25.042 --> 1:50:27.322 look for that same vulnerability in other parts of their 1:50:27.322 --> 1:50:30.562 own code base. 100% agree they will. Yeah. So they 1:50:30.562 --> 1:50:35.522 will fix one very specific instance of vulnerability. And could 1:50:35.522 --> 1:50:37.602 we make a large language model that goes and finds 1:50:37.602 --> 1:50:40.242 the rest? Sure. But they could they could have and 1:50:40.242 --> 1:50:42.892 they should have done that. And essentially all we're doing 1:50:42.892 --> 1:50:46.092 is we're making this huge, expensive model to pick up 1:50:46.092 --> 1:50:48.772 the slack. That was just we just decided we could 1:50:48.812 --> 1:50:50.372 have done it. We just decided not to. 1:50:50.772 --> 1:50:53.172 So. So yeah, this that's a perfect example of what 1:50:53.172 --> 1:50:55.412 I'm saying. So check this out. That thing I told 1:50:55.412 --> 1:51:00.732 you about that bacteriophages like for researchers, the best in 1:51:00.732 --> 1:51:03.652 the entire world. It turns out that they had a 1:51:03.692 --> 1:51:06.852 bias in their brain. They thought that, um, I'm going 1:51:06.892 --> 1:51:08.252 to mess this up because I don't know anything about 1:51:08.252 --> 1:51:10.372 the science, but there's something about a head and a 1:51:10.372 --> 1:51:13.932 tail of a bacteriophage. And like they combine in a 1:51:13.932 --> 1:51:16.132 certain way to be able to move and go spread. 1:51:16.572 --> 1:51:18.412 And so they were making this thing in their mind 1:51:18.452 --> 1:51:20.292 of like, well, it can't be this and it can't 1:51:20.292 --> 1:51:22.612 be this because heads can only go on tails the 1:51:22.612 --> 1:51:25.492 following way. So the I was like, can't you just 1:51:25.492 --> 1:51:29.532 do this instead? That's probably why it's happening. When they 1:51:29.532 --> 1:51:31.292 both looked at it, they were like, oh my God, 1:51:31.292 --> 1:51:35.092 that is so simple. So a mental human human bias 1:51:35.732 --> 1:51:38.572 caused them to miss this, which would have propelled science 1:51:38.702 --> 1:51:41.742 forward for so long? So it's exactly the same thing. 1:51:41.982 --> 1:51:46.702 So basically, did I make a discovery there? It actually 1:51:46.702 --> 1:51:51.102 just found a human error. But but at scale. So 1:51:51.102 --> 1:51:53.102 so now everyone in the in the world can use 1:51:53.102 --> 1:51:56.022 this model. And they can now do the same thing 1:51:56.062 --> 1:52:01.461 to all this hidden research that's sitting inside this, this, um, this, um, 1:52:01.462 --> 1:52:04.421 archive of, like, raw data that's just sitting there. This 1:52:04.422 --> 1:52:07.982 raw data, in my opinion, contains, hey, you know, if 1:52:07.982 --> 1:52:10.622 you put this molecule to to this thing in this 1:52:10.622 --> 1:52:12.462 part of the cell, it will actually just do this 1:52:12.502 --> 1:52:15.782 and like, it'll change aging completely. And anybody who sees 1:52:15.782 --> 1:52:18.942 that when the AI says it, they'll be like, that 1:52:18.942 --> 1:52:23.342 was obvious. That was completely obvious. So somebody like yourself 1:52:23.342 --> 1:52:27.381 could say, well, that wasn't actually intelligence. That was actually 1:52:27.382 --> 1:52:30.382 just noticing something and applying a rule. And what I'm 1:52:30.382 --> 1:52:34.342 saying is that that mentality is missing the point of 1:52:34.342 --> 1:52:38.352 the benefit of the AI, because whether it's pentesting or 1:52:38.392 --> 1:52:43.072 finding new drugs, or finding new interactions or helping couples 1:52:43.392 --> 1:52:47.272 or solving, you know, helping people get healthier heart, it 1:52:47.272 --> 1:52:51.312 doesn't matter if it's basic, if it should be basic 1:52:51.312 --> 1:52:53.952 or simple for a human. I agree it should be, 1:52:53.952 --> 1:52:58.552 but it's not so. So actually doing these complex things 1:52:58.552 --> 1:53:03.711 of pen testing and marriage counseling and cardiac cardiology, doing 1:53:03.712 --> 1:53:06.272 those things at scale as good or better than most 1:53:06.272 --> 1:53:09.112 people do them, most humans do them is still a 1:53:09.112 --> 1:53:10.552 massive boon to society. 1:53:12.952 --> 1:53:16.711 I, I agree with, uh, like what you just said 1:53:16.712 --> 1:53:19.671 in a vacuum. The issue I have, which I think 1:53:19.672 --> 1:53:23.752 we we discussed previously on text, is that in order 1:53:23.752 --> 1:53:26.952 for these systems to exist and sustain that, human knowledge 1:53:26.952 --> 1:53:29.631 has to remain right. Like, we can't set and forget 1:53:29.632 --> 1:53:32.392 this AI, and then suddenly we don't need to learn 1:53:32.392 --> 1:53:36.722 this entire subject. Right. Um, and one of the really 1:53:36.722 --> 1:53:39.522 big problems, which I've actually blogged about a lot, is 1:53:39.522 --> 1:53:43.961 the amount of data these models need in order to even, like, 1:53:44.322 --> 1:53:48.042 understand at a third grade level, a area of expertise 1:53:48.042 --> 1:53:51.562 is phenomenal. And what happens is once we start relying 1:53:51.562 --> 1:53:54.602 on the systems, we lose our own knowledge because we're 1:53:54.602 --> 1:53:57.522 now just deferring it to this system. But the system 1:53:57.522 --> 1:54:01.362 cannot sustain without this constant, uh, incoming flow of knowledge. 1:54:01.402 --> 1:54:05.122 And my big worry is that the more we automate 1:54:05.122 --> 1:54:08.042 with AI, the more of the, the more skill loss 1:54:08.042 --> 1:54:10.802 we get on our end, which then down the line 1:54:10.802 --> 1:54:13.282 impacts our ability to feed it back into the AI 1:54:13.642 --> 1:54:17.482 and keep these AIS, uh, like in tip top shape. Um, 1:54:17.482 --> 1:54:21.962 which is where we would need like an actual genuine AGI. 1:54:22.282 --> 1:54:24.722 We would need an AI that at some point it 1:54:24.722 --> 1:54:28.442 can learn for itself. It can just, um, it can 1:54:28.442 --> 1:54:31.402 teach itself new skills, which I don't think is possible. 1:54:31.602 --> 1:54:33.642 So then we run the risk of. sure in the 1:54:33.642 --> 1:54:37.242 short term. Can we, like, automate these bug bounties? Sure. 1:54:37.282 --> 1:54:41.202 Can we, uh, like, find the solution to this bacteria thing? Sure. 1:54:41.562 --> 1:54:43.602 But in like five, ten years, when we now no 1:54:43.602 --> 1:54:47.322 longer have any pen testers and, uh, no longer any 1:54:47.322 --> 1:54:51.002 biologists or whatever the word is for people who do bacteria, 1:54:51.602 --> 1:54:55.442 then what? What happens now when, like a new problem arises, 1:54:55.482 --> 1:54:57.402 our AI needs to be trained on it. It needs 1:54:57.402 --> 1:55:00.482 a model built to address that problem. And suddenly, like, 1:55:00.482 --> 1:55:03.362 no one can do the thing anymore. It's actually, uh. Yeah, 1:55:03.802 --> 1:55:05.442 it reminds me of. I don't remember the name of 1:55:05.442 --> 1:55:07.722 the movie, but there's a movie where they're, uh. I 1:55:07.722 --> 1:55:10.962 believe the aliens, they become so advanced that they forget 1:55:10.962 --> 1:55:12.642 how to do all the basics, and they have to 1:55:12.642 --> 1:55:15.842 contact humanity and be like, yo, we actually kind of 1:55:15.882 --> 1:55:18.122 forgot how to do, like, this basic shit. Can you, uh, 1:55:18.122 --> 1:55:21.002 can you help us? Can you, like, bail us out, please? 1:55:21.962 --> 1:55:27.962 Yeah, no, I agree. Yeah. It's an interesting variation of, uh, thing. 1:55:27.962 --> 1:55:29.642 I don't know if you saw the MIT paper that 1:55:29.642 --> 1:55:32.692 basically said people who are using AI and kind of 1:55:32.732 --> 1:55:37.092 relying on it, their brains actually changed. Like, you could 1:55:37.092 --> 1:55:39.812 physically see in their brains that they were less smart 1:55:40.052 --> 1:55:43.852 as a result of like. Using it as a crutch. Yeah, yeah, 1:55:43.852 --> 1:55:46.052 it was crazy. The sample size was really small. It 1:55:46.052 --> 1:55:50.852 was only 54 people, but it was the first example of, um. 1:55:51.212 --> 1:55:55.932 Like tangible difference of, like somebody who's thinking for themselves versus, uh, 1:55:56.092 --> 1:56:01.292 versus using AI. So I do take your point there. Um, well, 1:56:01.292 --> 1:56:03.732 I think we've been going for a while. Uh, any 1:56:03.732 --> 1:56:07.852 any final thoughts? I got to go get someone from Bart, but, um, 1:56:07.852 --> 1:56:10.572 any final thoughts? We could also do a second version, too. So. 1:56:11.452 --> 1:56:13.732 Yeah, I'm down to do a follow up. Yeah. No, my. 1:56:13.732 --> 1:56:15.732 I did not actually see that paper, but that that 1:56:15.732 --> 1:56:19.132 is terrifying. Like, if people are actually losing, like, critical 1:56:19.132 --> 1:56:22.332 thinking ability, that's even worse than just, like, knowledge loss. 1:56:22.492 --> 1:56:24.692 That's like full on. We're just going to end up 1:56:24.692 --> 1:56:29.052 with Idiocracy. So that is that is very concerning. Um, 1:56:29.142 --> 1:56:31.662 but yeah, like my, my primary concerns have been purely 1:56:31.702 --> 1:56:34.502 like the economic model of it. If you've got these 1:56:34.542 --> 1:56:37.742 eyes just ingesting all this data, they're essentially stealing it. Like, 1:56:37.742 --> 1:56:40.102 people don't like me using that term, but they are 1:56:40.102 --> 1:56:43.662 just taking copyrighted content. And the people who make that content, 1:56:43.662 --> 1:56:46.782 they make money from posting it from ad revenue. Um, 1:56:46.782 --> 1:56:48.662 so my belief has always been that they would just, 1:56:48.942 --> 1:56:51.622 just collapse under their own weight, they would cannibalize their 1:56:51.622 --> 1:56:55.182 source of information, and they would fail that way. But 1:56:55.182 --> 1:56:58.942 if they're also making humans dumber, that is an even 1:56:58.942 --> 1:57:02.142 more pressing problem, because then we're just I mean, we're 1:57:02.142 --> 1:57:04.582 just doomed at that point. So I'm I'm hoping that 1:57:04.582 --> 1:57:07.222 turns out to just be like, the sample size was 1:57:07.222 --> 1:57:10.102 too small, and maybe there was like some sampling bias 1:57:10.102 --> 1:57:14.062 where there's like a, a correlation between unintelligent people and 1:57:14.062 --> 1:57:17.902 just deferring their critical thinking to AI. Uh, but if 1:57:17.902 --> 1:57:20.382 that is the case that like, it's actually making intelligent 1:57:20.382 --> 1:57:24.262 people dumber. Um, um, I think we're going to have 1:57:24.262 --> 1:57:26.702 to unite to, to do something about that. 1:57:27.072 --> 1:57:29.992 Yeah. Yeah. I unfortunately, I don't think it's going to 1:57:29.992 --> 1:57:32.552 be a sample size issue. I think that's going to 1:57:32.552 --> 1:57:35.992 be reproducible. Uh, my explanation for it, though, is that 1:57:35.992 --> 1:57:39.632 there are some people who will do this. Arguably even 1:57:39.632 --> 1:57:42.672 most people will do this. They'll just like rely, rely, rely. 1:57:42.672 --> 1:57:45.152 And pretty soon they won't be able to think for themselves. 1:57:45.152 --> 1:57:47.992 And that's that's extremely troubling. But there is also some 1:57:47.992 --> 1:57:51.792 other group, um, which I hopefully will include myself in, 1:57:52.032 --> 1:57:54.792 who is going to like, train my eye to constantly 1:57:54.792 --> 1:57:57.192 badger me in, like sort of a Socratic type of 1:57:57.232 --> 1:58:00.672 way to make sure that never happens. Right? Because because 1:58:00.672 --> 1:58:03.752 I'm going to be actively like defending against this. Um, 1:58:03.752 --> 1:58:06.312 so I think people like that, they're going to accelerate, 1:58:06.312 --> 1:58:08.712 they're going to get smarter, even better at critical thinking. 1:58:09.072 --> 1:58:12.512 But that doesn't speak to the issue of like the other, um, 1:58:12.752 --> 1:58:13.712 larger percent. 1:58:14.552 --> 1:58:17.432 Yeah, I think that's going to be a very, very 1:58:17.432 --> 1:58:19.592 large percentage because it's the same. It's the same thing 1:58:19.592 --> 1:58:22.792 that makes social media work and. Yeah, and makes, uh, 1:58:22.792 --> 1:58:26.962 llms so attractive. It's the the instant gratification that I 1:58:26.962 --> 1:58:28.882 don't have to put in too much work to get 1:58:28.882 --> 1:58:31.682 the results that I want. Um, so I think that 1:58:31.722 --> 1:58:34.602 it's almost it's basically dopamine addiction to an extent. It's like, yeah, 1:58:34.642 --> 1:58:36.482 people just love the idea of I don't have to 1:58:36.482 --> 1:58:38.322 go and read a hundred books on programming. I can 1:58:38.322 --> 1:58:41.122 just make an app. So I think, uh, what's going 1:58:41.122 --> 1:58:43.122 to happen is sure there will be people like you 1:58:43.162 --> 1:58:45.922 who who do make their AI like, keep them in, 1:58:45.962 --> 1:58:49.722 like just challenge them. But I think for the overwhelming 1:58:49.722 --> 1:58:53.122 majority of the population that like instant gratification is just 1:58:53.122 --> 1:58:54.762 going to take them down that rabbit hole of I'm 1:58:54.762 --> 1:58:56.842 just going to I'm not even going to think about 1:58:56.842 --> 1:58:58.402 the question I was just asked. I'm just going to 1:58:58.402 --> 1:59:02.002 type it into ChatGPT. Yeah. Um, which now is giving 1:59:02.002 --> 1:59:03.242 me a lot of anxiety. 1:59:04.362 --> 1:59:07.642 Yeah. Well, Marcus, this has been super fun. Uh, I 1:59:07.642 --> 1:59:10.842 think the conversation was great. And, um. Yeah, we should 1:59:10.842 --> 1:59:11.842 talk about doing a follow up. 1:59:11.842 --> 1:59:14.882 Maybe just challenge them, but I think for them. 1:59:15.122 --> 1:59:16.362 And, uh, talk to you soon. 1:59:17.362 --> 1:59:18.882 Awesome. Thanks so much for having me on. 1:59:19.202 --> 1:59:19.962 All right. See you.