1 00:00:01,070 --> 00:00:04,430 S1: Welcome to Unsupervised Learning, a security, AI and meaning focused 2 00:00:04,430 --> 00:00:07,280 S1: podcast that looks at how best to thrive as humans 3 00:00:07,280 --> 00:00:11,539 S1: in a post AI world. It combines original ideas, analysis, 4 00:00:11,539 --> 00:00:14,750 S1: and mental models to bring not just the news, but 5 00:00:14,750 --> 00:00:21,950 S1: why it matters and how to respond. Hey what's up? 6 00:00:21,950 --> 00:00:26,960 S1: It's Daniel with Unsupervised learning. I'm building AI to upgrade humans, 7 00:00:26,960 --> 00:00:29,060 S1: and today I want to talk about a cool idea. 8 00:00:29,090 --> 00:00:32,419 S1: Basically trying to figure out how to predict how much 9 00:00:32,450 --> 00:00:36,890 S1: AI infrastructure we actually need and how far along we 10 00:00:36,890 --> 00:00:40,129 S1: are along that path. And I want to say something 11 00:00:40,130 --> 00:00:42,500 S1: about predictions real quick, which I've talked about before in 12 00:00:42,500 --> 00:00:45,290 S1: other videos, but I don't think it's possible to predict 13 00:00:45,320 --> 00:00:49,670 S1: tech really, the way that tech happens, who wins, who loses, 14 00:00:49,700 --> 00:00:54,080 S1: like the timelines? Like this stuff is like impossible to predict, basically. 15 00:00:54,080 --> 00:00:57,260 S1: So I wouldn't want you to think that I would 16 00:00:57,260 --> 00:01:00,510 S1: try to do that because I think that's foolish. What 17 00:01:00,510 --> 00:01:04,110 S1: I believe we can predict related to tech are the 18 00:01:04,110 --> 00:01:07,830 S1: human things that are related to human desires. So the 19 00:01:07,830 --> 00:01:11,730 S1: question is not whether we can predict tech itself, but 20 00:01:11,730 --> 00:01:14,820 S1: can we predict what we want from tech. And I 21 00:01:14,819 --> 00:01:18,089 S1: think that is a powerful way to kind of see 22 00:01:18,090 --> 00:01:21,300 S1: what might be happening and what could unfold in the 23 00:01:21,300 --> 00:01:24,180 S1: next 1 to 3 years or whatever. So I want 24 00:01:24,209 --> 00:01:27,300 S1: to specifically look at like, how far along are we 25 00:01:27,330 --> 00:01:32,100 S1: to building the infrastructure that we need for AI, right. So, um, 26 00:01:32,100 --> 00:01:35,699 S1: I own a bunch of, uh, Nvidia, um, I think 27 00:01:35,700 --> 00:01:39,600 S1: I've got some TSMC like, um, I've got stock related 28 00:01:39,600 --> 00:01:42,179 S1: to AI. I'm all in on AI, have been since 29 00:01:42,180 --> 00:01:47,280 S1: late 2022. Uh, I went independent to work in this field, um, 30 00:01:47,280 --> 00:01:52,500 S1: after being in security for like 25 years. So I'm very, 31 00:01:52,530 --> 00:01:55,380 S1: I guess, religious about the whole thing. So I want 32 00:01:55,380 --> 00:01:57,150 S1: you to understand that which you probably already know if 33 00:01:57,180 --> 00:01:59,860 S1: you've seen any of my videos? Um, but I don't 34 00:01:59,860 --> 00:02:03,220 S1: think that affects this particular thing that we're talking about. 35 00:02:03,250 --> 00:02:05,590 S1: I'm sure it does in some sense, because that's the 36 00:02:05,590 --> 00:02:08,500 S1: way bias works. But I think this argument I'm about 37 00:02:08,500 --> 00:02:12,970 S1: to make stands independent of that. So essentially what I'm 38 00:02:12,970 --> 00:02:17,140 S1: looking to figure out is how much infrastructure do we need? 39 00:02:17,169 --> 00:02:19,540 S1: How many GPUs do we need? How many data centers 40 00:02:19,540 --> 00:02:22,660 S1: do we need? How many AI companies and startups do 41 00:02:22,660 --> 00:02:24,790 S1: we need? And some of that depends on the actual 42 00:02:24,790 --> 00:02:28,690 S1: technical implementations, which we can't predict. So we don't need 43 00:02:28,690 --> 00:02:30,580 S1: to go too deep into that. But I'm just trying 44 00:02:30,580 --> 00:02:34,420 S1: to figure out, are we at like 13% of how 45 00:02:34,419 --> 00:02:37,120 S1: much I. We need a lot of people think we're 46 00:02:37,150 --> 00:02:41,410 S1: at 85%. For example, they're like, yeah, that was pretty 47 00:02:41,410 --> 00:02:45,519 S1: much it in 2023 and 2024. And now it's mostly 48 00:02:45,520 --> 00:02:47,860 S1: hype and it's just going to die down. So they 49 00:02:47,860 --> 00:02:51,070 S1: think it's like 87% and it's pretty much done. It 50 00:02:51,070 --> 00:02:54,730 S1: was a cycle. Whatever other people are like no we're 51 00:02:54,730 --> 00:02:57,400 S1: just getting started. We're at like 3%, but it's going 52 00:02:57,430 --> 00:02:59,950 S1: to grow over the next five years or whatever. And 53 00:02:59,950 --> 00:03:03,100 S1: I guess I'll just spoil what I think the answer is. 54 00:03:03,100 --> 00:03:07,780 S1: I think the answer is we're at like 0.000. I 55 00:03:07,780 --> 00:03:12,070 S1: don't know how many zeros, like eight zeros and a 1% 56 00:03:12,070 --> 00:03:16,120 S1: of where we need to be or where's more specifically 57 00:03:16,120 --> 00:03:21,369 S1: where humans will demand that we get. Okay. So that's 58 00:03:21,370 --> 00:03:23,020 S1: that's the point I want to make here. And the 59 00:03:23,020 --> 00:03:25,419 S1: way I want to get there is by asking a question, 60 00:03:25,419 --> 00:03:28,660 S1: what are we actually trying to do with AI? What 61 00:03:28,690 --> 00:03:30,610 S1: are we trying to do with AI? So I want 62 00:03:30,639 --> 00:03:34,300 S1: to walk through a few ideas there. So one cool 63 00:03:34,300 --> 00:03:38,530 S1: idea here is learning from the past and anticipating the future. 64 00:03:38,530 --> 00:03:40,690 S1: So we are sitting in the middle here in the 65 00:03:40,690 --> 00:03:44,140 S1: current present moment, and we could use AI to look 66 00:03:44,140 --> 00:03:47,140 S1: into the past, find things that went wrong for the 67 00:03:47,140 --> 00:03:50,620 S1: purpose of adjusting our current behavior right on the other side. 68 00:03:50,620 --> 00:03:55,600 S1: We could anticipate, try to anticipate Paint based on everything 69 00:03:55,600 --> 00:03:59,140 S1: we currently know. Try to anticipate what's going to happen 70 00:03:59,140 --> 00:04:03,070 S1: in the future. For what reason? Same exact reason. Adjust 71 00:04:03,070 --> 00:04:06,100 S1: our current behavior. So I think this is a paradigm. 72 00:04:06,100 --> 00:04:09,880 S1: This is an idea or a concept that's extremely powerful. 73 00:04:09,880 --> 00:04:12,610 S1: And so the question is how much more of this 74 00:04:12,640 --> 00:04:15,190 S1: do we need? How much more of this do we need? 75 00:04:15,190 --> 00:04:17,919 S1: Or how bad of a job are we currently doing 76 00:04:17,920 --> 00:04:20,410 S1: at this? And how much more can we gain if 77 00:04:20,440 --> 00:04:23,140 S1: we were doing it much, much better? So I won't 78 00:04:23,140 --> 00:04:26,650 S1: even give my answer, but probably kind of rhetorical. All right, 79 00:04:26,680 --> 00:04:29,890 S1: next idea I would say that for any given situation, 80 00:04:29,890 --> 00:04:33,220 S1: for any given operation or thing you're trying to do, 81 00:04:33,250 --> 00:04:36,790 S1: there is a current state and there is a desired state, 82 00:04:36,790 --> 00:04:39,429 S1: and there is a delta between those two. And I 83 00:04:39,430 --> 00:04:42,580 S1: think this is a really powerful concept. It automatically asks 84 00:04:42,610 --> 00:04:46,330 S1: a question. It demands a question how do we get 85 00:04:46,330 --> 00:04:50,920 S1: from the current state to the desired state? So I 86 00:04:50,920 --> 00:04:55,550 S1: ask you again, how many situations do we have worldwide, 87 00:04:55,580 --> 00:05:00,230 S1: human civilization wide, where we have a current situation that 88 00:05:00,230 --> 00:05:03,500 S1: we're not really happy with, and we either need help 89 00:05:03,500 --> 00:05:05,779 S1: coming up with a desired state, or we already have 90 00:05:05,779 --> 00:05:09,440 S1: a desired state clearly in mind. Or maybe we have 91 00:05:09,470 --> 00:05:12,409 S1: thoughts about a desired state, but we could use some 92 00:05:12,410 --> 00:05:15,200 S1: help articulating it. But either way, we want to get 93 00:05:15,200 --> 00:05:18,110 S1: to this desired state, which I could potentially help us 94 00:05:18,110 --> 00:05:21,260 S1: articulate the desired state. But either way, how are we 95 00:05:21,290 --> 00:05:26,990 S1: doing on understanding our current state perfectly, articulating our desired state, 96 00:05:26,990 --> 00:05:38,390 S1: and figuring out the delta between those two? Think education. Poverty. Climate. Health. Aging. Politics. War. 97 00:05:38,420 --> 00:05:42,200 S1: Just just think all the whole scope of human problems 98 00:05:42,200 --> 00:05:45,230 S1: over the course of history. How are we doing on this? 99 00:05:45,230 --> 00:05:49,220 S1: How much potential is there to do this better? That's 100 00:05:49,220 --> 00:05:51,650 S1: the second one. And the second one really brings up 101 00:05:51,650 --> 00:05:55,520 S1: this third one. What are the actions that we should 102 00:05:55,520 --> 00:05:59,870 S1: take to make the transition? This this is incredibly powerful. 103 00:05:59,900 --> 00:06:02,900 S1: Can we use AI to understand the current state? Can 104 00:06:02,930 --> 00:06:07,820 S1: we use AI to articulate and structure and build and 105 00:06:07,820 --> 00:06:12,500 S1: communicate the desired state? And most importantly, what's the plan? 106 00:06:12,500 --> 00:06:15,440 S1: How do we transition from one to the other? How 107 00:06:15,440 --> 00:06:18,380 S1: much do we need this middle piece? How good of 108 00:06:18,380 --> 00:06:22,190 S1: a job are we as humans doing at executing on 109 00:06:22,190 --> 00:06:25,490 S1: this middle piece? I would argue not very well at all. 110 00:06:25,490 --> 00:06:28,489 S1: And we haven't been for centuries. Well, I mean, you 111 00:06:28,490 --> 00:06:32,000 S1: could argue it could go way worse. Totally. Absolutely could 112 00:06:32,000 --> 00:06:35,690 S1: go way worse. So sure, we're not all dead yet, 113 00:06:35,690 --> 00:06:38,630 S1: so that's good. But if you look at the course 114 00:06:38,630 --> 00:06:41,690 S1: of history, it's just like folly after folly. And we're 115 00:06:41,690 --> 00:06:44,690 S1: currently in a whole bunch of it right now. And 116 00:06:44,690 --> 00:06:47,029 S1: it's a bad situation to be in. So I would 117 00:06:47,029 --> 00:06:49,310 S1: argue that this yellow piece right here in the middle 118 00:06:49,320 --> 00:06:55,020 S1: is extraordinary. It is. It has extraordinary potential for what 119 00:06:55,020 --> 00:06:57,900 S1: we wish we could do with that yellow to transition 120 00:06:57,900 --> 00:07:02,880 S1: to this desired state. So let's break these three down 121 00:07:02,880 --> 00:07:05,640 S1: even more. When you look at the current state, you 122 00:07:05,640 --> 00:07:08,850 S1: talk about, okay, let's understand the current state. How granular 123 00:07:08,880 --> 00:07:12,390 S1: are we talking about? Right. Because you could go from 124 00:07:12,420 --> 00:07:15,720 S1: like a coffee cup with like an amount of hot 125 00:07:15,720 --> 00:07:18,000 S1: liquid inside of it. And we could ask the question, 126 00:07:18,000 --> 00:07:20,640 S1: what is the current state of this liquid inside this 127 00:07:20,640 --> 00:07:23,070 S1: coffee cup? What's the current state of the coffee? Well, 128 00:07:23,070 --> 00:07:24,960 S1: you could ask lots of different people about this. You 129 00:07:24,960 --> 00:07:27,750 S1: could ask a philosopher, you could ask whatever. It's like, oh, 130 00:07:27,780 --> 00:07:31,800 S1: half full, half empty. Um, you could ask like a 131 00:07:31,830 --> 00:07:35,520 S1: chemist or a physicist. And it's like, well, um, what 132 00:07:35,520 --> 00:07:37,290 S1: do you want to know? Like how many atoms, like, 133 00:07:37,320 --> 00:07:40,410 S1: how excited are they? What the the location of all 134 00:07:40,410 --> 00:07:44,190 S1: the electrons. Certain parts of this question for a given 135 00:07:44,220 --> 00:07:47,880 S1: object are just not knowable. Like, what is the state 136 00:07:47,880 --> 00:07:53,160 S1: of all atoms and subatomic particles in, say, a pen 137 00:07:53,160 --> 00:07:56,730 S1: or in a cup of coffee or the state of 138 00:07:56,730 --> 00:08:00,720 S1: a human, for example, you say, how is he doing? 139 00:08:00,720 --> 00:08:03,840 S1: How is she doing? Well, what do you mean? I 140 00:08:03,840 --> 00:08:05,940 S1: can't give you a full rundown of every atom in 141 00:08:05,940 --> 00:08:09,360 S1: their body. So the question is, what level of depth 142 00:08:09,390 --> 00:08:12,150 S1: do we need to be able to answer that question? 143 00:08:12,150 --> 00:08:15,660 S1: And that comes down to how much telemetry can we get, 144 00:08:15,690 --> 00:08:18,809 S1: how much data can we get coming off of this situation. 145 00:08:18,810 --> 00:08:21,630 S1: So that's I've got an aura ring on, I've got 146 00:08:21,630 --> 00:08:24,600 S1: an Apple Watch on. This is telemetry coming off of 147 00:08:24,600 --> 00:08:28,980 S1: me so that you can ask an AI very soon 148 00:08:29,010 --> 00:08:33,090 S1: or actually now in my opinion, Apple is building life 149 00:08:33,090 --> 00:08:36,959 S1: OS essentially, in case you guys didn't know that. So 150 00:08:36,960 --> 00:08:40,860 S1: you will be able to ask yourself, how am I doing? Um, 151 00:08:40,860 --> 00:08:43,290 S1: and it will know. Okay, you're talking about health. You're 152 00:08:43,290 --> 00:08:45,750 S1: talking about financially, you're talking about education. What are you 153 00:08:45,750 --> 00:08:48,730 S1: talking about? But you have to have that telemetry coming in, right? 154 00:08:48,760 --> 00:08:51,910 S1: And you have to know what the limits are right for. 155 00:08:51,940 --> 00:08:54,640 S1: For a given object type, you have to know certain 156 00:08:54,640 --> 00:08:59,890 S1: questions and certain types of metrics that are just attainable. Okay. 157 00:08:59,890 --> 00:09:04,360 S1: Can we get someone's current heart rate? No, not 15 158 00:09:04,360 --> 00:09:07,840 S1: years ago. Not automatically, not every moment of the day. 159 00:09:07,840 --> 00:09:10,660 S1: We couldn't do that before. We can now. So that's 160 00:09:10,660 --> 00:09:15,910 S1: a technology change that enabled that level of metric to happen. Well, 161 00:09:15,940 --> 00:09:19,240 S1: Apple is currently gathering things like mood as well. How 162 00:09:19,270 --> 00:09:22,300 S1: are you feeling today. And it's associating those things with 163 00:09:22,300 --> 00:09:25,840 S1: other things like how much you've exercised and stuff like that. Right. 164 00:09:25,840 --> 00:09:28,060 S1: So you kind of see where this is all going 165 00:09:28,059 --> 00:09:32,410 S1: with life. OS now that's for a person. Okay. Now 166 00:09:32,410 --> 00:09:36,010 S1: let's talk about a family. What's the state of the family? Well, 167 00:09:36,010 --> 00:09:38,559 S1: what does that mean? Do we need all of those 168 00:09:38,559 --> 00:09:42,160 S1: metrics for each individual person? Sure. But you also want 169 00:09:42,190 --> 00:09:44,800 S1: to know the dynamics between the people. You also want 170 00:09:44,830 --> 00:09:48,080 S1: to know how the family as a whole is doing. Um, 171 00:09:48,080 --> 00:09:51,949 S1: what's the quality of life? What is the upward trajectory 172 00:09:51,950 --> 00:09:55,190 S1: look like? Right? This is the state of a family. 173 00:09:55,280 --> 00:09:58,880 S1: Now let's talk about business companies. What is the state 174 00:09:58,880 --> 00:10:00,800 S1: of a company? Well, what are we talking about? Its 175 00:10:00,830 --> 00:10:06,500 S1: IT infrastructure with thousands of Kubernetes pods running. What's the 176 00:10:06,500 --> 00:10:09,230 S1: current state of a Kubernetes pod right now? And also 177 00:10:09,230 --> 00:10:12,020 S1: right now and also right now, how often are we 178 00:10:12,020 --> 00:10:15,620 S1: polling these things both for a human and for a company? Right. 179 00:10:15,650 --> 00:10:18,050 S1: And a lot of these questions are going to come 180 00:10:18,050 --> 00:10:21,260 S1: down to, sure, we could poll the current state of 181 00:10:21,260 --> 00:10:26,060 S1: every single Kubernetes pod at, say, I don't know, Google, right? 182 00:10:26,059 --> 00:10:29,480 S1: Every single second, but that might cost billions of dollars 183 00:10:29,480 --> 00:10:32,420 S1: because it takes, you know, it's just expensive. And where 184 00:10:32,420 --> 00:10:34,579 S1: would you store the stuff in all these different questions. 185 00:10:34,580 --> 00:10:37,670 S1: So you've got like how granular, what is the update 186 00:10:37,670 --> 00:10:40,670 S1: frequency and how expensive is that to gather based on 187 00:10:40,670 --> 00:10:45,410 S1: the current tech. Right. And this is Tremendously interesting and 188 00:10:45,410 --> 00:10:49,219 S1: tremendously powerful, especially for something like a human. The most 189 00:10:49,220 --> 00:10:52,640 S1: important driver for all of this AI stuff one is 190 00:10:52,640 --> 00:10:55,400 S1: going to be business, no question. Right? And that's going 191 00:10:55,429 --> 00:10:58,370 S1: to have the money behind it or whatever. But ultimately, 192 00:10:58,370 --> 00:11:01,820 S1: the bigger thing is going to be humans using this 193 00:11:01,820 --> 00:11:06,530 S1: to feel better and be better and improve themselves. That's 194 00:11:06,530 --> 00:11:09,949 S1: that's the entire game here. What are people struggling with 195 00:11:09,950 --> 00:11:14,870 S1: right now? Job happiness. Can't get a job. Loneliness. A 196 00:11:14,870 --> 00:11:18,380 S1: lack of meaning in their lives. Okay. How do you 197 00:11:18,410 --> 00:11:20,990 S1: solve how how do you solve the current state desired 198 00:11:20,990 --> 00:11:24,679 S1: state situation for somebody who's lonely? And I got another 199 00:11:24,679 --> 00:11:27,679 S1: example of that coming up. But think about that. Think 200 00:11:27,679 --> 00:11:31,820 S1: about what it takes to sort of solve these problems. 201 00:11:31,850 --> 00:11:35,839 S1: Go from current state, desired state for a human, for 202 00:11:35,840 --> 00:11:39,650 S1: a company, and what level of context you need to 203 00:11:39,679 --> 00:11:42,270 S1: be able to answer those sorts of questions. now, I 204 00:11:42,270 --> 00:11:46,050 S1: would argue just like the previous ones, we are nowhere 205 00:11:46,080 --> 00:11:51,809 S1: near gathering enough context on the current state of anything. 206 00:11:51,840 --> 00:11:56,880 S1: Of anything. Okay, a park bench, a tree and at 207 00:11:56,880 --> 00:12:00,210 S1: the park, the state of a human that you care about. 208 00:12:00,240 --> 00:12:03,360 S1: We don't have nearly enough telemetry on that. We want 209 00:12:03,390 --> 00:12:05,520 S1: to change a business. We have no idea what's going 210 00:12:05,520 --> 00:12:08,280 S1: on in business. Most people who work there and actually 211 00:12:08,280 --> 00:12:10,260 S1: run the business, they have no idea what's going on 212 00:12:10,260 --> 00:12:12,540 S1: in the business. They don't know the level of mood 213 00:12:12,540 --> 00:12:16,679 S1: and happiness and the state. How much waste is happening 214 00:12:16,679 --> 00:12:20,939 S1: at any given time? Bottom line here is things like 215 00:12:20,940 --> 00:12:26,790 S1: context size, things like Rag technologies like these are absolutely 216 00:12:26,790 --> 00:12:30,720 S1: in their infancy because they have to be because of 217 00:12:30,720 --> 00:12:34,140 S1: the size of the scope of the problem that they 218 00:12:34,140 --> 00:12:38,190 S1: actually are being signed up to solve. Okay. Uh, Altman 219 00:12:38,190 --> 00:12:40,680 S1: talked about this a while back. He was like, look, 220 00:12:40,710 --> 00:12:44,280 S1: at some point, context is basically going to be infinite, 221 00:12:44,280 --> 00:12:49,620 S1: and obviously nothing's infinite. So that doesn't really mean infinite. 222 00:12:49,620 --> 00:12:52,860 S1: It means functionally infinite. So there's going to be this 223 00:12:52,860 --> 00:12:57,990 S1: competition between context size going into models versus Rag, going 224 00:12:58,020 --> 00:13:02,640 S1: into models versus whatever, whatever comes out later and makes 225 00:13:02,640 --> 00:13:07,199 S1: maybe those not as important. But either way, what you 226 00:13:07,200 --> 00:13:09,900 S1: have to do is you have to get this current 227 00:13:09,900 --> 00:13:13,530 S1: state for a thing that matters to humans. Forget tech 228 00:13:13,559 --> 00:13:16,890 S1: is about a thing that matters to humans into the 229 00:13:16,890 --> 00:13:20,340 S1: brain of the AI, so that all it's understanding of 230 00:13:20,340 --> 00:13:23,910 S1: the world can be applied to this larger situation. Now 231 00:13:23,940 --> 00:13:26,640 S1: think about this as a human. Imagine somebody who's 40 232 00:13:26,640 --> 00:13:30,270 S1: years old and has had this amazing life and, you know, 233 00:13:30,300 --> 00:13:34,620 S1: the hardships and people died and they fought in wars 234 00:13:34,620 --> 00:13:37,770 S1: and they traveled to Costa Rica and they did ayahuasca, 235 00:13:37,770 --> 00:13:40,089 S1: and they learned all these things. They know thousands of 236 00:13:40,090 --> 00:13:43,570 S1: people and they've had this wonderful life, and they're trying 237 00:13:43,570 --> 00:13:47,050 S1: to craft even a better life. And you want to 238 00:13:47,080 --> 00:13:49,630 S1: ask an AI, you know, tell me what I can 239 00:13:49,630 --> 00:13:52,780 S1: work on. Tell me about myself. What do you suggest 240 00:13:52,780 --> 00:13:56,410 S1: for me? I'm looking at doing this career or this career. 241 00:13:56,410 --> 00:13:58,870 S1: I'm looking at moving here. I'm looking at marrying this 242 00:13:58,870 --> 00:14:02,260 S1: girl or not marrying the squirrel. Um. What should I do? 243 00:14:02,290 --> 00:14:05,920 S1: What does the. I need to know about that person 244 00:14:05,920 --> 00:14:10,120 S1: to make that decision? Or to offer advice in making 245 00:14:10,120 --> 00:14:14,950 S1: that decision? Ideally, it has everything. It's got the genome. 246 00:14:14,950 --> 00:14:18,760 S1: It's it's got all your medical records growing up. Ideally 247 00:14:18,760 --> 00:14:23,200 S1: it has journal entries. Ideally, it's been recording you at 248 00:14:23,200 --> 00:14:26,530 S1: a deep level of state gathering for your entire life. 249 00:14:26,530 --> 00:14:29,229 S1: So imagine this is 200 years from now and this 250 00:14:29,230 --> 00:14:31,660 S1: person is 40 years old. But every moment of their 251 00:14:31,660 --> 00:14:34,780 S1: life has been captured. And the very moment that you 252 00:14:34,780 --> 00:14:38,230 S1: ask a question, what should I do? Which career? Which 253 00:14:38,230 --> 00:14:41,800 S1: country do I move to? Which guy do I marry? Um, 254 00:14:41,800 --> 00:14:46,000 S1: the entirety of their life up to that moment would 255 00:14:46,000 --> 00:14:49,000 S1: be the context for answering that question. This is the 256 00:14:49,000 --> 00:14:51,850 S1: key point for this entire thing. Think of the AI 257 00:14:51,850 --> 00:14:57,220 S1: infrastructure that is required to have the entire context of 258 00:14:57,220 --> 00:15:00,070 S1: the moment, right up to the point when the question 259 00:15:00,070 --> 00:15:05,290 S1: is asked to be jammed into that AI's mind. But 260 00:15:05,290 --> 00:15:07,930 S1: don't just imagine it for this 40 year old guy 261 00:15:07,930 --> 00:15:11,950 S1: in Costa Rica. Imagine it for a giant business with 262 00:15:11,950 --> 00:15:17,560 S1: 10,000 employees who's been in business for 120 years, how 263 00:15:17,560 --> 00:15:20,560 S1: much context, at a deep level, can we grab from 264 00:15:20,560 --> 00:15:24,160 S1: that entire 120 years and put it into what should 265 00:15:24,160 --> 00:15:28,390 S1: I do now that is big. That is massive. And 266 00:15:28,390 --> 00:15:31,300 S1: here's the crazy part. It can be summarized to a 267 00:15:31,300 --> 00:15:34,030 S1: ten page text document. It can be summarized to a 268 00:15:34,030 --> 00:15:38,510 S1: 1010 page text document. And guess what? It would be 269 00:15:38,510 --> 00:15:42,380 S1: good if you could summarize it to a 200 page report. 270 00:15:42,380 --> 00:15:45,710 S1: It would be really good. It could also be terabytes 271 00:15:45,710 --> 00:15:48,710 S1: of data. Petabytes of data. How quickly can you get 272 00:15:48,710 --> 00:15:52,790 S1: petabytes of data into an AI's brain to snap and 273 00:15:52,790 --> 00:15:56,450 S1: answer the question in a really, really powerful way for 274 00:15:56,450 --> 00:15:59,480 S1: a business, for a country, for a city, for a human, 275 00:15:59,480 --> 00:16:03,290 S1: for a family. Think of the infrastructure that is required 276 00:16:03,290 --> 00:16:06,170 S1: to do that, and think about the fact that this 277 00:16:06,170 --> 00:16:09,890 S1: is what humans want. This is what humans will demand. 278 00:16:09,920 --> 00:16:13,520 S1: This is why I'm saying forget the tech. Forget the 279 00:16:13,520 --> 00:16:17,240 S1: tech does not matter. What matters is what humans want. 280 00:16:17,270 --> 00:16:23,630 S1: Humans want the most amazing answer ever to that question. 281 00:16:23,660 --> 00:16:26,480 S1: Do I do this merger with this company? Do I 282 00:16:26,510 --> 00:16:30,470 S1: hire this person as my CFO? And the amount of 283 00:16:30,470 --> 00:16:33,800 S1: data that is needed to get a better and better 284 00:16:33,800 --> 00:16:37,670 S1: and better answer to that question just keeps going up 285 00:16:37,670 --> 00:16:41,630 S1: along with the capabilities of the tech. Right. So this 286 00:16:41,630 --> 00:16:45,170 S1: is what's so powerful about this paradigm of thinking about 287 00:16:45,170 --> 00:16:49,370 S1: this is that there's kind of no end. There's no 288 00:16:49,370 --> 00:16:53,330 S1: foreseeable end coming to how much better it gets when 289 00:16:53,330 --> 00:16:57,620 S1: you give it more. And that's why okay, 200 K 290 00:16:57,650 --> 00:17:01,489 S1: tokens right now. So that's a lot. No it's not 291 00:17:01,520 --> 00:17:03,950 S1: it's not a lot. It's not a lot because watch this. 292 00:17:03,950 --> 00:17:06,800 S1: The thing a human wants to do is to say hey, 293 00:17:06,800 --> 00:17:11,120 S1: I've got an idea for a concept for this really cool, um, 294 00:17:11,150 --> 00:17:14,750 S1: mini series that's going to be on Netflix. It's got anime, 295 00:17:14,780 --> 00:17:18,859 S1: it's got Harry Potter stuff, but it's in this stylized thing, 296 00:17:18,890 --> 00:17:22,399 S1: kind of like The Last Samurai or, um, or what 297 00:17:22,400 --> 00:17:25,850 S1: was that? It was called Shogun, the recent animated thing. Anyway, 298 00:17:25,850 --> 00:17:27,470 S1: you give it all these things, you're like, I want 299 00:17:27,470 --> 00:17:30,050 S1: this art. I want the concept of Star Wars, but 300 00:17:30,050 --> 00:17:32,520 S1: it's got to be Harry Potter. But I also want 301 00:17:32,550 --> 00:17:35,640 S1: like deep romance. But I want it to be really, 302 00:17:35,640 --> 00:17:38,460 S1: really gritty. And it definitely, you know, got to be 303 00:17:38,460 --> 00:17:40,770 S1: over 21 to even watch this thing. And there's also 304 00:17:40,770 --> 00:17:43,080 S1: violence and there's also sex and there's also all these 305 00:17:43,080 --> 00:17:45,840 S1: different things. And you basically hit it with that and 306 00:17:45,840 --> 00:17:47,850 S1: you're like, yeah. But I also want it to be 307 00:17:47,850 --> 00:17:51,090 S1: kind of like Tolstoy. I really like Tolstoy. And it's 308 00:17:51,090 --> 00:17:54,900 S1: going to go and write you and build you a 309 00:17:54,900 --> 00:17:59,700 S1: complete movie, a complete series, and do all these different 310 00:17:59,700 --> 00:18:01,890 S1: things with agents, all the stuff that's coming out. That's 311 00:18:01,890 --> 00:18:04,770 S1: kind of inevitable, right? How much does it need to 312 00:18:04,800 --> 00:18:07,050 S1: know to be able to do that perfectly? It's got 313 00:18:07,050 --> 00:18:09,390 S1: to be able to read into everything you're saying. It's 314 00:18:09,390 --> 00:18:12,629 S1: got to have all the capabilities of actually doing the 315 00:18:12,630 --> 00:18:15,270 S1: video and the art creation and all this stuff, but 316 00:18:15,270 --> 00:18:18,030 S1: the more it knows, the better this thing gets. And 317 00:18:18,030 --> 00:18:21,540 S1: that doesn't really stop at any time soon. Right? So 318 00:18:21,540 --> 00:18:24,360 S1: this is needed for business. This is needed for creativity. 319 00:18:24,359 --> 00:18:28,050 S1: This is needed for human thriving at an individual and 320 00:18:28,050 --> 00:18:31,740 S1: a society level. There is no end to how much 321 00:18:31,740 --> 00:18:35,040 S1: we will demand this thing. I'll give you another example here. 322 00:18:35,430 --> 00:18:39,810 S1: Because of this assassination that happened with the head of United, 323 00:18:40,020 --> 00:18:43,889 S1: UnitedHealth and I talked about this in in my book, 324 00:18:43,890 --> 00:18:47,310 S1: The Real Internet of Things in like 20, 2016. It 325 00:18:47,310 --> 00:18:50,070 S1: came out and I've recently done a whole series on 326 00:18:50,100 --> 00:18:53,220 S1: like modernizing it. And it's got video and everything, so 327 00:18:53,220 --> 00:18:55,409 S1: you should check it out. But the point is, what 328 00:18:55,410 --> 00:18:57,780 S1: I talked about in that book was walking down the 329 00:18:57,780 --> 00:19:01,649 S1: street and having sensors all around you. This is the 330 00:19:01,650 --> 00:19:04,410 S1: vision that I'm seeing for everything. You have sensors all 331 00:19:04,410 --> 00:19:06,960 S1: around you, the current state of the world. This thing 332 00:19:06,960 --> 00:19:09,750 S1: on the left here, this current state of the world, 333 00:19:09,750 --> 00:19:13,380 S1: is the thing that your eye is always monitoring for you. 334 00:19:13,410 --> 00:19:17,040 S1: Your Da, your digital assistant. Right? It could see behind you. 335 00:19:17,070 --> 00:19:19,649 S1: You could see above you. Why is it seeing above you? 336 00:19:19,680 --> 00:19:23,550 S1: This is dystopian, but this is real. Because of drones, 337 00:19:23,550 --> 00:19:26,340 S1: that's why. Why is it seeing behind you? Because someone 338 00:19:26,340 --> 00:19:29,410 S1: might walk up behind you with a big backpack on 339 00:19:29,410 --> 00:19:32,200 S1: and a mask and pull out a gun and shoot you. 340 00:19:32,200 --> 00:19:35,770 S1: So what your eye has to be doing is watching 341 00:19:35,770 --> 00:19:38,859 S1: everything all the time. And guess what? If you're a 342 00:19:38,859 --> 00:19:41,530 S1: VIP or you just have some money, you're going to 343 00:19:41,530 --> 00:19:45,850 S1: have little drones flying around looking at everything, looking down. Hey, 344 00:19:45,850 --> 00:19:48,070 S1: why is that car moving like that? Hey, moved. Move 345 00:19:48,100 --> 00:19:49,689 S1: to the left across the street. I don't want you 346 00:19:49,690 --> 00:19:51,969 S1: over here. Your little ear piece is going to be 347 00:19:51,970 --> 00:19:55,629 S1: giving you these little guidance things. It hears your stomach rumble. 348 00:19:55,630 --> 00:19:58,060 S1: It's like, hey, there's Thai food up there. It's your 349 00:19:58,060 --> 00:20:02,260 S1: favorite restaurant. I already pinged Jeremiah. He's got your favorite table. 350 00:20:02,260 --> 00:20:04,180 S1: I turned on table tennis. So when you get there, 351 00:20:04,180 --> 00:20:08,199 S1: table tennis is on. It is monitoring the state of 352 00:20:08,200 --> 00:20:12,340 S1: the world and changing the state of that world in 353 00:20:12,340 --> 00:20:14,830 S1: preemptively watching for the state of the world to make 354 00:20:14,859 --> 00:20:18,310 S1: sure it doesn't turn into an undesired state for you 355 00:20:18,340 --> 00:20:21,910 S1: 24 over seven like every second. Boom, boom, boom boom. 356 00:20:21,910 --> 00:20:25,629 S1: It's reading every API. It's looking at every person. It's 357 00:20:25,630 --> 00:20:28,220 S1: pulling up their daemon to see if there if there's 358 00:20:28,250 --> 00:20:31,520 S1: information on them, like should we go introduce ourselves to them? 359 00:20:31,520 --> 00:20:34,729 S1: This is the parsing of the current state of the 360 00:20:34,730 --> 00:20:38,840 S1: world around us at all times. That once again only 361 00:20:38,840 --> 00:20:41,990 S1: gets better and better with more tech, right? You could 362 00:20:41,990 --> 00:20:47,060 S1: do this a little bit in 2025 and 2026 and 27. 363 00:20:47,060 --> 00:20:49,310 S1: That'll be pretty good. But wait till you see it 364 00:20:49,310 --> 00:20:51,170 S1: in five years. Wait till you see it. In ten 365 00:20:51,170 --> 00:20:54,440 S1: years it will be absolute sci fi stuff. And again, 366 00:20:54,440 --> 00:20:57,590 S1: you don't need to think about the tech. You only 367 00:20:57,590 --> 00:21:00,889 S1: need to think about the fact that humans wish they 368 00:21:00,890 --> 00:21:04,939 S1: could see around themselves at all times. They wish they 369 00:21:04,940 --> 00:21:08,990 S1: had that tech on all of their kids at all times. 370 00:21:08,990 --> 00:21:13,760 S1: So when some white kid jumps out of a jumps 371 00:21:13,760 --> 00:21:16,160 S1: out of a pickup in the front of the school 372 00:21:16,160 --> 00:21:18,830 S1: and runs into the thing, and he looks like he's 373 00:21:18,830 --> 00:21:22,310 S1: over 18, but probably 19 or 20, and he's got 374 00:21:22,310 --> 00:21:26,550 S1: a backpack. Suddenly the earpiece in your kid's ear and 375 00:21:26,550 --> 00:21:29,129 S1: in the ear of the principal and everyone else. Boom! 376 00:21:29,160 --> 00:21:32,879 S1: There's data happening. Why? Because there's drones flying over. This 377 00:21:32,880 --> 00:21:36,840 S1: is a this is a monitored state of the world 378 00:21:36,840 --> 00:21:40,440 S1: that I parse and hand to the people who care 379 00:21:40,440 --> 00:21:43,470 S1: about that state. And going back to the first point, 380 00:21:43,470 --> 00:21:46,830 S1: there is so much you can gather every everything's going 381 00:21:46,859 --> 00:21:48,419 S1: to have an API. You're going to be able to 382 00:21:48,450 --> 00:21:51,570 S1: pull all this data. But okay, why am I pulling 383 00:21:51,570 --> 00:21:54,300 S1: gigabytes of data all the time? Can I even process it? 384 00:21:54,330 --> 00:21:57,000 S1: The answer is no, you can't, which means we won't 385 00:21:57,000 --> 00:21:59,070 S1: pull that yet. But that will push us to be 386 00:21:59,100 --> 00:22:01,350 S1: able to process it. Oh, now we can. Now we'll 387 00:22:01,350 --> 00:22:04,650 S1: pull more data. This will just keep ratcheting up and 388 00:22:04,650 --> 00:22:09,540 S1: keep ratcheting up. And until until you eventually have something 389 00:22:09,540 --> 00:22:13,740 S1: like Neuralink and you just kind of know the state 390 00:22:13,740 --> 00:22:16,439 S1: the same way you see color. You will know the 391 00:22:16,440 --> 00:22:20,670 S1: state of danger. Right? And I believe all this is inevitable, 392 00:22:20,670 --> 00:22:22,710 S1: but it's so far in the future. Like you don't 393 00:22:22,710 --> 00:22:24,810 S1: need to worry about it. But what I'm trying to 394 00:22:24,810 --> 00:22:28,500 S1: do is show you that it's obvious that that sort 395 00:22:28,500 --> 00:22:32,940 S1: of push is happening from humans, because that's what we demand, right? 396 00:22:32,940 --> 00:22:35,970 S1: So that's kind of where we're going, um, and how 397 00:22:35,970 --> 00:22:38,250 S1: it will happen and who will do it or whatever. 398 00:22:38,250 --> 00:22:41,760 S1: Who knows? Nobody knows. Could be a giant single corporation 399 00:22:41,760 --> 00:22:43,800 S1: that does everything. And or it could be a million 400 00:22:43,800 --> 00:22:47,580 S1: different startups. Who knows? It's impossible to predict. The point is, 401 00:22:47,580 --> 00:22:50,340 S1: we will demand this and we will demand it gets 402 00:22:50,340 --> 00:22:53,520 S1: better and better. And the deeper you go on how 403 00:22:53,520 --> 00:22:57,990 S1: granular and the more often the updates to the frequency 404 00:22:57,990 --> 00:23:00,960 S1: are demanded, and the more the cost goes down, the 405 00:23:00,960 --> 00:23:06,390 S1: more we will have. It's this thing just goes crazy. Okay, 406 00:23:06,480 --> 00:23:10,020 S1: now we talk about what's the desired state. Now we 407 00:23:10,020 --> 00:23:12,180 S1: talk about how do we get from here to there? 408 00:23:12,210 --> 00:23:15,840 S1: What's the planning? Right? How do we make a I'm 409 00:23:15,840 --> 00:23:17,790 S1: reading good to great right now. How do we make 410 00:23:17,790 --> 00:23:20,580 S1: this business from good to great. What are the changes 411 00:23:20,580 --> 00:23:23,790 S1: we make. Right. That's also an AI eye piece. All 412 00:23:23,790 --> 00:23:26,939 S1: of these, all of these combined are going to push 413 00:23:26,940 --> 00:23:31,050 S1: this thing from us barely starting this year, last year, 414 00:23:31,080 --> 00:23:35,160 S1: year before to it's not a hockey stick. It is 415 00:23:35,160 --> 00:23:38,639 S1: straight up into the sky in how much context we 416 00:23:38,640 --> 00:23:41,670 S1: need and how much memory we need and how much 417 00:23:41,670 --> 00:23:46,500 S1: processing we need right now. It could be that new 418 00:23:46,500 --> 00:23:50,699 S1: processors come out so we could do a 100,000 times 419 00:23:50,700 --> 00:23:56,699 S1: as much processing, with 100,000 times less effort and resources 420 00:23:56,700 --> 00:23:59,159 S1: and stuff like that. Doesn't matter. We'll just ask for 421 00:23:59,160 --> 00:24:02,940 S1: 100,000 times more. We'll be like, oh, cool. So so 422 00:24:02,940 --> 00:24:07,950 S1: you can pull, um, 40,000 metrics on me as a 423 00:24:07,950 --> 00:24:11,580 S1: human every second. Cool. Can you get the state of 424 00:24:11,580 --> 00:24:14,400 S1: every molecule in my body? And they'll be like, no, 425 00:24:14,400 --> 00:24:17,219 S1: that's impossible. Cool. Someone will go work on it. And 426 00:24:17,220 --> 00:24:20,609 S1: suddenly the current hardware is not good enough. And we 427 00:24:20,609 --> 00:24:22,810 S1: could make better predictions if we knew the state of 428 00:24:22,810 --> 00:24:25,630 S1: every molecule in the body. And guess what? It just 429 00:24:25,630 --> 00:24:28,750 S1: keeps going. It just keeps going. That's the point of this. 430 00:24:28,750 --> 00:24:32,980 S1: So these are the three basic pieces current state, desired state, 431 00:24:32,980 --> 00:24:38,830 S1: and the actions, processes and SOPs. SOP is a standard 432 00:24:38,830 --> 00:24:42,850 S1: operating procedure. I think this is a really, really powerful concept. 433 00:24:42,850 --> 00:24:46,240 S1: I think these three combined are going to be like 434 00:24:46,240 --> 00:24:49,840 S1: a framework for making improvements to things. I did a 435 00:24:49,840 --> 00:24:55,209 S1: piece in early 2023, like March of 2023, called Pspca 436 00:24:55,240 --> 00:25:00,070 S1: State Policy Questions and Actions. State is state of the world. 437 00:25:00,070 --> 00:25:04,480 S1: Exactly like this policy is the desired state of your 438 00:25:04,480 --> 00:25:08,260 S1: company or your life, or whatever questions is like what 439 00:25:08,260 --> 00:25:12,040 S1: are we asking to this I body of knowledge and 440 00:25:12,040 --> 00:25:14,409 S1: then action is what do we do as a result 441 00:25:14,410 --> 00:25:16,750 S1: of learning the answer. Right. So I think that's a 442 00:25:16,750 --> 00:25:19,850 S1: really powerful structure for thinking about how to improve anything, 443 00:25:19,850 --> 00:25:22,430 S1: and it's all based on the stuff we've been talking about. 444 00:25:22,430 --> 00:25:25,879 S1: So some examples of this. A lonely human wants to 445 00:25:25,910 --> 00:25:29,060 S1: be a happy human. So the recommendation is going to 446 00:25:29,060 --> 00:25:31,159 S1: be to connect. And all of these are going to 447 00:25:31,160 --> 00:25:33,979 S1: be extremely granular with lots of different recommendations going all 448 00:25:33,980 --> 00:25:36,560 S1: the way down. Right. And of course this will all 449 00:25:36,560 --> 00:25:39,830 S1: be managed by a Da AI of some sort. You've 450 00:25:39,830 --> 00:25:41,780 S1: got a dying business. You wish you had a thriving 451 00:25:41,780 --> 00:25:44,030 S1: business that was actually making money. What are you going 452 00:25:44,030 --> 00:25:46,669 S1: to do? You're going to increase revenue by doing the 453 00:25:46,670 --> 00:25:50,750 S1: following things. This flow here of understand the current thing, 454 00:25:50,780 --> 00:25:53,480 S1: understand where you want to be and have AI figure 455 00:25:53,510 --> 00:25:55,879 S1: out the stuff in the middle. At first it's going 456 00:25:55,910 --> 00:25:58,909 S1: to be recommending things to us right as it is now, 457 00:25:58,910 --> 00:26:02,930 S1: but very soon and already starting. It's going to just 458 00:26:02,930 --> 00:26:04,580 S1: be like, do you want me to do that for you? 459 00:26:04,580 --> 00:26:07,399 S1: And the businesses are going to be like, yes, absolutely. 460 00:26:07,400 --> 00:26:09,409 S1: That means I don't have to hire anyone. In fact, 461 00:26:09,410 --> 00:26:10,609 S1: that means I can get rid of a lot of 462 00:26:10,609 --> 00:26:13,370 S1: people because you could just do that for me. So 463 00:26:13,369 --> 00:26:15,800 S1: what are the predictions that we can make based on this? 464 00:26:15,800 --> 00:26:19,190 S1: I think of the three. The state piece is the 465 00:26:19,190 --> 00:26:23,780 S1: most important. Just because you can't really do anything unless 466 00:26:23,780 --> 00:26:28,700 S1: you understand what's currently happening, right? It's hard to it's 467 00:26:28,700 --> 00:26:32,270 S1: hard to talk about improvement if you don't know what's wrong. Um, 468 00:26:32,270 --> 00:26:34,760 S1: so state of the human state of a company. State 469 00:26:34,760 --> 00:26:37,250 S1: of a family. Cup of coffee. We've talked about this 470 00:26:37,280 --> 00:26:42,229 S1: infinitely complex changing constantly, and that's why the stuff is 471 00:26:42,230 --> 00:26:45,650 S1: going to scale so crazily, in my opinion. And who 472 00:26:45,650 --> 00:26:48,170 S1: knows if Nvidia is going to win. I think they'll 473 00:26:48,170 --> 00:26:50,450 S1: probably win for at least a couple of years, but 474 00:26:50,450 --> 00:26:54,080 S1: I have no idea they could crash tomorrow. Or they 475 00:26:54,080 --> 00:26:57,980 S1: could whatever, go to the stratosphere in six months? No idea. 476 00:26:58,010 --> 00:27:01,370 S1: So this this brings us to the question how much 477 00:27:01,369 --> 00:27:04,189 S1: context size do we need? How many tokens do we 478 00:27:04,220 --> 00:27:06,380 S1: need to be able to put into context? How big 479 00:27:06,380 --> 00:27:09,379 S1: are Rag systems going to get? How many GPUs do 480 00:27:09,380 --> 00:27:12,230 S1: we need? How many startups do we need in the 481 00:27:12,230 --> 00:27:16,280 S1: AI space? How much I do we actually need? That 482 00:27:16,280 --> 00:27:18,710 S1: is ultimately what we're trying to answer here. And my 483 00:27:18,710 --> 00:27:22,580 S1: answer to that is enough to constantly poll the state 484 00:27:22,580 --> 00:27:26,389 S1: of everything we care about, and then take actions to 485 00:27:26,420 --> 00:27:29,119 S1: fix it to get it into the desired state. And 486 00:27:29,119 --> 00:27:34,070 S1: I think that is a metric crap ton of AI. 487 00:27:34,070 --> 00:27:36,680 S1: And that's why I think we're at like, like I said, 488 00:27:36,710 --> 00:27:40,640 S1: point ten zeros and a one on where it's actually going. 489 00:27:40,670 --> 00:27:44,090 S1: See you in the next one. Unsupervised learning is produced 490 00:27:44,090 --> 00:27:47,150 S1: and edited by Daniel Miessler on a Neumann U87 AI 491 00:27:47,150 --> 00:27:51,380 S1: microphone using Hindenburg. Intro and outro music is by Zomby 492 00:27:51,380 --> 00:27:54,380 S1: with a Y, and to get the text and links 493 00:27:54,380 --> 00:27:56,629 S1: from this episode, sign up for the newsletter version of 494 00:27:56,630 --> 00:28:02,270 S1: the show at Daniel missler.com/newsletter. We'll see you next time.