1 00:00:00,050 --> 00:00:03,440 S1: Whether you're starting or scaling your company's security program, demonstrating 2 00:00:03,470 --> 00:00:06,980 S1: top notch security practices and establishing trust is more important 3 00:00:06,980 --> 00:00:12,110 S1: than ever. Vanta automates compliance for Soc2, ISO 27,001 and more, 4 00:00:12,110 --> 00:00:16,070 S1: saving you time and money while helping you build customer trust. Plus, 5 00:00:16,070 --> 00:00:20,329 S1: you can streamline security reviews by automating questionnaires and demonstrating 6 00:00:20,329 --> 00:00:24,050 S1: your security posture with a customer facing trust center, all 7 00:00:24,050 --> 00:00:28,610 S1: powered by advanced AI. Over 7000 global companies like Atlassian, 8 00:00:28,610 --> 00:00:31,750 S1: Flow Health and Quora use Vanta to manage risk and 9 00:00:31,750 --> 00:00:35,740 S1: prove security in real time. Get $1,000 off Vanta when 10 00:00:35,740 --> 00:00:42,159 S1: you go to Vanta comm slash unsupervised. That's vanta.com/supervised for 11 00:00:42,159 --> 00:00:47,379 S1: $1,000 off. Welcome to Unsupervised Learning, a security, AI, and 12 00:00:47,380 --> 00:00:50,110 S1: meaning focused podcast that looks at how best to thrive 13 00:00:50,110 --> 00:00:55,050 S1: as humans in a post AI world. It combines original ideas, analysis, 14 00:00:55,050 --> 00:00:58,290 S1: and mental models to bring not just the news, but 15 00:00:58,290 --> 00:01:05,880 S1: why it matters and how to respond. All right, welcome 16 00:01:05,880 --> 00:01:08,370 S1: to unsupervised learning. This is Daniel. All right. What do 17 00:01:08,370 --> 00:01:11,819 S1: we have here? All right. So saw this funny thing. 18 00:01:11,819 --> 00:01:18,260 S1: So basically, um, some terrorist funding news. Al Qaeda rebrands 19 00:01:18,260 --> 00:01:23,360 S1: as I. Qaeda raises a $1 billion. Cool. I am 20 00:01:23,360 --> 00:01:28,910 S1: rereading Alex Hormozi. Two business books. Uh, really? I think 21 00:01:28,910 --> 00:01:32,479 S1: Alex Hormozi is, like the best business content generator right now, 22 00:01:32,480 --> 00:01:35,479 S1: at least for my type of business, although he's like 23 00:01:35,480 --> 00:01:39,140 S1: a gym person. Um, so it's not the same type 24 00:01:39,140 --> 00:01:44,470 S1: of business, but he's just extraordinary at, like, user acquisition, 25 00:01:44,590 --> 00:01:47,650 S1: like all sorts of stuff. And I love the format 26 00:01:47,650 --> 00:01:49,390 S1: of his content. Like he's doing a whole bunch of 27 00:01:49,390 --> 00:01:53,440 S1: shorts now. He's just making content probably, I don't know, 28 00:01:53,440 --> 00:01:56,050 S1: it's got to be making 4 or 5 pieces a day. 29 00:01:56,200 --> 00:01:59,230 S1: It's just so much and it's really good. My buddy 30 00:01:59,230 --> 00:02:03,730 S1: Jacoby is an Army veteran, really cool security guy, and 31 00:02:03,730 --> 00:02:06,060 S1: he's struggling really hard right now and he's got a 32 00:02:06,090 --> 00:02:09,540 S1: gofund me. So if you could, uh, add to that, 33 00:02:09,540 --> 00:02:12,480 S1: I would really appreciate it. Fabric made it to the 34 00:02:12,480 --> 00:02:16,500 S1: front page of Hacker News. Uh, next run of augmented 35 00:02:16,500 --> 00:02:19,590 S1: is on July 26th. You want to go sign up there? 36 00:02:19,620 --> 00:02:23,100 S1: Got to spend some in-person time with a couple of 37 00:02:23,100 --> 00:02:27,690 S1: fellow entrepreneurs creators over the weekend. Doing this stuff in 38 00:02:27,690 --> 00:02:31,970 S1: person is invaluable. Like, we're talking about wanting to do 39 00:02:31,970 --> 00:02:35,270 S1: a digital version of The thing that we built, and 40 00:02:35,270 --> 00:02:38,510 S1: I'm reticent to do that because I. I want it 41 00:02:38,510 --> 00:02:42,440 S1: to happen only in person, but we can make it 42 00:02:42,440 --> 00:02:45,200 S1: digitally and then just interact with it in person. So 43 00:02:45,200 --> 00:02:47,570 S1: that's what we're going to do, because it's too valuable 44 00:02:47,570 --> 00:02:51,350 S1: not to be digital, but highly recommend setting up some 45 00:02:51,350 --> 00:02:55,630 S1: similar sessions with your friends because it's really, really powerful. 46 00:02:55,780 --> 00:02:59,079 S1: And my friend Monica Verma is running our CSO masterclass 47 00:02:59,080 --> 00:03:01,630 S1: soon and she also has a newsletter, so go check 48 00:03:01,630 --> 00:03:04,450 S1: those out. My work just finished a resource I've been 49 00:03:04,450 --> 00:03:07,750 S1: working on for over a week called Raid. Real world 50 00:03:07,750 --> 00:03:13,450 S1: AI definitions, not redundant array of inexpensive disks, which is 51 00:03:13,450 --> 00:03:15,549 S1: the first thing I think of whenever I hear Raid. 52 00:03:15,550 --> 00:03:20,370 S1: Either that or Bug Killer. But yeah, real world AI definitions. 53 00:03:20,490 --> 00:03:22,980 S1: And I'm actually going to go into it and talk 54 00:03:22,980 --> 00:03:25,860 S1: about it. So I'm going to go ahead and jump 55 00:03:25,860 --> 00:03:29,310 S1: in here. Okay. So I see a lot of definitions 56 00:03:29,310 --> 00:03:31,380 S1: for different AI terms out there. And I want to 57 00:03:31,380 --> 00:03:34,680 S1: put my own thoughts into a long form version. And 58 00:03:34,680 --> 00:03:37,260 S1: that's basically why I did this. And it's mostly for 59 00:03:37,260 --> 00:03:41,340 S1: me because I love this quote. I forget the name 60 00:03:41,340 --> 00:03:44,570 S1: of the woman. Uh, I need to memorize that. But 61 00:03:44,570 --> 00:03:48,680 S1: she basically says, um, if you are thinking without writing, 62 00:03:48,680 --> 00:03:51,440 S1: you only think that you're thinking. And I think that 63 00:03:51,440 --> 00:03:55,610 S1: is such a powerful statement, because when you are writing 64 00:03:55,610 --> 00:03:57,680 S1: and this is, um, this is why I call this 65 00:03:57,680 --> 00:04:03,080 S1: a Feynman approach to learning, it's basically Feynman said, if 66 00:04:03,080 --> 00:04:05,630 S1: you think you know something, try to teach it and 67 00:04:05,630 --> 00:04:08,000 S1: you will figure out pretty quickly that you probably don't 68 00:04:08,000 --> 00:04:10,460 S1: know it as well as you thought you did. And 69 00:04:10,460 --> 00:04:13,940 S1: I've always known this for a very long time. So 70 00:04:13,940 --> 00:04:19,880 S1: starting back in 1999, I started writing tutorials to myself 71 00:04:19,880 --> 00:04:23,210 S1: to explain things. So when I was learning tech, I 72 00:04:23,210 --> 00:04:25,729 S1: literally would try to learn the tech from textbooks and 73 00:04:25,730 --> 00:04:27,620 S1: then it wouldn't make sense to me. So I would 74 00:04:27,620 --> 00:04:30,330 S1: go write my own explanation and then I would just 75 00:04:30,330 --> 00:04:33,810 S1: use that whenever I wanted to remember something. And that 76 00:04:33,810 --> 00:04:36,750 S1: turned into my tutorials page, which is how I got 77 00:04:36,750 --> 00:04:40,260 S1: started on the internet. So table of contents here, the 78 00:04:40,260 --> 00:04:44,610 S1: expanded definitions table, the one liner definitions table, and then 79 00:04:44,610 --> 00:04:48,900 S1: I break out AGI and ASI into two different sections. 80 00:04:48,900 --> 00:04:51,630 S1: So we're going to start with the one liner AI 81 00:04:51,630 --> 00:04:55,640 S1: definitions table. And this is coming off of Hamel Hussein's 82 00:04:55,670 --> 00:05:00,289 S1: I bullshit knife which gave violently short definitions for a 83 00:05:00,290 --> 00:05:03,440 S1: bunch of I terms. And this is my expanded version 84 00:05:03,440 --> 00:05:07,609 S1: of that. Okay. So this is made to like crystallize 85 00:05:07,610 --> 00:05:10,760 S1: in your brain what these things are in very, very 86 00:05:10,760 --> 00:05:14,270 S1: crisp conversational form. And we're going to go into more 87 00:05:14,270 --> 00:05:17,740 S1: detail in the further ones. But I tech that does 88 00:05:17,740 --> 00:05:22,270 S1: cognitive tasks that only humans could do before. And the 89 00:05:22,270 --> 00:05:25,900 S1: reason I have that only humans could do before is 90 00:05:25,900 --> 00:05:29,260 S1: because that's a moving bar. So before there was a 91 00:05:29,260 --> 00:05:35,470 S1: chess playing robot or a chess playing AI, everyone said, well, yeah, 92 00:05:35,470 --> 00:05:37,870 S1: but it can't play chess. It never will be able 93 00:05:37,870 --> 00:05:40,299 S1: to play chess because that's a human thing. And of 94 00:05:40,300 --> 00:05:43,920 S1: course you could do that now. Image identification. Being able 95 00:05:43,920 --> 00:05:46,740 S1: to tell the difference between a dog and a cat, 96 00:05:46,740 --> 00:05:49,290 S1: that was a very, very hard problem for a very 97 00:05:49,290 --> 00:05:51,660 S1: long time. You have to think in terms of like 98 00:05:51,660 --> 00:05:55,080 S1: what currently is a thing that only humans can do 99 00:05:55,080 --> 00:05:58,800 S1: that I cannot do. Yet that is the barrier which 100 00:05:58,800 --> 00:06:02,370 S1: determines AI or not, because the moment you cross over 101 00:06:02,370 --> 00:06:06,000 S1: into that and I suddenly becomes available to do that, 102 00:06:06,000 --> 00:06:09,479 S1: that's what's currently called I. Now, what's interesting is as 103 00:06:09,480 --> 00:06:12,779 S1: it moves further and further out, people forget that that's 104 00:06:12,779 --> 00:06:16,049 S1: I like image recognition. A lot of people are like, 105 00:06:16,050 --> 00:06:19,110 S1: that's not AI, that's ML. Well, it's still AI, first 106 00:06:19,110 --> 00:06:22,140 S1: of all, because machine learning is a subset of AI, 107 00:06:22,140 --> 00:06:25,890 S1: but also because it's doing something that only humans could 108 00:06:25,890 --> 00:06:30,140 S1: do before. So I, I love that definition. Very flexible 109 00:06:30,290 --> 00:06:34,370 S1: machine learning AI that can improve just by seeing more data. 110 00:06:34,370 --> 00:06:36,770 S1: And we're going to talk about this more, uh, later 111 00:06:36,770 --> 00:06:41,630 S1: on prompting clearly articulate what you want from an AI rag. 112 00:06:41,630 --> 00:06:45,080 S1: Provide context to an AI that's too big or expensive 113 00:06:45,080 --> 00:06:48,140 S1: to fit in a prompt, an agent, an AI component 114 00:06:48,140 --> 00:06:51,650 S1: that does more than just LM call to respond. So 115 00:06:51,650 --> 00:06:55,260 S1: it's not just call and response, that's just regular. I, 116 00:06:55,260 --> 00:06:58,110 S1: an agent, does more than that, taking on more of 117 00:06:58,110 --> 00:07:01,049 S1: the work chain of thought. Tell the I to walk 118 00:07:01,050 --> 00:07:04,440 S1: through its thinking and steps. Zero shot ask an AI 119 00:07:04,470 --> 00:07:08,010 S1: to do something without any examples. Multi-shot. Ask an AI 120 00:07:08,040 --> 00:07:12,180 S1: to do something and provide multiple examples. Prompt injection tricking 121 00:07:12,180 --> 00:07:16,410 S1: an AI into doing something bad. Jailbreaking, bypassing security controls 122 00:07:16,410 --> 00:07:19,100 S1: to get full execution ability, or at least as full 123 00:07:19,100 --> 00:07:23,840 S1: as possible, doesn't necessarily need to be full AGI general 124 00:07:23,840 --> 00:07:27,950 S1: AI smart enough to replace an $80,000 white collar worker, 125 00:07:27,950 --> 00:07:33,170 S1: and ASI, which is superintelligence. General AI that's smarter and 126 00:07:33,170 --> 00:07:36,050 S1: or more capable than any human. So I think that's 127 00:07:36,050 --> 00:07:39,020 S1: a pretty clean list for thinking about these concepts. And 128 00:07:39,020 --> 00:07:41,510 S1: now I want to go into more detail for each. 129 00:07:41,530 --> 00:07:44,950 S1: So AI. AI is technology that does cognitive tasks or 130 00:07:44,950 --> 00:07:48,010 S1: work that could previously only be done by humans. So 131 00:07:48,010 --> 00:07:51,130 S1: cognitive tasks or work, and I might clean that up 132 00:07:51,130 --> 00:07:55,630 S1: at some point and just make it work or cognitive tasks. Anyway, 133 00:07:55,630 --> 00:07:58,000 S1: lots of different ways to define AI. So this is 134 00:07:58,000 --> 00:08:00,610 S1: going to be controversial. And this is what I talked about. 135 00:08:00,610 --> 00:08:03,040 S1: Well yeah of course I can do this. But it 136 00:08:03,040 --> 00:08:05,170 S1: still can't do this and it probably never will be 137 00:08:05,170 --> 00:08:07,650 S1: able to. And then the narrator voice says, yeah, that 138 00:08:07,650 --> 00:08:10,230 S1: happened seven months later. And this is why I like 139 00:08:10,230 --> 00:08:13,770 S1: the definition is because it moves over time. Machine learning 140 00:08:13,800 --> 00:08:16,920 S1: a subset of AI that enables a system to learn 141 00:08:16,920 --> 00:08:20,400 S1: from data alone, rather than needing to be explicitly programmed. 142 00:08:20,400 --> 00:08:24,990 S1: This is a super exciting to me as a definition. Actually, 143 00:08:24,990 --> 00:08:28,679 S1: as a concept. I can't believe the program is the same, 144 00:08:28,680 --> 00:08:32,100 S1: but you change the data and the program gets smarter. 145 00:08:32,130 --> 00:08:35,160 S1: That blows my mind. And it's like the most important. 146 00:08:35,160 --> 00:08:37,950 S1: In fact, if you want to go tighter learning from 147 00:08:37,950 --> 00:08:41,340 S1: data alone or a technology system that's able to learn 148 00:08:41,340 --> 00:08:44,069 S1: from data alone, like lots of different ways you can 149 00:08:44,070 --> 00:08:47,580 S1: skin this thing. Prompt engineering. Prompt engineering is the art 150 00:08:47,580 --> 00:08:51,180 S1: and science of using language, usually text, to get AI 151 00:08:51,210 --> 00:08:53,730 S1: to do precisely what you want it to do. Yeah, 152 00:08:53,760 --> 00:08:56,410 S1: a lot of people think prompt engineering is like this 153 00:08:56,410 --> 00:08:59,590 S1: really special thing. Others think it's not important at all. 154 00:08:59,590 --> 00:09:03,490 S1: I think it's absolutely an art and a science because 155 00:09:03,490 --> 00:09:07,150 S1: it's more about clear thinking than about text itself. Just 156 00:09:07,150 --> 00:09:10,330 S1: like writing, the best prompt engineer is the same. It 157 00:09:10,330 --> 00:09:13,330 S1: comes from deeply understanding the problem and be able to 158 00:09:13,330 --> 00:09:16,150 S1: break out your instructions to the AI in a very 159 00:09:16,150 --> 00:09:19,550 S1: methodical and clear way, So the clearer you can think 160 00:09:19,550 --> 00:09:23,599 S1: and the clearer you can think about problems and describe methodologies. 161 00:09:23,600 --> 00:09:25,640 S1: That's that's what's going to make you really good at 162 00:09:25,640 --> 00:09:28,760 S1: prompt engineering. Don't think of it as an AI thing. 163 00:09:28,760 --> 00:09:32,960 S1: Think of it as a thinking thing and a writing thing. Retrieval. 164 00:09:32,960 --> 00:09:36,920 S1: Augmented generation Rag is the process of taking large quantities 165 00:09:36,920 --> 00:09:39,470 S1: of data, which are either too large or too expensive 166 00:09:39,470 --> 00:09:41,990 S1: to put in a prompt directly, and making that data 167 00:09:42,010 --> 00:09:46,510 S1: usable as vectorized embeddings to AI at runtime. And I 168 00:09:46,510 --> 00:09:49,720 S1: say it's important to understand that Rag is a hack 169 00:09:49,720 --> 00:09:52,750 S1: that solves a specific problem, which is that people and 170 00:09:52,750 --> 00:09:56,650 S1: companies have way too much data gigabytes, terabytes, petabytes of 171 00:09:56,650 --> 00:09:59,140 S1: data that they want their AI to be aware of 172 00:09:59,140 --> 00:10:02,650 S1: when performing tasks. But I can only handle small amounts 173 00:10:02,650 --> 00:10:05,740 S1: of that data per interaction. And that's what Rag is for. 174 00:10:05,740 --> 00:10:08,189 S1: So the solution we come up with is to use 175 00:10:08,190 --> 00:10:12,090 S1: embeddings and vector databases to encode relevant information and send 176 00:10:12,090 --> 00:10:14,550 S1: that along with the prompts. And it's not really clear 177 00:10:14,550 --> 00:10:17,160 S1: what the successor is going to be. But one option 178 00:10:17,160 --> 00:10:20,520 S1: is to just have massive context windows and put the 179 00:10:20,520 --> 00:10:23,189 S1: raw content in there. But then you have cost problems, 180 00:10:23,190 --> 00:10:24,990 S1: so you would need that to be very cheap for 181 00:10:25,020 --> 00:10:28,560 S1: that to work. Agents. An AI agent is an AI 182 00:10:28,590 --> 00:10:32,540 S1: component that interprets instructions and takes on more of the 183 00:10:32,540 --> 00:10:37,430 S1: work in a total AI workflow than just LLM response, 184 00:10:37,429 --> 00:10:43,729 S1: for example, executing functions, performing data lookups, etc. before passing 185 00:10:43,730 --> 00:10:47,960 S1: on results. Yeah, so there are lots of different definitions 186 00:10:47,960 --> 00:10:51,079 S1: for agent. I think the trick is to go back 187 00:10:51,080 --> 00:10:55,510 S1: to the Latin, the agents or Arjun's, which is to 188 00:10:55,510 --> 00:10:59,020 S1: do or to act. So I think eventually this is 189 00:10:59,020 --> 00:11:01,210 S1: going to be like an AI component that has its 190 00:11:01,210 --> 00:11:04,210 S1: own mission and goals and can use its resources and 191 00:11:04,210 --> 00:11:07,630 S1: capabilities to accomplish them in a self-directed way. That is 192 00:11:07,630 --> 00:11:11,740 S1: the true nature of agent or agents from the Latin. 193 00:11:11,740 --> 00:11:14,650 S1: But I think in this current version, I think the 194 00:11:14,650 --> 00:11:17,230 S1: best way to say it is that anything that acts 195 00:11:17,230 --> 00:11:20,250 S1: on behalf of the mission and something that performs multiple 196 00:11:20,250 --> 00:11:23,910 S1: steps towards the final goal, and that's why I'm using 197 00:11:23,910 --> 00:11:26,730 S1: this definition for agent chain of thought, a way of 198 00:11:26,730 --> 00:11:29,040 S1: interacting with AI in which you don't just say what 199 00:11:29,040 --> 00:11:30,990 S1: you want, but you give the steps that you would 200 00:11:30,990 --> 00:11:33,660 S1: take to accomplish the task. So you're basically teaching it 201 00:11:33,660 --> 00:11:36,449 S1: how you think, which which is essentially teaching it how 202 00:11:36,450 --> 00:11:39,059 S1: to think like a human. And then it uses that 203 00:11:39,059 --> 00:11:41,699 S1: template to sort of do that every time it's going 204 00:11:41,700 --> 00:11:44,000 S1: to solve the same type of problem. And to me, 205 00:11:44,000 --> 00:11:46,100 S1: it's kind of an example of what we talked about 206 00:11:46,100 --> 00:11:50,059 S1: in prompt engineering, clear thinking. It's walking the AI through 207 00:11:50,059 --> 00:11:52,910 S1: how you think when you're solving a problem yourself. Prompt 208 00:11:52,910 --> 00:11:56,300 S1: injection a method of using language, usually text, to get 209 00:11:56,300 --> 00:11:58,460 S1: AI to do something it's not supposed to do. And 210 00:11:58,460 --> 00:12:01,310 S1: a lot of people confuse prompt injection with jailbreaking. And 211 00:12:01,309 --> 00:12:03,290 S1: I think the best way to think about this. And 212 00:12:03,290 --> 00:12:06,530 S1: thanks to Jason Haddox for talking through this with me, 213 00:12:06,530 --> 00:12:09,800 S1: because he had a very similar definition, uh, is to 214 00:12:09,800 --> 00:12:13,610 S1: say prompt injection is a method. It's an attack avenue. 215 00:12:13,610 --> 00:12:17,300 S1: It's not a payload, whereas jailbreaking is a goal. We're 216 00:12:17,300 --> 00:12:20,900 S1: trying to bypass the security on a system. Prompt injection 217 00:12:20,900 --> 00:12:23,840 S1: is not bypassing the security on a system. Well, it's 218 00:12:23,840 --> 00:12:26,780 S1: a method for doing that. It's a way of tricking 219 00:12:26,780 --> 00:12:29,359 S1: the eye into doing something it's not supposed to do. 220 00:12:29,360 --> 00:12:32,790 S1: But that could be telling a joke that it's not 221 00:12:32,790 --> 00:12:35,340 S1: supposed to tell. That could be showing a picture it's 222 00:12:35,340 --> 00:12:37,710 S1: not supposed to show. It doesn't mean it's bypassing the 223 00:12:37,710 --> 00:12:40,590 S1: security on a system so you can execute commands. So 224 00:12:40,590 --> 00:12:42,990 S1: it could be data exfil, it could be code execution, 225 00:12:42,990 --> 00:12:46,260 S1: it could be image generation. It could be anything. So easiest, 226 00:12:46,260 --> 00:12:48,390 S1: cleanest way, at least in my mind, to think about this. 227 00:12:48,390 --> 00:12:51,630 S1: Prompt injection is a method. Jailbreaking is a goal. All right. 228 00:12:51,630 --> 00:12:54,920 S1: We talked about jailbreaking. Going to pass that one. Artificial 229 00:12:54,920 --> 00:12:58,939 S1: general intelligence. The ability for an AI system, whether a model, 230 00:12:58,940 --> 00:13:01,700 S1: a product or a system to fully do the job 231 00:13:01,700 --> 00:13:04,849 S1: of an average US based knowledge worker. In other words, 232 00:13:04,850 --> 00:13:07,580 S1: it's not just competent at doing a specific thing, which 233 00:13:07,580 --> 00:13:12,170 S1: is called narrow AI oftentimes, but many different things, which 234 00:13:12,170 --> 00:13:15,260 S1: is why it's called general. So basically it's some combination 235 00:13:15,260 --> 00:13:19,000 S1: of sufficiently general and sufficiently competent, and the amounts of 236 00:13:19,000 --> 00:13:21,760 S1: which are, you know, going to be debated so generally 237 00:13:21,760 --> 00:13:25,870 S1: competent at many other cognitive tasks, sample efficient so we 238 00:13:25,870 --> 00:13:29,020 S1: can learn quickly from like one example or a low 239 00:13:29,020 --> 00:13:31,449 S1: number of examples. And the system should be able to 240 00:13:31,450 --> 00:13:35,380 S1: apply new information, concepts and skills that learns to additional 241 00:13:35,380 --> 00:13:39,430 S1: new problems. So it's not just generally intelligent from its training, 242 00:13:39,429 --> 00:13:42,340 S1: but when it learns new things, it's generally intelligent at 243 00:13:42,360 --> 00:13:44,970 S1: applying those new things that learn to everything else in 244 00:13:44,970 --> 00:13:48,870 S1: the future. And I like this definition focusing on the 245 00:13:48,870 --> 00:13:54,390 S1: worker replacing a worker, because I think this is arguably 246 00:13:54,390 --> 00:13:58,020 S1: what matters most to humans when we talk about AI discussion, 247 00:13:58,020 --> 00:14:01,230 S1: which is the future of humans in a world of AI, 248 00:14:01,260 --> 00:14:05,520 S1: regular people don't care about processing speed or model size 249 00:14:05,520 --> 00:14:07,949 S1: or model weights or any of that crap. What they 250 00:14:07,950 --> 00:14:10,080 S1: care about is if and when any of this is 251 00:14:10,080 --> 00:14:13,380 S1: going to actually affect them. And that means job replacement mostly, 252 00:14:13,380 --> 00:14:16,320 S1: or at least that's the most important thing. So these 253 00:14:16,320 --> 00:14:20,280 S1: are the levels that I see within that job replacement 254 00:14:20,280 --> 00:14:23,790 S1: of this average white collar worker. So first level is 255 00:14:23,790 --> 00:14:27,239 S1: better but with significant drawbacks. And I'm not going to 256 00:14:27,240 --> 00:14:29,850 S1: go through all of these. But it's like the interface 257 00:14:29,850 --> 00:14:33,910 S1: the language, the confusion, the errors, flexibility, basically. Basically, real 258 00:14:33,910 --> 00:14:38,020 S1: world work environments are kludgy and lame and they change 259 00:14:38,020 --> 00:14:41,350 S1: all the time. And it's basically a giant mess. This 260 00:14:41,350 --> 00:14:44,650 S1: level one is basically saying, yes, we brought in an 261 00:14:44,650 --> 00:14:47,710 S1: AI worker to replace that person, or we didn't hire 262 00:14:47,710 --> 00:14:50,770 S1: a person. We used an AI worker instead for for 263 00:14:50,770 --> 00:14:53,530 S1: growth if you want to be nice. So that's what happened. 264 00:14:53,530 --> 00:14:57,720 S1: We brought in this level one AGI one. And here's 265 00:14:57,720 --> 00:15:01,470 S1: the problem. It's kludgy. So so the problem is that 266 00:15:01,470 --> 00:15:03,900 S1: they're doing the work. But you have to clean up 267 00:15:03,900 --> 00:15:07,950 S1: after them. So the interface proprietary cumbersome the language that 268 00:15:07,950 --> 00:15:10,740 S1: you have to use. It's still kind of like AI language. 269 00:15:10,740 --> 00:15:14,010 S1: It's almost like you're prompting. You're not just normally talking 270 00:15:14,010 --> 00:15:16,890 S1: to them. Um, it gets confused a decent amount of 271 00:15:16,890 --> 00:15:19,820 S1: the time. There are frequent mistakes which need to be 272 00:15:19,820 --> 00:15:23,120 S1: fixed by humans, and if you want to change what 273 00:15:23,120 --> 00:15:26,720 S1: this thing is doing, it needs to be largely retooled. 274 00:15:27,260 --> 00:15:30,020 S1: Not not like in a major way, but you've got 275 00:15:30,020 --> 00:15:31,790 S1: to have a conversation with it. You've got to change 276 00:15:31,790 --> 00:15:34,520 S1: some documentation, you've got to do some stuff. So going 277 00:15:34,520 --> 00:15:38,000 S1: back to the top level of this one better but 278 00:15:38,000 --> 00:15:41,150 S1: with significant drawbacks. So it is right at the line 279 00:15:41,150 --> 00:15:43,330 S1: of being too much work and not good enough to 280 00:15:43,330 --> 00:15:49,150 S1: replace a human right. That's the trick. It's just barely better. Barely. 281 00:15:49,150 --> 00:15:51,520 S1: And sometimes you can't tell if it's better or not. 282 00:15:51,520 --> 00:15:54,700 S1: It might be too much work. Next one. It removes 283 00:15:54,700 --> 00:15:57,400 S1: a lot of those problems. So it's like in the middle. 284 00:15:57,400 --> 00:16:03,940 S1: It's like decent, competent but imperfect. So mostly normal workflows, 285 00:16:03,940 --> 00:16:08,160 S1: mostly human with exceptions. Doesn't get confused very often. Fewer 286 00:16:08,160 --> 00:16:10,920 S1: mistakes and the mistakes aren't as bad, and it just 287 00:16:10,920 --> 00:16:14,250 S1: decently well to new direction from leaders. So that's like 288 00:16:14,250 --> 00:16:19,770 S1: the middle level. And then AGI three is full worker replacement. 289 00:16:19,770 --> 00:16:26,070 S1: So this can replace this average 80 K worker in 2022. 290 00:16:26,100 --> 00:16:29,610 S1: White collar worker in the US. It can fully replace them. 291 00:16:29,610 --> 00:16:32,630 S1: So when you talk to them, it's you talk to 292 00:16:32,630 --> 00:16:34,940 S1: them just like any other employee. You could text them, 293 00:16:34,940 --> 00:16:37,130 S1: you could voice them, you could do video with them. 294 00:16:37,130 --> 00:16:39,410 S1: It's just like an employee. You talk to them exactly 295 00:16:39,410 --> 00:16:42,950 S1: like an employee. They're confused the same amount or less 296 00:16:42,950 --> 00:16:46,490 S1: than the $80,000 employee. Keep in mind, humans have all 297 00:16:46,490 --> 00:16:49,250 S1: these problems too, right? Humans also get confused and need 298 00:16:49,250 --> 00:16:51,230 S1: to be retasked or whatever. So it's not like we're 299 00:16:51,230 --> 00:16:54,200 S1: competing with perfection here. Same or fewer mistakes than an 300 00:16:54,200 --> 00:16:58,210 S1: $80,000 employee. Employees also make mistakes and just as flexible 301 00:16:58,210 --> 00:17:02,050 S1: or more than $80,000 employee employees also have trouble like 302 00:17:02,050 --> 00:17:04,510 S1: completely being retests. It's like I thought it was working 303 00:17:04,510 --> 00:17:06,520 S1: on that. What are you doing? Like, so this is 304 00:17:06,520 --> 00:17:09,790 S1: basically saying it's just as good as them or or better. 305 00:17:09,790 --> 00:17:14,649 S1: So this is the top level of complete worker replacement. 306 00:17:14,650 --> 00:17:16,420 S1: I think this is going to take a decent amount 307 00:17:16,420 --> 00:17:18,550 S1: of time. And then the next question is like okay, 308 00:17:18,550 --> 00:17:21,940 S1: but what if it's like that? It's a full replacement 309 00:17:21,940 --> 00:17:24,730 S1: for that worker. So it's AGI three, but with an 310 00:17:24,730 --> 00:17:28,930 S1: IQ of 150. Well that's that's really powerful. And I'm 311 00:17:28,930 --> 00:17:32,440 S1: not working on that axis. I'm only working on how 312 00:17:32,440 --> 00:17:35,710 S1: good of a worker are they. And that's the three 313 00:17:35,710 --> 00:17:39,010 S1: levels for AGI okay. Let's do AC a level of 314 00:17:39,010 --> 00:17:43,180 S1: general AI that's smarter and more capable than any human 315 00:17:43,180 --> 00:17:45,830 S1: that's ever lived. So this is interesting for a number 316 00:17:45,830 --> 00:17:49,160 S1: of reasons. One, it's a threshold that sits above AGI 317 00:17:49,160 --> 00:17:52,100 S1: and people don't really agree on that definition. And the 318 00:17:52,100 --> 00:17:54,709 S1: second thing is, at least as I'm defining it, it 319 00:17:54,710 --> 00:17:58,490 S1: has a massive range. And third, it blends with AGI 320 00:17:58,490 --> 00:18:02,960 S1: because AGI just really means general incompetent. And ASI is 321 00:18:02,960 --> 00:18:08,139 S1: also general incompetent. So really ASI is also AGI. It's 322 00:18:08,140 --> 00:18:12,040 S1: just uh, at a level beyond any human. So AGI 323 00:18:12,040 --> 00:18:15,489 S1: is replacement for a human worker making SDK in the 324 00:18:15,490 --> 00:18:20,260 S1: US in 2022, which is pretty AI, prision AI and 325 00:18:20,260 --> 00:18:25,449 S1: ASI is more smart and more competent than any human. 326 00:18:25,450 --> 00:18:28,659 S1: And the model I'm using there is like John von Neumann, 327 00:18:28,660 --> 00:18:33,900 S1: because unlike Einstein and Newton, he was, uh, very broad. 328 00:18:33,930 --> 00:18:37,980 S1: He was a generalist, right. So he did game theory, physics, computing, 329 00:18:37,980 --> 00:18:41,580 S1: lots of other stuff. But I don't think being smarter 330 00:18:41,580 --> 00:18:44,070 S1: is the only thing that matters, because I think it 331 00:18:44,070 --> 00:18:46,260 S1: comes down to multiple things. So if I look at 332 00:18:46,260 --> 00:18:48,720 S1: all the different things that I'm putting into ASI, it's 333 00:18:48,720 --> 00:18:52,080 S1: like your ability to model abstractions of reality at these 334 00:18:52,080 --> 00:18:57,590 S1: different levels. It's the ability to take action, to perceive, understand, improve, solve, create, destroy. 335 00:18:57,590 --> 00:19:00,620 S1: And then it's like, okay, what fields are we innovating in? 336 00:19:00,619 --> 00:19:05,000 S1: What problems are we able to solve and what scale 337 00:19:05,030 --> 00:19:08,630 S1: are we actually working? You know, thinking about, I guess 338 00:19:08,630 --> 00:19:11,540 S1: I should add universe or whatever, but all the way 339 00:19:11,540 --> 00:19:15,109 S1: from Quark to Galaxy is like the scale. So what's 340 00:19:15,109 --> 00:19:17,330 S1: really cool about this is you can now turn these 341 00:19:17,330 --> 00:19:20,859 S1: into functional phrases. So you could say, like an AI 342 00:19:20,890 --> 00:19:26,710 S1: capable of curing aging by creating new chemicals that affect DNA. Okay. 343 00:19:26,710 --> 00:19:30,160 S1: So we have a problem here creating new chemicals. That's 344 00:19:30,160 --> 00:19:34,330 S1: like a net new thing, right? So that you're seeing this, 345 00:19:34,330 --> 00:19:37,810 S1: you know, verbs and nouns here, maintaining a full city 346 00:19:37,810 --> 00:19:42,129 S1: by monitoring and adjusting all public resources in real time. 347 00:19:42,130 --> 00:19:44,740 S1: And the AI capable of taking over a country by 348 00:19:44,740 --> 00:19:49,119 S1: manufacturing a drone army and a new energy based weapon. 349 00:19:49,119 --> 00:19:51,520 S1: So this is like new science here, and an AI 350 00:19:51,550 --> 00:19:55,960 S1: capable of faster than light travel by discovering completely new physics. 351 00:19:55,960 --> 00:19:59,320 S1: So that's like a level above, right? So let's look 352 00:19:59,320 --> 00:20:03,430 S1: at the different levels here. AC one superior, an AI 353 00:20:03,460 --> 00:20:07,790 S1: capable of making incremental improvements in multiple fields and managing 354 00:20:07,790 --> 00:20:11,750 S1: up to city size entities on its own. So here 355 00:20:11,750 --> 00:20:13,730 S1: are the things that it has smarter and more capable 356 00:20:13,730 --> 00:20:17,300 S1: than any human on many topics. Able to move progress 357 00:20:17,300 --> 00:20:21,439 S1: forward in multiple scientific fields. Able to recommend novel solutions 358 00:20:21,440 --> 00:20:25,070 S1: to many of our main problems. Able to copy and 359 00:20:25,070 --> 00:20:28,970 S1: surpass the creativity of many top artists. Able to fully 360 00:20:28,970 --> 00:20:32,889 S1: manage a large company or city itself and keep, keep 361 00:20:32,890 --> 00:20:36,970 S1: in mind, copy and surpass some of the best artists. Now, 362 00:20:36,970 --> 00:20:40,510 S1: what this really means is like in a superhuman way, right? 363 00:20:40,510 --> 00:20:44,020 S1: So it's not just like better in one area. It's like, okay, 364 00:20:44,020 --> 00:20:48,310 S1: it's Eminem, but it's way better than Eminem. It's Kendrick Lamar, 365 00:20:48,310 --> 00:20:52,360 S1: but five times better, right? Same with these types of things. 366 00:20:52,359 --> 00:20:56,910 S1: Novel solutions, moving progress forward, which obviously no human has done. 367 00:20:56,910 --> 00:20:59,790 S1: So they're doing it better, but it doesn't have these. 368 00:20:59,790 --> 00:21:04,200 S1: It can't create. Net new physics or net new material science. 369 00:21:04,200 --> 00:21:07,500 S1: It's unable to fundamentally improve itself by orders of magnitude. 370 00:21:07,500 --> 00:21:12,150 S1: So next phase dominant okay. First one was superior. Next 371 00:21:12,150 --> 00:21:14,879 S1: one is dominance. This one has all of AC one. 372 00:21:14,880 --> 00:21:17,550 S1: But it's also able to completely change how we see 373 00:21:17,550 --> 00:21:21,440 S1: multiple fields able to completely solve most of our current problems. 374 00:21:21,470 --> 00:21:25,370 S1: Able to fully manage a country by itself. So this 375 00:21:25,369 --> 00:21:27,770 S1: is a huge jump over the first one. Able to 376 00:21:27,770 --> 00:21:31,760 S1: fundamentally improve itself by orders of magnitude, but still can't create. 377 00:21:31,760 --> 00:21:34,100 S1: Net new physics or run an entire planet. And then 378 00:21:34,100 --> 00:21:38,030 S1: you have AC three, another massive jump AI capable of 379 00:21:38,030 --> 00:21:43,429 S1: creating net new physics, completely new materials, manipulation of near 380 00:21:43,450 --> 00:21:47,830 S1: fundamental reality or fundamental reality. If it's possible to do 381 00:21:47,830 --> 00:21:52,270 S1: that and can run an entire planet. And keep in mind, 382 00:21:52,270 --> 00:21:55,780 S1: I might add AC for um, if I have a 383 00:21:55,780 --> 00:22:00,850 S1: like a clear picture of what that looks like. Um, Yoshi. No. Rishi, 384 00:22:00,880 --> 00:22:05,379 S1: my buddy Rishi mentioned, um, what about someone? An AI 385 00:22:05,380 --> 00:22:08,409 S1: that's smarter than all of humanity put together? And I 386 00:22:08,410 --> 00:22:12,850 S1: was like, well, isn't that confusing knowledge with intelligence and reason? 387 00:22:12,850 --> 00:22:15,280 S1: But I haven't heard back from him yet. Other things 388 00:22:15,280 --> 00:22:18,460 S1: to consider here. So this one adds the ability to 389 00:22:18,460 --> 00:22:22,360 S1: modify reality at a fundamental or near fundamental level, able 390 00:22:22,359 --> 00:22:26,350 S1: to manage an entire planet simultaneously. And maybe its primary 391 00:22:26,350 --> 00:22:30,310 S1: concerns become like sun expansion, which is going to like 392 00:22:30,310 --> 00:22:33,530 S1: eat up the earth or populating the galaxy and beyond, 393 00:22:33,530 --> 00:22:35,990 S1: or the heat death of the universe, which is like 394 00:22:35,990 --> 00:22:39,170 S1: escaping this reality. Yeah, or breaking out of the simulation. 395 00:22:39,170 --> 00:22:41,959 S1: I should add that. So as a summary, I terms 396 00:22:41,960 --> 00:22:45,320 S1: are confusing. It's nice to have simple, practical versions. It's 397 00:22:45,320 --> 00:22:48,620 S1: useful to crystallize your own definitions on paper, both for 398 00:22:48,619 --> 00:22:50,900 S1: your own reference and to see if your definitions are 399 00:22:50,900 --> 00:22:55,130 S1: consistent with each other. And I think these eye definitions 400 00:22:55,130 --> 00:22:59,350 S1: work best when they're human focused and practically worded, because 401 00:22:59,350 --> 00:23:01,629 S1: it is very easy to try to be technical with 402 00:23:01,630 --> 00:23:05,410 S1: these definitions, and you instantly open up like this giant 403 00:23:05,410 --> 00:23:09,580 S1: mess the moment you go a level down in technicality. 404 00:23:09,580 --> 00:23:13,449 S1: And what you'll find is if you get 20 AI 405 00:23:13,450 --> 00:23:17,230 S1: experts in in a room and a bunch of ML experts, 406 00:23:17,230 --> 00:23:20,160 S1: they've been doing this 30 years and you ask them 407 00:23:20,160 --> 00:23:23,100 S1: their opinions. They don't agree. The textbooks don't agree, the 408 00:23:23,100 --> 00:23:26,190 S1: experts don't agree. The moment you go a layer down, 409 00:23:26,190 --> 00:23:30,540 S1: it's a giant mess. And that giant mess stops conversations 410 00:23:30,540 --> 00:23:33,360 S1: from happening because they end up, they'll get two hours 411 00:23:33,359 --> 00:23:35,340 S1: into a podcast and be like, oh, that's what you 412 00:23:35,340 --> 00:23:37,560 S1: meant by I. That's not what I mean by I. 413 00:23:37,560 --> 00:23:41,280 S1: And I'm like, okay, let's stop messing with let's stop 414 00:23:41,280 --> 00:23:44,300 S1: wasting two hours in a conversation because we're not talking 415 00:23:44,300 --> 00:23:47,990 S1: about the same thing. Let's level up to an abstraction 416 00:23:47,990 --> 00:23:51,230 S1: level where it's useful in a conversation. That's why it's 417 00:23:51,230 --> 00:23:54,290 S1: called real world AI definition. All right. So that's that one. 418 00:23:54,290 --> 00:23:57,439 S1: All right. Stories. Cloudflare has a new free tool to 419 00:23:57,440 --> 00:24:00,680 S1: stop bots from scraping websites. Remember I told you about 420 00:24:00,680 --> 00:24:03,620 S1: Cloudflare how they just do like the coolest stuff and 421 00:24:03,619 --> 00:24:06,950 S1: they just Cloudflare is a beast. I am telling you. 422 00:24:06,950 --> 00:24:10,930 S1: They just they see a problem that is annoying everybody 423 00:24:10,930 --> 00:24:13,270 S1: and they go and just pour a little bit of 424 00:24:13,270 --> 00:24:16,270 S1: like cement in that seam or that crack or that 425 00:24:16,270 --> 00:24:19,149 S1: hole or whatever, and they're just like, we're going to 426 00:24:19,150 --> 00:24:21,669 S1: solve this. And if no one else does it, they're 427 00:24:21,670 --> 00:24:23,710 S1: just going to be the solution for this. Like they're 428 00:24:23,710 --> 00:24:26,170 S1: doing this with Captcha. They're doing this now with AI 429 00:24:26,200 --> 00:24:30,040 S1: scraping like somebody over there is going, this is going 430 00:24:30,040 --> 00:24:31,810 S1: to be a problem in the future or this is 431 00:24:31,830 --> 00:24:34,560 S1: already becoming a problem. Let's get there first. And I 432 00:24:34,560 --> 00:24:37,530 S1: massively respect it. And before too long, the whole whole 433 00:24:37,530 --> 00:24:40,530 S1: internet's going to be Cloudflare, because over the course of time, 434 00:24:40,530 --> 00:24:44,070 S1: they've made like 500 of these tools at some point. Right. 435 00:24:44,070 --> 00:24:47,430 S1: And it's like all those random little tools that they 436 00:24:47,430 --> 00:24:51,060 S1: made become the structure of the internet. Twilio says, uh, 437 00:24:51,060 --> 00:24:55,500 S1: somebody's got phone numbers because of an unsecured auth API endpoint, 438 00:24:55,500 --> 00:24:59,869 S1: 33 million phone numbers. Russian experts say they fully analyzed 439 00:24:59,869 --> 00:25:05,210 S1: the structure of the American attack Miss Atacms missile, and 440 00:25:05,210 --> 00:25:09,140 S1: they believe that's going to help them fighting Ukrainian missiles, 441 00:25:09,140 --> 00:25:13,100 S1: which is not good. Thanks to tines for sponsoring. North 442 00:25:13,100 --> 00:25:15,649 S1: Korea has switched its state TV broadcast from a Chinese 443 00:25:15,650 --> 00:25:18,320 S1: satellite to a Russian one. And now fewer people can 444 00:25:18,320 --> 00:25:22,369 S1: watch Hezbollah launch over 200 rockets and drones at Israel 445 00:25:22,369 --> 00:25:26,930 S1: after Israel killed a senior Hezbollah official. Thanks to nudge 446 00:25:26,930 --> 00:25:31,280 S1: security for sponsoring. US intelligence community is diving into generative 447 00:25:31,280 --> 00:25:36,050 S1: AI to enhance intelligence operations, using it for search, discovery, 448 00:25:36,050 --> 00:25:41,149 S1: counterargument generation. Uh, I'm going to make so much Intel stuff, 449 00:25:41,150 --> 00:25:44,700 S1: I cannot wait to build that product in substrate. I 450 00:25:44,700 --> 00:25:47,970 S1: cannot wait. There's some analysis saying I cannot be funny 451 00:25:47,970 --> 00:25:52,980 S1: from Anya, Jeremy Greenwald, and I think I disagree. And 452 00:25:52,980 --> 00:25:56,459 S1: here's an example. My buddy Joseph posted this one. It's 453 00:25:56,460 --> 00:26:00,330 S1: a glyph app and it makes memes. Look at this 454 00:26:00,330 --> 00:26:05,160 S1: bug bounty, bro. Like I legitimately laughed at some of these. 455 00:26:05,160 --> 00:26:10,190 S1: Show me show you my leet skills. Use the inspect 456 00:26:10,190 --> 00:26:14,270 S1: element I only hack ethically dos school website in eighth grade. 457 00:26:14,390 --> 00:26:18,230 S1: I mean, first of all, these are real. They're good 458 00:26:18,230 --> 00:26:20,629 S1: in terms of like they actually apply to the field 459 00:26:20,630 --> 00:26:24,619 S1: that they're making fun of. And, uh, my POC is flawless. 460 00:26:24,650 --> 00:26:28,970 S1: It's a Rick roll link. I generated this. Okay. I've 461 00:26:28,970 --> 00:26:31,490 S1: been saying for a while that being funny is a 462 00:26:31,490 --> 00:26:37,209 S1: major milestone in human AI, in AI becoming real. And 463 00:26:37,210 --> 00:26:39,580 S1: just like a lot of other things in AI, there 464 00:26:39,580 --> 00:26:41,440 S1: are so many people who are like, oh yeah, it can. 465 00:26:41,440 --> 00:26:43,270 S1: It could tell a cat from a dog, but it's 466 00:26:43,270 --> 00:26:46,120 S1: never going to make me laugh. Whatever. The stuff's been 467 00:26:46,119 --> 00:26:49,990 S1: out for like 18 minutes, this just happened at the 468 00:26:49,990 --> 00:26:53,500 S1: end of 2022. Like this is still day one. I'm 469 00:26:53,500 --> 00:26:56,359 S1: telling you right now, there will be funny I very 470 00:26:56,359 --> 00:26:58,880 S1: soon and this is the first version of it. I mean, 471 00:26:58,880 --> 00:27:02,870 S1: this is already funny. I technically the real standard is 472 00:27:02,869 --> 00:27:06,050 S1: going to be a stand up routine, a deepfake of 473 00:27:06,050 --> 00:27:10,190 S1: a face, or actually a deepfake. Okay, I am watching 474 00:27:10,190 --> 00:27:14,090 S1: a stage. There is a deepfake, uh, person standing there. 475 00:27:14,450 --> 00:27:16,430 S1: It doesn't even have to convince me. It could be 476 00:27:16,430 --> 00:27:20,679 S1: like uncanny valley. But if there's a deepfake person standing 477 00:27:20,680 --> 00:27:24,490 S1: there talking, doing a stand up routine and it makes 478 00:27:24,490 --> 00:27:26,649 S1: me laugh the way it makes me laugh when I 479 00:27:26,650 --> 00:27:30,399 S1: watch other stand ups, that is a major milestone. That 480 00:27:30,400 --> 00:27:35,290 S1: is extraordinary. I don't know how long that's going to. Oh, 481 00:27:35,320 --> 00:27:37,840 S1: I should do that on a prediction market. Yeah, I'm 482 00:27:37,840 --> 00:27:39,820 S1: curious about that. I'm guessing like a year and a 483 00:27:39,820 --> 00:27:41,800 S1: half to two years. I think it's going to be 484 00:27:41,800 --> 00:27:44,970 S1: a little while because that one's really, really difficult. But 485 00:27:44,970 --> 00:27:48,000 S1: it could be, honestly, it could be opus, it could 486 00:27:48,000 --> 00:27:50,970 S1: be opus 35 or GPT five. It could happen in 487 00:27:50,970 --> 00:27:53,820 S1: this next iteration, which is going to be around the 488 00:27:53,820 --> 00:27:56,520 S1: end of this year or beginning of next year. Okay. 489 00:27:56,520 --> 00:27:59,580 S1: Nvidia is set to make 12 billion by selling over 490 00:27:59,580 --> 00:28:04,590 S1: a million HG, H20 GPUs. That sounded impressive. I thought 491 00:28:04,590 --> 00:28:06,570 S1: it was like, how are you going to send those? 492 00:28:06,570 --> 00:28:09,119 S1: I thought we had export controls, but these are within 493 00:28:09,119 --> 00:28:11,850 S1: the export controls, so they must be like pretty nerfed. 494 00:28:11,850 --> 00:28:15,090 S1: Apple might soon announce a deal to bring Google Gemini 495 00:28:15,090 --> 00:28:18,899 S1: to iOS 18, so that means they're working with all three. 496 00:28:18,900 --> 00:28:22,500 S1: They already announced OpenAI in the announcement. They mentioned meta 497 00:28:22,530 --> 00:28:25,109 S1: a while back and now Google. And by the way, 498 00:28:25,109 --> 00:28:30,570 S1: I'm bothered with Apple and also with OpenAI because they 499 00:28:30,570 --> 00:28:33,850 S1: both did demos of things that aren't in the betas. 500 00:28:33,850 --> 00:28:37,119 S1: Some reporting is saying this Siri stuff, and all the 501 00:28:37,119 --> 00:28:39,880 S1: AI stuff won't even be until next spring. And I'm like, 502 00:28:39,880 --> 00:28:41,950 S1: stop showing me demos of stuff that won't be out 503 00:28:41,950 --> 00:28:45,790 S1: for months, if ever. But that's just me being annoyed. 504 00:28:45,790 --> 00:28:49,480 S1: Greece is moving to a six day working week. Weird. Oh, 505 00:28:49,480 --> 00:28:52,480 S1: this is actually for a different story. That's a mistake here. 506 00:28:53,020 --> 00:28:58,080 S1: Do not like, um, this was about somebody, uh, Apple 507 00:28:58,080 --> 00:29:03,210 S1: putting cameras on AirPods, which I was hoping was going 508 00:29:03,210 --> 00:29:05,970 S1: to be actual cameras. But of course, that's going to 509 00:29:05,970 --> 00:29:09,630 S1: be more headgear that's going to require like, uh, like 510 00:29:09,630 --> 00:29:12,570 S1: a collar or like a neck piece or like a 511 00:29:12,570 --> 00:29:16,860 S1: heavy earphone or headphone because you need to power this, uh, 512 00:29:16,860 --> 00:29:19,260 S1: this camera that's running all the time. So you're going 513 00:29:19,260 --> 00:29:22,490 S1: to have battery issues, plus lenses and stuff like that. 514 00:29:22,490 --> 00:29:25,430 S1: But what I really want is I want the AirPods 515 00:29:25,430 --> 00:29:28,190 S1: or this headset or whatever it is, but it's got 516 00:29:28,190 --> 00:29:29,750 S1: to be every day. But I want it to see 517 00:29:29,750 --> 00:29:32,750 S1: behind me. That's super critical. If you look at my 518 00:29:32,750 --> 00:29:34,940 S1: big eye piece, I want it to see in front 519 00:29:34,940 --> 00:29:37,310 S1: of me. I want it to see behind me. It'd 520 00:29:37,340 --> 00:29:39,380 S1: be nice if it saw to the sides and it 521 00:29:39,380 --> 00:29:42,740 S1: could connect them for 360. That would be nice. But 522 00:29:42,740 --> 00:29:46,390 S1: most importantly, in front and behind And then that all 523 00:29:46,390 --> 00:29:48,910 S1: being processed and it's like, hey, there's a dangerous thing 524 00:29:48,910 --> 00:29:50,800 S1: in front of you. Hey, watch out for this. This 525 00:29:50,800 --> 00:29:53,860 S1: car is coming in your direction. Whatever. Someone's sneaking up 526 00:29:53,860 --> 00:29:57,610 S1: behind you. Be great for women. Like walking at night. 527 00:29:57,730 --> 00:30:00,730 S1: It's like this person is following you. Yeah, but no, 528 00:30:00,730 --> 00:30:04,570 S1: the thing that actually is getting shipped, it's for gesture controls. 529 00:30:04,570 --> 00:30:07,000 S1: So you get like, you know, go like this and 530 00:30:07,000 --> 00:30:10,080 S1: you could do things with your AirPods. That's the rumor. 531 00:30:10,080 --> 00:30:12,840 S1: David Brooks sat down with Steve Bannon to understand his 532 00:30:12,840 --> 00:30:16,680 S1: vision for the global populist movement. Um, number of U.S. 533 00:30:16,680 --> 00:30:20,100 S1: high school graduates is expected to peak in 25 and 534 00:30:20,100 --> 00:30:23,670 S1: then decline for years. I don't understand this. Schools and 535 00:30:23,670 --> 00:30:26,700 S1: colleges are closing. Faculty members are being laid off. I, 536 00:30:26,910 --> 00:30:30,270 S1: I what I don't understand why there's going to be 537 00:30:30,270 --> 00:30:33,830 S1: fewer people graduating. Is there a few people going into 538 00:30:33,830 --> 00:30:37,400 S1: the system? Is the population declining? I'm not sure exactly 539 00:30:37,400 --> 00:30:39,860 S1: what's going on here. US dollar just hit a 38 540 00:30:39,860 --> 00:30:43,550 S1: year peak against the yen. Peoples whose eyes dilated more 541 00:30:43,550 --> 00:30:47,060 S1: performed better on tests of working memory, suggesting that pupil 542 00:30:47,060 --> 00:30:49,010 S1: size is linked to how well we can process and 543 00:30:49,010 --> 00:30:54,080 S1: remember information. And this is from Scientific American. The FDA 544 00:30:54,080 --> 00:30:58,900 S1: approved Eli Lilly's Alzheimer's drug Kazuna. Why was that hard 545 00:30:58,900 --> 00:31:04,090 S1: to pronounce? Kazuna. Donanemab is the real name which shows 546 00:31:04,090 --> 00:31:08,320 S1: disease progression, slows disease progression by about a third. These 547 00:31:08,320 --> 00:31:11,920 S1: drug companies are killing it with Ozempic and now Alzheimer's. 548 00:31:11,920 --> 00:31:15,370 S1: High work in progress is killing your business, innovation and morale. 549 00:31:15,400 --> 00:31:18,700 S1: The more tasks you juggle, the slower everything moves. Yeah, 550 00:31:18,700 --> 00:31:22,620 S1: tell me about it. I'm struggling with this right now. Ideas. 551 00:31:22,620 --> 00:31:26,310 S1: The I class gap. There's a separation that is really 552 00:31:26,310 --> 00:31:30,090 S1: freaking me out about AI right now, which is dramatically 553 00:31:30,090 --> 00:31:37,470 S1: increased separation between social groups or classes, socioeconomic classes because 554 00:31:37,470 --> 00:31:40,500 S1: of AI. So basically one small group at the top 555 00:31:40,500 --> 00:31:43,500 S1: of like 1% or whatever, how you want to call 556 00:31:43,500 --> 00:31:46,040 S1: it 5%, we'll use AI to have a staff of 557 00:31:46,040 --> 00:31:51,140 S1: 10,000 executive assistants, tutors and analysts working constantly to run 558 00:31:51,140 --> 00:31:55,670 S1: their businesses, optimize their entire lives, accountants doing their paperwork, 559 00:31:55,670 --> 00:31:59,870 S1: filing their taxes, optimizing everything. Just think about that. A 560 00:31:59,870 --> 00:32:02,600 S1: lot of people, this top 1%, are going to have 561 00:32:02,600 --> 00:32:06,780 S1: like a team of 10 or 50 or 100 or 562 00:32:06,780 --> 00:32:10,959 S1: 10,000 executive assistants, which are basically independent eyes, all working 563 00:32:10,960 --> 00:32:14,890 S1: towards the same goals, optimizing everything in your life. That's 564 00:32:14,890 --> 00:32:17,590 S1: the top 1%. And a lot of other people either 565 00:32:17,590 --> 00:32:19,900 S1: won't be using AI at all, or they'll only be 566 00:32:19,900 --> 00:32:25,150 S1: using it for gaming, watching media porn entertainment, watching videos, 567 00:32:25,150 --> 00:32:31,270 S1: watching Netflix, just wasting time like slightly more efficiently using AI. 568 00:32:31,270 --> 00:32:34,150 S1: And this is largely the same split as between like 569 00:32:34,150 --> 00:32:37,660 S1: voracious readers and everyone else today. But it's going to 570 00:32:37,660 --> 00:32:40,780 S1: be way worse with AI, because AI is going to 571 00:32:40,780 --> 00:32:44,680 S1: lift the top 1% way more. Even more than being 572 00:32:44,680 --> 00:32:48,970 S1: a heavy reader. Discovery self publishing a tech book mma AI. 573 00:32:48,970 --> 00:32:52,960 S1: This guy is predicting MMA fight outcomes with very high 574 00:32:52,960 --> 00:32:57,110 S1: accuracy just using AI, and all he's doing is he's 575 00:32:57,110 --> 00:32:59,540 S1: looking at the, um, the people that they fought in 576 00:32:59,540 --> 00:33:03,560 S1: the past. Sam Parr launched Sam's List, a database of CPAs, 577 00:33:03,560 --> 00:33:07,670 S1: accountants and tax strategists, and he's already made $32,000 in 578 00:33:07,670 --> 00:33:11,510 S1: just two months. And illustrated transformer visual intuitive guide to 579 00:33:11,510 --> 00:33:15,770 S1: understanding how Transformers work. 11 labs has a new voice isolator. 580 00:33:15,770 --> 00:33:21,110 S1: This thing's scary. Rachel Toback, my buddy, talked about this, uh, talking. 581 00:33:21,110 --> 00:33:25,060 S1: I forgot how to pronounce this. Um, talking. Yeah, I 582 00:33:25,060 --> 00:33:27,190 S1: feel like I should know how to pronounce that. Anyway, 583 00:33:27,520 --> 00:33:31,420 S1: Ursula K Le Guin did a translation of this, which 584 00:33:31,420 --> 00:33:34,300 S1: I've not read this thing yet. I really need to 585 00:33:34,300 --> 00:33:36,520 S1: read this thing in full. And the recommendation of the 586 00:33:36,520 --> 00:33:41,080 S1: week is ground news. Ground news. This is like the 587 00:33:41,080 --> 00:33:45,660 S1: best website right now for news, especially going into this, uh, 588 00:33:45,660 --> 00:33:52,470 S1: political election series, uh, situation in the US. So it 589 00:33:52,470 --> 00:33:57,060 S1: shows you bias in how stories are covered. In fact, 590 00:33:57,060 --> 00:33:59,430 S1: let me just pull this up. It's the whole purpose 591 00:33:59,430 --> 00:34:02,820 S1: of being on, uh, on video. Look at this thing. Okay. 592 00:34:02,820 --> 00:34:07,530 S1: I'm logging in, maybe and doxing myself while I log in. 593 00:34:07,530 --> 00:34:11,089 S1: That's cool. Okay, look at these. You see these outlines? 594 00:34:11,090 --> 00:34:14,480 S1: It's got an outline for the type of coverage that 595 00:34:14,480 --> 00:34:19,940 S1: it's mostly getting. Look at this bar. Left 17% center 16%. 596 00:34:19,940 --> 00:34:24,140 S1: Right 67% okay. And you have these bars all over 597 00:34:24,140 --> 00:34:28,640 S1: the place. And what's really cool blind spot. Look at 598 00:34:28,640 --> 00:34:32,360 S1: this blind spot thing for the left. This is exposing 599 00:34:32,360 --> 00:34:34,960 S1: you to things that the right is talking about, that 600 00:34:34,960 --> 00:34:37,509 S1: the left is not. And same for over here, for 601 00:34:37,510 --> 00:34:40,299 S1: the right exposing you to things that the left is 602 00:34:40,300 --> 00:34:43,270 S1: talking about, that the right is not. So this this 603 00:34:43,270 --> 00:34:46,900 S1: helps you basically balance yourself, or at least have more 604 00:34:46,900 --> 00:34:51,009 S1: of an option of balancing yourself. Highly recommended. Especially if 605 00:34:51,010 --> 00:34:53,740 S1: you know somebody who claims to be wanting to be 606 00:34:53,739 --> 00:34:56,200 S1: less biased. And they're like, well, I'm just looking for 607 00:34:56,200 --> 00:34:59,759 S1: a good source or whatever. Send them this. Okay. And 608 00:34:59,760 --> 00:35:02,730 S1: the aphorism of the week, clear writing and clear speaking 609 00:35:02,730 --> 00:35:07,320 S1: are simply outputs of clear thinking. Naval clear writing and 610 00:35:07,320 --> 00:35:13,650 S1: clear speaking are simply outputs of clear thinking. Naval Unsupervised 611 00:35:13,650 --> 00:35:16,170 S1: Learning is produced and edited by Daniel Miller on a 612 00:35:16,170 --> 00:35:21,029 S1: Neumann U87 AI microphone using Hindenburg. Intro and outro music 613 00:35:21,030 --> 00:35:24,110 S1: is by Zomby with the Y, and to get the 614 00:35:24,110 --> 00:35:26,180 S1: text and links from this episode, sign up for the 615 00:35:26,180 --> 00:35:31,820 S1: newsletter version of the show at Daniel miessler.com/newsletter. We'll see 616 00:35:31,820 --> 00:35:32,569 S1: you next time.