WEBVTT - AI Mania Is Eviscerating Global Decision-Making w/ Nik Suresh 0:00:02.480 --> 0:00:06.200 Also media greeting, Hello and salutations. 0:00:06.240 --> 0:00:19.439 I am at Zetron and this is better offline. 0:00:20.360 --> 0:00:22.880 We are back. You need to subscribe to the newsletter, 0:00:22.920 --> 0:00:25.760 you need to download the podcast you're already listening to 0:00:25.800 --> 0:00:28.639 take your friend's phones downloaded on there. You will have 0:00:28.720 --> 0:00:30.880 failed me if you don't do this, not really, I 0:00:30.920 --> 0:00:33.879 love you all dearly. But look today I am joined 0:00:33.880 --> 0:00:37.040 by the incredible writer and consultant Nick Siesh, who two 0:00:37.080 --> 0:00:39.040 weeks ago put out one of the best pieces on 0:00:39.120 --> 0:00:41.519 AI I have ever read, and not just because it confirms 0:00:41.560 --> 0:00:44.400 my biases and it's called AI mania, is a viscerating 0:00:44.440 --> 0:00:46.479 global decision making Nick, thank you for joining me. 0:00:47.159 --> 0:00:48.800 Hey, thank you so much for having me on again. 0:00:49.440 --> 0:00:52.080 This piece was terrifying. I just want to be really 0:00:52.120 --> 0:00:54.880 clear that, like the things that I read in here 0:00:55.520 --> 0:00:57.320 a lot of them, I was like, I assumed this 0:00:57.520 --> 0:01:00.680 was happening. I assumed that it was the bad. But 0:01:01.560 --> 0:01:04.560 to summarize the piece in short, it appears that just 0:01:04.760 --> 0:01:10.160 AI is this bordering on religious thing within business now 0:01:10.440 --> 0:01:14.640 that people get into, regardless of whether it even does anything. 0:01:16.120 --> 0:01:19.040 Yeah, yeah, and you know, I think the thing that 0:01:19.120 --> 0:01:22.080 got it so widely circulated is even that people who 0:01:22.600 --> 0:01:25.039 use AI in their work and say they're seeing benefits. 0:01:25.440 --> 0:01:27.600 A lot of the more sensible people are looking at 0:01:27.600 --> 0:01:31.360 their colleagues and going, what are you guys doing? Right? Like, 0:01:31.920 --> 0:01:33.880 I think I'm a little bit faster at my job, 0:01:33.920 --> 0:01:37.040 And here we have my CEO saying, let's just scrap 0:01:37.200 --> 0:01:39.920 every other project. We'll remove all the employees. We're gonna 0:01:39.920 --> 0:01:43.520 have one hundred x productivity. And even sensible people that 0:01:43.760 --> 0:01:46.080 enjoy it are getting a little bit worried by the 0:01:46.160 --> 0:01:47.200 level themania's gotten to. 0:01:48.040 --> 0:01:51.360 Yeah, so give me, give me an example here. The 0:01:51.400 --> 0:01:54.600 first section is called AI investments are generally total failures. 0:01:54.600 --> 0:01:57.320 And I realize the title really does say it all there, 0:01:57.720 --> 0:02:00.880 but can you break that down for me? Yeah. 0:02:00.920 --> 0:02:04.880 So it's interesting because you know, most projects at companies 0:02:04.960 --> 0:02:08.400 do fail, especially big companies. They're just bad at running 0:02:08.440 --> 0:02:10.680 tech projects. I've written about this well before the AI 0:02:10.800 --> 0:02:13.800 thing was big. What is fascinating is that they have 0:02:13.880 --> 0:02:17.320 an even higher failure rate for the AI projects. And again, 0:02:17.720 --> 0:02:20.280 like just to we don't need to get into an 0:02:20.400 --> 0:02:22.640 argument with people about whether it works or not, right, 0:02:22.680 --> 0:02:26.120 we can just start with the basis of normal it 0:02:26.400 --> 0:02:30.119 projects fail a bunch, right, AI projects are novel technology. 0:02:30.240 --> 0:02:33.480 So these are failing even more. And the protective factor 0:02:33.480 --> 0:02:36.000 at a companies normally that they don't want to spend 0:02:36.040 --> 0:02:37.919 all their money on it. They actually want to spend 0:02:38.120 --> 0:02:40.600 the smallest amount of it possible. Right, it's mostly a 0:02:40.639 --> 0:02:43.440 thing you're it's a cost for the business. Right With AI, 0:02:43.520 --> 0:02:46.720 this is totally inverted, and companies are going, I'd like 0:02:46.800 --> 0:02:48.920 to take as much budget as I can feasibly get, 0:02:49.480 --> 0:02:51.119 and I'm going to put it all into this thing 0:02:51.160 --> 0:02:53.000 that I know I'm bad at, and I want to 0:02:53.000 --> 0:02:55.960 do a harder version of that, but I can't admit 0:02:56.000 --> 0:02:58.160 to anyone I've just wasted millions and millions of dollars. 0:02:58.560 --> 0:03:01.760 So it's sort of like all the organizations in the 0:03:01.760 --> 0:03:04.040 planet are going all in on the thing they know 0:03:04.080 --> 0:03:06.560 they're bad at, playing it on hard mode, and then 0:03:06.639 --> 0:03:08.720 sort of just lying about it after the fact. 0:03:08.919 --> 0:03:11.280 And when you say lying, what do you mean that? 0:03:11.320 --> 0:03:13.960 They just being like, Yeah, this is making me one 0:03:14.040 --> 0:03:18.880 bazillion more productive percents. I'm so good at AI. I 0:03:18.919 --> 0:03:23.040 love it. Like what is that? Like? What are they doing? 0:03:23.320 --> 0:03:24.960 What are they actually do? Get this crap? 0:03:25.760 --> 0:03:28.200 Well, okay, so it happens in a couple of ways. 0:03:28.280 --> 0:03:31.680 Right when I say lying, that sounds like very intentional. 0:03:31.720 --> 0:03:33.640 And there are people that lie, because there are people 0:03:33.639 --> 0:03:37.880 that lie everywhere, but usually it's two things. Like internally 0:03:37.960 --> 0:03:40.440 it can be a lot of delusion. There are a 0:03:40.440 --> 0:03:44.120 lot of executives out there who have funded AI projects 0:03:44.640 --> 0:03:48.280 and they don't realize that the whole organization has sort 0:03:48.320 --> 0:03:52.760 of shifted into presenting them with the rosiest view of 0:03:52.760 --> 0:03:55.080 what is happening that is possible. Like if you're a 0:03:55.120 --> 0:03:58.080 CEO who has heard good things about AI, and let's 0:03:58.080 --> 0:04:01.240 say you've got two guys underneath you gunning for promotion, Right, 0:04:01.400 --> 0:04:03.480 those two guys are now at an arms race to 0:04:03.560 --> 0:04:06.040 talk about their AI productivity and how much AI they're 0:04:06.120 --> 0:04:08.320 rolling it across the business, Like that is the goal. 0:04:08.360 --> 0:04:11.760 It's not to make money, it's to be the most AI, 0:04:12.360 --> 0:04:15.480 whatever the fuck that means. Right, So it's pretty nonsensic, 0:04:15.840 --> 0:04:20.000 But that's that's self creation. It's it's very so good 0:04:20.080 --> 0:04:23.560 much And again it's it's an extreme version of what 0:04:23.600 --> 0:04:26.640 we already see, right, because people already lie to CEOs 0:04:26.640 --> 0:04:28.640 for promotions. This is just that on steroids. 0:04:29.440 --> 0:04:32.039 So you mentioned in here you say you've seen zero 0:04:32.120 --> 0:04:35.320 percent success in a year and a half. Really is 0:04:35.320 --> 0:04:37.599 it really zero? You've seen nothing? I just want to 0:04:37.600 --> 0:04:40.080 be my owner is going to eat this up, I realize, 0:04:40.120 --> 0:04:42.880 but like I just want to make sure, yes we 0:04:42.920 --> 0:04:46.520 should you know, add one asterisk there, because some people 0:04:46.520 --> 0:04:50.800 misunderstood that me seeing zero percent doesn't mean like it's 0:04:50.880 --> 0:04:53.360 zero percent across the whole world, because I haven't seen 0:04:53.400 --> 0:04:56.320 every project in the world. Sadly, some people don't have 0:04:56.480 --> 0:04:58.400 enough reading comprehension for that, so you have to lay 0:04:58.400 --> 0:05:00.599 it out or you get like an AI boo who's like, 0:05:00.640 --> 0:05:04.919 it's not literally zero. Some must be getting some results somewhere. 0:05:05.440 --> 0:05:08.880 I have seen nothing, and I've been doing nothing but 0:05:09.160 --> 0:05:12.040 sales and on sites at companies like all across Australia 0:05:12.040 --> 0:05:14.400 in the US for a year and a half. Now 0:05:14.480 --> 0:05:15.760 I don't get brought in to do them. 0:05:15.760 --> 0:05:18.000 They're like things I see on the side and I go, well, 0:05:18.040 --> 0:05:19.440 that's that is not working. 0:05:20.520 --> 0:05:20.680 Well. 0:05:20.720 --> 0:05:23.359 The thing I think about when I hear that people like, actually, 0:05:23.600 --> 0:05:27.520 you haven't saw every software thing. It's just this is 0:05:27.560 --> 0:05:30.880 how AI boosters work. They want you, as the defender 0:05:30.960 --> 0:05:34.599 of reality, to be like prove unilaterally this is the 0:05:34.680 --> 0:05:37.479 AI is terrible. If it's so bad, prove it. But 0:05:37.520 --> 0:05:40.440 when you ask them to prove anything, they look at 0:05:40.480 --> 0:05:44.560 you and go, well, you know, just the numbers, just 0:05:44.640 --> 0:05:48.040 the biggest, the number number I made so big this year, 0:05:48.279 --> 0:05:51.480 It's the biggest I've ever seen. And they refuse to 0:05:51.520 --> 0:05:53.640 show you anything real. I think I've been sent the 0:05:53.640 --> 0:05:57.640 fucking Erdos problem thing fifteen times now, and the next 0:05:57.680 --> 0:05:59.800 time someone sends it to me, I'm gonna draw a gut. 0:05:59.839 --> 0:06:02.640 That is a joke. That's a joke. I'm not gonna 0:06:02.880 --> 0:06:04.520 not gonna do anything with a gun, but I am 0:06:04.600 --> 0:06:09.799 just gonna possibly imagine the saw from Jigsaw, sorry, Jigsaw 0:06:09.800 --> 0:06:11.320 from saur Even they're gonna get me in the comments 0:06:11.360 --> 0:06:13.640 with that one, just kind of tying them to radiate 0:06:13.680 --> 0:06:16.080 and say, sure, sure, tell me what you think the 0:06:16.200 --> 0:06:19.039 Erdos problem means. Tell me what it means, Tell me 0:06:19.080 --> 0:06:21.440 what you think eros means, because that's the thing. These 0:06:21.440 --> 0:06:24.480 people don't have anything tangible to share, and your peace 0:06:24.600 --> 0:06:26.440 kind of I read it in an airport and I 0:06:26.480 --> 0:06:28.919 was just sitting there going like fucking hell. And I 0:06:28.920 --> 0:06:31.039 did get a few looks because I was so surprised 0:06:31.040 --> 0:06:35.440 by it. But it feels like your peace crystallizes everything. 0:06:35.480 --> 0:06:38.520 I've been thinking for two years that just this is 0:06:38.600 --> 0:06:42.200 this weird death of stylein style situation where people are 0:06:42.279 --> 0:06:45.120 just saying, God, damn, I love AI because if they 0:06:45.160 --> 0:06:48.280 don't say it, they'll make their boss feel bad, their 0:06:48.680 --> 0:06:51.840 customers feel bad, and their colleagues feel bad despite none 0:06:51.839 --> 0:06:53.919 of them getting anything out of AI. 0:06:55.320 --> 0:06:57.400 Yeah, you know, the simple way to think about it 0:06:57.440 --> 0:06:59.800 is that any given business, you know, you have one 0:07:00.040 --> 0:07:02.599 problem generally that is bottlenecking how you're making most of 0:07:02.600 --> 0:07:06.039 your money, and whatever you think AI can do, it's 0:07:06.120 --> 0:07:08.599 really only the correct thing to deploy that solves that 0:07:08.680 --> 0:07:11.720 one problem. It doesn't matter if you write more code, 0:07:11.920 --> 0:07:14.120 if writing code isn't the way to make more money. 0:07:15.360 --> 0:07:18.960 And I mean it's basically just like as simple as that, right, 0:07:19.000 --> 0:07:21.000 people are going all in on a thing that doesn't 0:07:21.040 --> 0:07:23.160 really make sense sort of for context they're inhabiting. 0:07:23.720 --> 0:07:25.560 So you mentioned within the piece that you had to 0:07:25.600 --> 0:07:29.120 basically stop offering AI services or a specific one forgive 0:07:29.120 --> 0:07:31.240 me if I forgive me if I misread that, but 0:07:31.280 --> 0:07:36.440 you just won't mention AI services within software projects because 0:07:36.480 --> 0:07:38.120 people go nutrageous over them. 0:07:38.720 --> 0:07:41.680 Yeah. What we found was, you know, we're a relatively 0:07:41.720 --> 0:07:44.440 small business and we optimize a lot for working with people. 0:07:44.840 --> 0:07:46.440 We entry a software consultancy. 0:07:46.520 --> 0:07:49.080 Just to be clear, Yes, we're a software consultancy and 0:07:49.320 --> 0:07:52.200 data specifically. So my postgrad degree was a Masters in 0:07:52.240 --> 0:07:55.760 data science. Like I was into AI nominally before the 0:07:56.000 --> 0:07:59.880 whole LM thing took off. And yeah, what we found 0:08:00.200 --> 0:08:03.800 was whenever we mentioned AI, two things would happen. And 0:08:03.840 --> 0:08:07.320 the one I mentioned in the piece was people would 0:08:07.360 --> 0:08:11.680 become rabidly interested in buying from us. But they were 0:08:11.720 --> 0:08:15.080 also really bad customers, Like they made bad decisions, they 0:08:15.080 --> 0:08:17.600 had cash flow issues, they didn't think clearly, Like, these 0:08:17.600 --> 0:08:20.560 were not people we wanted to work with as a 0:08:20.560 --> 0:08:24.040 matter of lifestyle. But also, you know, it's like a 0:08:24.040 --> 0:08:27.120 lot of startups and stuff, the sort of customer where 0:08:27.160 --> 0:08:29.520 used to take all the money upfront because you're worried 0:08:29.560 --> 0:08:31.320 they're just going to scam you, Like, these are not 0:08:31.560 --> 0:08:34.000 people you want to be in bed with. And the 0:08:34.000 --> 0:08:37.120 people we really liked, we would mention it just as 0:08:37.160 --> 0:08:40.880 an aside, and they would immediately become so disinterested because 0:08:40.880 --> 0:08:43.360 they just go, well, you know a lot of people 0:08:43.400 --> 0:08:44.840 are trying to scam you with this right now, I'm 0:08:44.840 --> 0:08:47.840 just not interested. So we just don't talk about it 0:08:47.840 --> 0:08:49.760 at all in any context. We don't like doing it. 0:08:49.840 --> 0:08:53.400 The work's really boring anyway. But yeah, like it's just 0:08:54.600 --> 0:08:57.320 here's another thing I can tell you. Our clients that 0:08:57.400 --> 0:09:00.120 don't talk about AI at all are like far in 0:09:00.120 --> 0:09:02.720 a way, the most profitable because they're focused on whatever 0:09:02.760 --> 0:09:05.199 their biggest problem is at any given moment. It's very 0:09:05.320 --> 0:09:07.280 very simple, Like this is not rocket science. 0:09:07.360 --> 0:09:09.679 Yeah, it's just like it's like the companies that are 0:09:09.679 --> 0:09:12.600 doing well are focusing on solving business issues. The companies 0:09:12.640 --> 0:09:16.160 that are doing badly are the ones getting very expensive 0:09:16.160 --> 0:09:17.439 digital fidget spinners. 0:09:18.280 --> 0:09:20.880 Yeah, like it's it's as simple as that. I don't 0:09:20.920 --> 0:09:25.160 know many companies where llms unblock, like whatever their core 0:09:25.200 --> 0:09:29.240 issue is, there's stuff I can think, like maybe customer support, 0:09:29.240 --> 0:09:32.240 but I don't know much about customer support anything where 0:09:32.400 --> 0:09:35.200 I have some subject matter expertise. I'm like, an LLM 0:09:35.280 --> 0:09:36.960 is not the most important thing you need right now. 0:09:37.760 --> 0:09:41.520 I can't think of it with two, three, four boobless 0:09:41.559 --> 0:09:44.960 years into this bullshit, and I am a spreadsheet hog 0:09:45.240 --> 0:09:47.680 and it doesn't do spreadsheets very well. I will say, 0:09:47.679 --> 0:09:49.480 people can be people will be very mad at me 0:09:49.520 --> 0:09:53.360 for this. I've found one use of llm's one. And 0:09:53.480 --> 0:09:55.480 I think I've told you about this the Bloomberg terminal, 0:09:55.920 --> 0:09:58.559 because the way that you the way that you request 0:09:58.559 --> 0:10:00.800 stuff on the Bloomberg terminal is the three Python. I 0:10:00.840 --> 0:10:03.840 don't write Python, so you'd be like Bloomberg, l of 0:10:03.880 --> 0:10:06.040 them tell me out to get this, and it gives 0:10:06.040 --> 0:10:09.440 you the code and it runs it boom simple. Probably 0:10:09.480 --> 0:10:13.200 the least you like, literally just sexy search. Well done, 0:10:13.440 --> 0:10:16.240 that one point one trillion dollars of capex. I'm sure 0:10:16.280 --> 0:10:19.160 that was all worth there. But what's funny is the 0:10:19.679 --> 0:10:22.640 example you've given. The piece of the AI chatbot thing 0:10:22.679 --> 0:10:25.560 for Snowflake feels like the things that LM should be 0:10:25.600 --> 0:10:30.360 good at, which is a database request, right, Like, yeah, 0:10:30.920 --> 0:10:33.160 why doesn't it work? Surely that is the thing that 0:10:33.160 --> 0:10:36.040 this should be good at. What is the issue there 0:10:36.160 --> 0:10:39.160 and why don't LLLM solve them? Yeah? 0:10:39.280 --> 0:10:41.960 The complexity there is you know, I don't know much 0:10:41.960 --> 0:10:45.000 about Bloomberg terminals, although it's very funny everything you've told 0:10:45.000 --> 0:10:45.480 me about. 0:10:45.800 --> 0:10:48.680 If you tell it's an autism portal, it's so good. 0:10:50.080 --> 0:10:53.599 Yeah, So you know, I imagine they're fairly constrained in 0:10:53.679 --> 0:10:56.080 what they can do. Yeah, For example, you know, you 0:10:56.120 --> 0:10:59.000 mentioned writing Python, which means someone has written a Python 0:10:59.080 --> 0:11:04.160 library that interfaces with the Bloomberg terminal and constrains what 0:11:04.320 --> 0:11:06.760 sorts of things you can ask it. So if it's 0:11:06.880 --> 0:11:08.920 very well designed and you put all the text in there, 0:11:09.040 --> 0:11:11.400 the same thing that lets me a human interrogate it with, 0:11:11.640 --> 0:11:14.080 let an LM do it. What's very hard at a 0:11:14.120 --> 0:11:16.040 business is a lot of that stuff is just not 0:11:16.120 --> 0:11:18.520 written down, or it exists in someone's head or the 0:11:18.520 --> 0:11:22.040 real world is very, very complicated, and the very people 0:11:22.080 --> 0:11:25.120 who are Let's use this example, right, why does cortex 0:11:25.160 --> 0:11:27.360 not work? So you go in and you're the CFO 0:11:27.400 --> 0:11:29.520 and you type in I would like to know much 0:11:29.559 --> 0:11:31.920 revenue we made last year. It doesn't know all the 0:11:31.960 --> 0:11:35.240 little nuances of your business, and for something really important 0:11:35.280 --> 0:11:37.440 you need it to be correct. But if you're making 0:11:37.520 --> 0:11:40.400 use of an LLM and you're the CFO, you can't 0:11:40.520 --> 0:11:43.959 debug whatever it put out. And that's okay when the 0:11:44.040 --> 0:11:46.880 number is totally wrong, because you'll go, hey, this isn't correct. 0:11:47.440 --> 0:11:49.720 If it's off by like five percent or ten percent, 0:11:49.800 --> 0:11:54.200 you probably don't know and exactly would like with your 0:11:54.240 --> 0:11:56.960 Bloomberg terminal. Right, if you looked up something really important. 0:11:57.280 --> 0:11:59.040 You might reach out to a friend and say, is 0:11:59.320 --> 0:12:01.200 this Python court because I'm about to spend a lot 0:12:01.240 --> 0:12:01.800 of money on this. 0:12:02.320 --> 0:12:04.760 Yeah. And to be clear, when you make a request 0:12:04.800 --> 0:12:06.880 on the terminal, it does the Python. It's just the 0:12:06.880 --> 0:12:11.000 Python code to request Bloomberg's actual data, like it hasn't. 0:12:11.040 --> 0:12:13.280 I imagine that I don't write Python, but I imagine 0:12:13.280 --> 0:12:15.360 there is a library in this, that this is just 0:12:15.440 --> 0:12:17.360 a layer on top of it so that people like 0:12:17.400 --> 0:12:21.080 me can actually use it without just punching the typing 0:12:21.160 --> 0:12:23.960 in like three letter things and making things pop up 0:12:24.000 --> 0:12:25.760 and then spending hours fucking with it, which of course 0:12:25.800 --> 0:12:27.679 I never did before. I use the LMS side homes 0:12:28.080 --> 0:12:28.440 of course. 0:12:28.640 --> 0:12:30.079 Yeah, yeah, that's pretty great. 0:12:30.400 --> 0:12:33.840 But with Cortex, the search side, it seems that the 0:12:33.880 --> 0:12:37.720 problem is also that is it making stuff up? Because 0:12:37.800 --> 0:12:42.040 I'm not being an ignoramus here. Surely looking over a 0:12:42.080 --> 0:12:44.560 big data based on information and pulling stuff out is 0:12:44.559 --> 0:12:45.959 what an l of them is good at. Or is 0:12:46.000 --> 0:12:48.560 the suggestion that the databased layer is not something ALMS 0:12:48.559 --> 0:12:49.040 can solve. 0:12:49.920 --> 0:12:51.120 It's just very complicated. 0:12:51.120 --> 0:12:51.320 You know. 0:12:51.360 --> 0:12:54.120 So we were at a client site yesterday and you 0:12:54.120 --> 0:12:56.800 know they have this ledger internally, which tracks like all 0:12:56.800 --> 0:12:59.920 the transactions coming in and out, and there's like forty 0:13:00.040 --> 0:13:03.160 different account types stored in all these different tables, and 0:13:03.240 --> 0:13:06.520 in different situations you want to use different types of 0:13:06.520 --> 0:13:10.679 transactions to calculate certain things around like service utilization, and 0:13:10.679 --> 0:13:12.640 and LLM looks at that and it just goes, well, 0:13:12.679 --> 0:13:14.240 I haven't worked here for eight years, so I don't 0:13:14.240 --> 0:13:15.400 know how to run this calculation. 0:13:15.920 --> 0:13:18.800 Like it's so the data is so it's that the 0:13:18.920 --> 0:13:21.839 data isn't labeled properly and all these others. So it's 0:13:21.920 --> 0:13:25.719 how the data is presented and looked over. Yeah, yeah, you. 0:13:25.640 --> 0:13:28.440 Can even label it properly. But there's a reason we 0:13:28.520 --> 0:13:31.400 have human specialists. It's the same reason I don't get 0:13:31.400 --> 0:13:34.200 an LLLM to write all my contracts. I could, but 0:13:34.240 --> 0:13:36.120 it's that I go to my lawyer Eed, and I go, 0:13:36.880 --> 0:13:38.200 what should I actually do here? 0:13:38.280 --> 0:13:41.679 You know, like the Yeah, it's experience versus data. 0:13:42.120 --> 0:13:43.800 Right, you know, Like when we started the business, he 0:13:43.840 --> 0:13:45.559 wrote a contract for me that you know, it's like 0:13:45.640 --> 0:13:47.760 a standard one and it has things in there like 0:13:47.840 --> 0:13:52.720 I wouldn't have thought of, like, for example, the agreement 0:13:52.760 --> 0:13:55.800 we have with clients on how long we take between 0:13:55.840 --> 0:13:56.920 email responses and. 0:13:57.160 --> 0:14:01.840 Ye, when I'm still doing PR stuff, that's exactly That's 0:14:01.880 --> 0:14:04.040 exactly what I had in there, like basic things that 0:14:04.320 --> 0:14:07.120 you wouldn't think of unless you'd experienced the thing. But 0:14:07.200 --> 0:14:11.000 it's funny because even now, even people who are anti LLM, 0:14:11.200 --> 0:14:13.720 I will hear them say, but it's good for searching 0:14:13.800 --> 0:14:16.880 large databases, and it sounds like it isn't like AI 0:14:17.000 --> 0:14:17.800 is just not good at that. 0:14:19.240 --> 0:14:22.440 I specialize in this specific area, which is just like 0:14:22.480 --> 0:14:26.240 big information retrieval. It's pretty good at you know, the 0:14:27.600 --> 0:14:31.640 a database specifically in tech means something. It means systems 0:14:31.640 --> 0:14:34.480 like Postcrest or whichever, which means nothing to normal people. 0:14:35.440 --> 0:14:38.880 I would say they're pretty good at reading gigantic amounts 0:14:38.880 --> 0:14:41.880 of texts and then pointing a human at the correct place. 0:14:42.280 --> 0:14:44.200 So a good use is if we're like in a 0:14:44.240 --> 0:14:46.440 new client site and they have a million lines of 0:14:46.880 --> 0:14:50.400 some fucking thing, I can go, I need to look 0:14:50.440 --> 0:14:53.440 for the places where this logic might happen, and they'll 0:14:53.440 --> 0:14:55.680 point me at them and it's still wrong sometimes, but 0:14:55.680 --> 0:14:59.360 at least I can look. That's not useless. But it's 0:14:59.400 --> 0:15:00.920 like how much what I pay for that? 0:15:01.720 --> 0:15:02.240 I don't know? 0:15:02.440 --> 0:15:05.119 Maybe like twenty dollars a month. 0:15:06.200 --> 0:15:08.680 But that's not that. But that's the thing. It's like, 0:15:08.760 --> 0:15:11.000 even with the terminal, it's the same thing. It will 0:15:11.040 --> 0:15:13.760 get things wrong. It never gets the numbers wrong because 0:15:13.760 --> 0:15:16.680 when it runs the numbers, it's pulling them from a 0:15:16.720 --> 0:15:19.480 specific place. And I still tell it, if I ever 0:15:19.520 --> 0:15:22.280 touch the LM stuff, run it again, run it again, 0:15:22.560 --> 0:15:24.280 just like and then I'll google it and I'll look 0:15:24.320 --> 0:15:26.520 up the stuff because I don't trust anything that touches them. 0:15:26.960 --> 0:15:29.800 But still, it's just it is a phishing expedition. You 0:15:29.840 --> 0:15:32.280 can't rely on it to have any logic, which I 0:15:32.320 --> 0:15:35.120 think is what most people want from search when the 0:15:35.200 --> 0:15:38.160 human beings are pretty terrible about asking for exactly what 0:15:38.200 --> 0:15:41.040 they want, and I feel like it just falls down. 0:15:41.280 --> 0:15:42.960 And the reason I keep going on about this stuff 0:15:43.000 --> 0:15:47.120 is because searching big databases and making decisions feels like 0:15:47.240 --> 0:15:50.600 ninety percent of the LLLM integrations I've heard about, like, 0:15:50.680 --> 0:15:53.360 it's just how do I do this? And what does 0:15:53.400 --> 0:15:56.080 this mean? Is how people are kind of seem to 0:15:56.120 --> 0:15:58.360 at least be describing how it's useful in their business, 0:15:59.080 --> 0:16:01.160 but it can't do that, which is bad. 0:16:01.280 --> 0:16:03.960 Yeah, it at the end of the day, as long 0:16:03.960 --> 0:16:06.520 as we're constrained by having to go in and check 0:16:06.560 --> 0:16:08.920 whether the results are correct, even if we grant that 0:16:08.960 --> 0:16:11.360 they're correct most of the time, which is not always true. 0:16:11.400 --> 0:16:15.200 It's very domain specific. That just means you can't get 0:16:15.200 --> 0:16:18.119 one hundred times faster. And that's what's justifying the bajillion 0:16:18.160 --> 0:16:21.400 million infinity dollars of investment. Right if we went, hey, 0:16:21.400 --> 0:16:26.680 this makes you fifty percent faster, Like, all right, is 0:16:26.680 --> 0:16:30.000 that worth a trillion dollars? I I don't think so. 0:16:30.080 --> 0:16:33.000 I don't think it's worth trillion dollars, Like the market 0:16:33.000 --> 0:16:35.240 thesis depends on the psychosis claim. 0:16:36.080 --> 0:16:39.280 Also, I mean it's meant to be artificial intelligence, just 0:16:39.320 --> 0:16:41.800 to be clear, that's the intelligence part is in the name. 0:16:42.840 --> 0:16:45.960 But everything was saying makes it sound really fucking stupid, 0:16:46.840 --> 0:16:48.960 just like it's just it's not smart to do. It 0:16:48.960 --> 0:16:52.480 doesn't know anything and it does dumb guys stuff. Oh 0:16:52.520 --> 0:16:55.240 you didn't ask me in exactly the way I needed 0:16:55.280 --> 0:16:58.840 to be asked. I don't understand, mate. Sorry, I'm not 0:16:58.880 --> 0:17:01.760 going to be helpful at all. And also, what we 0:17:01.840 --> 0:17:04.840 are describing is very simple stuff. It's not a trillion dollar, 0:17:04.840 --> 0:17:08.359 it's not even a fifty billion tam industry. It's just 0:17:08.920 --> 0:17:11.520 even if it got everything was saying right, it still 0:17:11.560 --> 0:17:13.360 wouldn't be that exciting. 0:17:15.400 --> 0:17:17.840 Yeah, you know, I think it was. It was very 0:17:17.840 --> 0:17:22.160 interesting when it came out. It was like pretty novel. 0:17:22.160 --> 0:17:24.200 When GPT three came out, I went with the hell 0:17:24.280 --> 0:17:26.399 like this was not available when I was at university. 0:17:26.440 --> 0:17:30.760 This is fucking wild. I think this really sort of overstate. 0:17:30.840 --> 0:17:34.720 It's welcome, and it's just so interesting because we've taken 0:17:34.920 --> 0:17:37.520 a technology that I mean, if you're willing to set 0:17:37.560 --> 0:17:40.760 aside many many ethical claims around stealing from artists and 0:17:40.840 --> 0:17:43.080 environmental things, like if we just handwave all of that, 0:17:43.640 --> 0:17:45.439 which I'm not happy to do, but in this context 0:17:45.480 --> 0:17:46.360 for an argument, I will. 0:17:46.480 --> 0:17:47.240 Yeah, yeah, if you. 0:17:47.200 --> 0:17:50.680 Handwave all of that, It's like, all right, we've made 0:17:50.680 --> 0:17:54.560 something pretty neat, and then we've still somehow completely managed 0:17:54.560 --> 0:17:57.240 to fuck everything up. Like all we had to do 0:17:57.400 --> 0:17:59.520 was be reasonable for ten seconds about how we use 0:17:59.560 --> 0:18:02.000 this in advent than it, and we have completely failed. 0:18:02.040 --> 0:18:04.560 Everyone has gone insane. 0:18:14.560 --> 0:18:17.840 Yeah, it's and I'm trying. It's like if everyone did 0:18:17.920 --> 0:18:19.800 platform as a service and take it to a church, 0:18:19.880 --> 0:18:21.480 It's like, Wow, you could just put a single line 0:18:21.480 --> 0:18:25.080 of code. Now your entire ecosystem is working perfectly on this. Yeah, 0:18:25.080 --> 0:18:27.200 it sounds like the promise that was made with pass 0:18:27.280 --> 0:18:29.000 and then just didn't happen. I worked on a few 0:18:29.000 --> 0:18:32.000 of those clients. It was fucking terrible, explained to journalists. 0:18:32.400 --> 0:18:35.720 Platform was a service meant. But again, no, everyone didn't 0:18:35.720 --> 0:18:38.480 suddenly say that no one was running their own computers anymore, 0:18:38.480 --> 0:18:41.040 no one was running their own databases anymore. And it's 0:18:41.119 --> 0:18:44.040 just it feels like it's completely escaped reality and you 0:18:44.080 --> 0:18:46.959 are very much in the trenches with this. And one 0:18:47.000 --> 0:18:48.560 thing I did want to ask about as well, was 0:18:49.200 --> 0:18:52.960 this my favorite section of the piece, heretics will be shot. 0:18:54.320 --> 0:18:56.719 And the way you describe it is that there is 0:18:56.880 --> 0:19:02.560 this genuine religiosity behind it, this this aggressive kind of 0:19:03.160 --> 0:19:06.520 tech evangelism that says, if you were not in with 0:19:06.600 --> 0:19:08.639 the AI stuff, you were out of the crowd. 0:19:09.880 --> 0:19:13.080 Yeah, and it's actually kind of scary to watch. Like, 0:19:13.160 --> 0:19:15.560 normally when I write something, I'm mostly angry about it, 0:19:15.680 --> 0:19:17.280 and you also wasn't a lot of anger in that 0:19:17.320 --> 0:19:21.160 piece because I was more wowly realizing. Yeah, I was like, hey, 0:19:21.600 --> 0:19:24.560 there are not adults running our hospitals, you know. It 0:19:24.600 --> 0:19:28.360 turns out, I think the worst thing was realizing that 0:19:28.440 --> 0:19:31.560 they were always sort of immature and childlike and not 0:19:31.680 --> 0:19:34.760 responsible or brave, like they've always been that way. This 0:19:34.840 --> 0:19:37.480 has just allowed them to coordinate around a thing that 0:19:37.520 --> 0:19:40.160 allows them to behave this way at the same time. Yeah, 0:19:40.200 --> 0:19:42.360 and anyone that does not get on board, like God, 0:19:42.400 --> 0:19:45.280 I've seen some pretty wild stuff. We're sort of at 0:19:45.320 --> 0:19:48.119 the point where if you tell as CEO at maybe 0:19:48.320 --> 0:19:52.120 most companies, especially big ones, right if they're worth over 0:19:52.160 --> 0:19:55.600 a billion dollars, if you tell certain people, certain executives 0:19:55.680 --> 0:19:57.879