1 00:00:00,200 --> 00:00:03,320 S1: All right. Welcome to unsupervised learning. It's Daniel Meisler. All right. 2 00:00:03,320 --> 00:00:05,750 S1: What do we got here? All right. So got a 3 00:00:05,750 --> 00:00:09,500 S1: new definition of prompt engineering that I'm basically using. Now 4 00:00:09,500 --> 00:00:12,110 S1: when I talk to people. And I can share that 5 00:00:12,110 --> 00:00:15,380 S1: a little bit later. AI's impact on the job market. 6 00:00:15,380 --> 00:00:17,960 S1: Working on a number of new talks. I didn't think 7 00:00:17,960 --> 00:00:19,369 S1: I was going to do this. I thought I was 8 00:00:19,370 --> 00:00:21,020 S1: actually going to be like 2 or 3 for the 9 00:00:21,020 --> 00:00:24,080 S1: whole year, but it's like turning into a lot more. 10 00:00:24,079 --> 00:00:26,840 S1: So curious to see where that's going to end up. 11 00:00:26,840 --> 00:00:29,030 S1: It's going to end up being like 4 or 5 talks. 12 00:00:29,030 --> 00:00:32,480 S1: What I like about this system that I'm using, though, 13 00:00:32,479 --> 00:00:35,270 S1: is that I'm only speaking about stuff that I normally 14 00:00:35,270 --> 00:00:38,870 S1: talk about anyway, and that's what essentially what the newsletter 15 00:00:38,870 --> 00:00:41,390 S1: is and what the podcast is. It's like stuff I'm 16 00:00:41,390 --> 00:00:44,480 S1: already talking about and it's the same for talks, it's 17 00:00:44,479 --> 00:00:48,320 S1: the same for presentations, it's the same for blogs. It's 18 00:00:48,320 --> 00:00:51,440 S1: the same for tweets. Like I'm not doing a strategy 19 00:00:51,440 --> 00:00:57,380 S1: that relates to particular media or syndication destinations. I'm not 20 00:00:57,380 --> 00:01:00,290 S1: thinking at all about that, actually. And I've talked about 21 00:01:00,290 --> 00:01:04,190 S1: this multiple times before, but people keep bringing it up, like, 22 00:01:04,190 --> 00:01:06,289 S1: how do you come up with ideas or whatever? First 23 00:01:06,290 --> 00:01:08,060 S1: of all, the way I come up with ideas is 24 00:01:08,060 --> 00:01:11,000 S1: I read a crap ton of stuff, like all the time. 25 00:01:11,000 --> 00:01:14,330 S1: I'm always reading mostly books, but also articles and stuff 26 00:01:14,330 --> 00:01:18,589 S1: like that, and social media and other builders and creators 27 00:01:18,590 --> 00:01:21,589 S1: in AI and security and stuff like that. So that 28 00:01:21,590 --> 00:01:24,020 S1: is the main way that I get ideas is by 29 00:01:24,020 --> 00:01:27,200 S1: having tons and tons of inputs. And the metaphor that 30 00:01:27,200 --> 00:01:29,720 S1: I use for this before was like a particle accelerator, 31 00:01:29,720 --> 00:01:32,750 S1: like I'm just like the atom, the tiny little brain 32 00:01:32,750 --> 00:01:35,990 S1: in the middle. But I'm being bombarded by the coolest 33 00:01:35,990 --> 00:01:41,780 S1: ideas from like Greek classics to like ancient philosophy productivity 34 00:01:41,780 --> 00:01:46,220 S1: books to AI stuff to whatever. And when that stuff 35 00:01:46,220 --> 00:01:49,070 S1: hits my brain. If you've ever seen those outputs from 36 00:01:49,070 --> 00:01:54,290 S1: one of these particle accelerators, it's like this super colorful, amazing, 37 00:01:54,290 --> 00:01:57,980 S1: like explosion of things that come out of it. And 38 00:01:57,980 --> 00:02:00,890 S1: those are all the subatomic particles. And what's interesting about 39 00:02:00,890 --> 00:02:04,309 S1: this analogy is those only last for this very small 40 00:02:04,310 --> 00:02:07,220 S1: amount of time. And that's why I'm so obsessed with capturing, 41 00:02:07,220 --> 00:02:10,370 S1: because if I'm having a conversation or I'm out for 42 00:02:10,370 --> 00:02:12,680 S1: a walk and it's funny because going out for a 43 00:02:12,680 --> 00:02:15,770 S1: walk or taking a shower with no tech is very 44 00:02:15,770 --> 00:02:20,180 S1: similar to a particle accelerator, where you're actually capturing ideas 45 00:02:20,180 --> 00:02:22,820 S1: from really smart things, because what happens is your brain 46 00:02:22,820 --> 00:02:25,250 S1: will start freaking out and it's like, hey, let's think 47 00:02:25,250 --> 00:02:27,320 S1: about some stuff. Hey, what about that one time where 48 00:02:27,320 --> 00:02:30,440 S1: you read that one thing and it starts just churning 49 00:02:30,440 --> 00:02:32,690 S1: and coming up with cool ideas, which is why you 50 00:02:32,690 --> 00:02:35,450 S1: have ideas in the shower, because your brain is freaking 51 00:02:35,450 --> 00:02:37,700 S1: out because it doesn't have any input. So it starts 52 00:02:37,700 --> 00:02:40,760 S1: like just generating things. I wonder if we're doing the 53 00:02:40,760 --> 00:02:43,430 S1: same things in dreams or whatever. I'm not sure. Maybe 54 00:02:43,430 --> 00:02:46,760 S1: a similar mechanism, but the point is, I have ideas 55 00:02:46,760 --> 00:02:50,150 S1: because I consume a lot and the ideas don't even 56 00:02:50,150 --> 00:02:53,240 S1: necessarily have anything to do with what I'm consuming. They 57 00:02:53,240 --> 00:02:55,820 S1: tend to be peripheral to what I'm consuming, and that's 58 00:02:55,820 --> 00:02:59,690 S1: why I read deeply but broadly. Right? Because I'll switch 59 00:02:59,690 --> 00:03:02,780 S1: to fiction and there'll be a cool idea in the fiction, 60 00:03:02,780 --> 00:03:05,570 S1: and then suddenly the fiction is the peanut butter. And 61 00:03:05,570 --> 00:03:09,200 S1: this nonfiction AI thing that I'm working on or coding 62 00:03:09,200 --> 00:03:12,470 S1: on is like the chocolate, and they mix together and 63 00:03:12,470 --> 00:03:14,299 S1: all of a sudden I have a cool idea. And 64 00:03:14,300 --> 00:03:16,940 S1: this is why I really believe that this whole idea 65 00:03:16,940 --> 00:03:19,910 S1: of like, oh, people are super smart or people are 66 00:03:19,910 --> 00:03:22,370 S1: super genius or whatever, look at this person who has 67 00:03:22,370 --> 00:03:26,300 S1: all the ideas. Nope. The ideas come from inputs. If 68 00:03:26,300 --> 00:03:28,490 S1: you stop reading and I know this for a fact 69 00:03:28,490 --> 00:03:31,910 S1: for myself, if I stop reading within a few months, 70 00:03:31,910 --> 00:03:35,210 S1: even a couple of weeks, like my brain will just 71 00:03:35,210 --> 00:03:38,420 S1: slow down. Like I will not have the same number 72 00:03:38,420 --> 00:03:41,480 S1: or quality of ideas coming in. And this is why 73 00:03:41,510 --> 00:03:44,600 S1: people like Buffett and Charlie Munger and all these people, 74 00:03:44,600 --> 00:03:47,029 S1: they're like, look, if you want to be smart, you 75 00:03:47,030 --> 00:03:49,850 S1: need to read. You need to be consuming really high 76 00:03:49,850 --> 00:03:52,970 S1: quality stuff at a very high rate of speed. And 77 00:03:52,970 --> 00:03:55,160 S1: that's not always the case with like the rate you 78 00:03:55,160 --> 00:03:58,340 S1: can actually get really high quality ideas. Riva has talked 79 00:03:58,340 --> 00:04:02,390 S1: about this where like slow reading, a deep book like 80 00:04:02,390 --> 00:04:07,340 S1: the Bible, or like a classic from the Greeks or something, whatever. 81 00:04:07,340 --> 00:04:10,310 S1: You could slow read and try to memorize and like 82 00:04:10,310 --> 00:04:13,460 S1: go super deep with the thing, try to fully extract 83 00:04:13,460 --> 00:04:16,219 S1: all the wisdom from it and do that on a 84 00:04:16,220 --> 00:04:19,400 S1: regular basis. And you still get lots of ideas. But 85 00:04:19,400 --> 00:04:22,790 S1: what you can't do is do nothing and think nothing, 86 00:04:22,790 --> 00:04:26,239 S1: and consume nothing and just consume. You don't want to 87 00:04:26,240 --> 00:04:29,600 S1: just be like watching Netflix or something. Of course, there 88 00:04:29,600 --> 00:04:31,580 S1: could be some good stuff on Netflix where you actually 89 00:04:31,580 --> 00:04:34,460 S1: get some ideas, but in general, that is a thing 90 00:04:34,460 --> 00:04:37,850 S1: that calms your mind and like sedates it right? You're 91 00:04:37,850 --> 00:04:41,870 S1: not getting high quality content coming in. It's more like relaxation. 92 00:04:41,870 --> 00:04:46,550 S1: So all right, cool giant diatribe. And we've got through 93 00:04:46,550 --> 00:04:49,159 S1: one bullet in the damn newsletter. All right. Publish a 94 00:04:49,160 --> 00:04:53,930 S1: new official pattern template in fabric. So official Pattern Template 95 00:04:53,930 --> 00:04:57,020 S1: is the name of it. And it's basically all my 96 00:04:57,020 --> 00:05:01,200 S1: latest formats and instructions. And like the structure of the 97 00:05:01,230 --> 00:05:05,190 S1: of the pattern or a prompt really an I prompt 98 00:05:05,190 --> 00:05:07,830 S1: that works best for me. And I did a few 99 00:05:07,830 --> 00:05:11,010 S1: updates here. I've got a lot more, I would say 100 00:05:11,010 --> 00:05:15,210 S1: comments or descriptions of the sections. I broke out more sections. 101 00:05:15,210 --> 00:05:19,830 S1: I added a examples and negative examples section, and I 102 00:05:19,830 --> 00:05:22,290 S1: also broke out the identity and the goal. So a 103 00:05:22,290 --> 00:05:25,560 S1: number of upgrades in there and people have been asking like, okay, 104 00:05:25,560 --> 00:05:28,170 S1: how do you make a perfect prompt? So I made this, 105 00:05:28,170 --> 00:05:30,780 S1: that basically shows that. And of course that's going to 106 00:05:30,779 --> 00:05:33,839 S1: be updated constantly. Right. Because this is like there's no 107 00:05:33,839 --> 00:05:37,170 S1: testing for these things. So everything I'm doing is like 108 00:05:37,170 --> 00:05:40,890 S1: it's like it's like placebo. If it works for me, 109 00:05:40,890 --> 00:05:42,810 S1: I'm like, yeah, this is really cool. I like it, 110 00:05:42,810 --> 00:05:45,270 S1: and I continue to do it because it works for me. 111 00:05:45,270 --> 00:05:47,729 S1: And when I test it against other things, I am 112 00:05:47,730 --> 00:05:52,380 S1: doing a B testing. But first of all, LMS are non-deterministic, 113 00:05:52,380 --> 00:05:55,440 S1: so I can get different answers for whatever it's like. 114 00:05:55,440 --> 00:05:57,270 S1: In order to do this properly, you need to run 115 00:05:57,270 --> 00:06:00,870 S1: it like tens of times, dozens of times, hundreds of times. 116 00:06:00,870 --> 00:06:02,700 S1: And of course you have to worry about cost here, 117 00:06:02,700 --> 00:06:05,130 S1: but then you have to have a really good objective 118 00:06:05,130 --> 00:06:08,520 S1: rating system that's looking at those and actually finding the 119 00:06:08,520 --> 00:06:11,790 S1: differences between these prompt techniques. And a bunch of papers 120 00:06:11,790 --> 00:06:14,370 S1: have tried to do this. I have not necessarily agreed 121 00:06:14,370 --> 00:06:18,090 S1: with the results of those papers. And again, that could 122 00:06:18,089 --> 00:06:20,250 S1: be because the whole thing is a moving target in 123 00:06:20,250 --> 00:06:23,280 S1: the first place, right? Bottom line is this what works 124 00:06:23,279 --> 00:06:26,490 S1: for me? It is the official pattern template in fabric 125 00:06:26,490 --> 00:06:29,910 S1: right now. Did my buddy Jason Haddix red blue, purple 126 00:06:29,910 --> 00:06:33,030 S1: I class last week. It was fantastic. It was Monday, 127 00:06:33,240 --> 00:06:36,450 S1: no Thursday and Friday of last week. It was so good. 128 00:06:36,450 --> 00:06:39,630 S1: Jason is definitely the best teacher that I know and 129 00:06:39,630 --> 00:06:42,330 S1: it was like, you're learning stuff. But he's also just 130 00:06:42,330 --> 00:06:46,560 S1: hanging out. He's talking or we're talking all his friends. 131 00:06:46,560 --> 00:06:49,919 S1: Different people like myself were also contributing to the class 132 00:06:49,920 --> 00:06:52,440 S1: a little bit. He would ask a question, we would 133 00:06:52,440 --> 00:06:55,500 S1: give an answer. Jason would chime in on that. And 134 00:06:55,500 --> 00:06:58,080 S1: the people taking the class were also super smart. So 135 00:06:58,080 --> 00:07:00,419 S1: they were contributing lots of stuff, and it was just 136 00:07:00,420 --> 00:07:03,599 S1: like it was just great. It was absolutely great. Both 137 00:07:03,600 --> 00:07:06,089 S1: days of content really enjoyed, and the next time it 138 00:07:06,089 --> 00:07:08,669 S1: comes out, you should definitely sign up. So I got 139 00:07:08,670 --> 00:07:12,870 S1: a new sponsor conversation here with Abhishek from Material Security. 140 00:07:12,900 --> 00:07:15,450 S1: He's the CEO over there and we talked about a 141 00:07:15,450 --> 00:07:18,720 S1: bunch of cool stuff. It was really a great conversation. 142 00:07:18,720 --> 00:07:21,420 S1: One of my favorite topics was like, why are product 143 00:07:21,420 --> 00:07:24,210 S1: managers so good at being CEOs? And so he gave 144 00:07:24,210 --> 00:07:27,210 S1: a great answer to that. I got another piece up 145 00:07:27,210 --> 00:07:30,540 S1: here on getting a perfect sound from your microphone. And 146 00:07:30,540 --> 00:07:33,600 S1: this sound that you're currently listening to is a good 147 00:07:33,600 --> 00:07:36,780 S1: version of that, although I don't do post-production for this 148 00:07:36,780 --> 00:07:40,560 S1: particular thing because this is an OBS video going directly 149 00:07:40,560 --> 00:07:44,070 S1: into YouTube. So I'm not doing post-production on here, but 150 00:07:44,070 --> 00:07:47,580 S1: it should be a pretty decent initial sound. I got 151 00:07:47,580 --> 00:07:50,490 S1: a political one here. It's called The left's. It's political, 152 00:07:50,490 --> 00:07:52,770 S1: but it's centrist, so I don't think it should bother 153 00:07:52,770 --> 00:07:55,740 S1: you unless you're crazy left or crazy right? But you 154 00:07:55,740 --> 00:07:58,350 S1: should check that out. Lots of people from the left 155 00:07:58,350 --> 00:08:00,630 S1: and the right have reached out and said that they 156 00:08:00,630 --> 00:08:04,140 S1: loved it. So that should be a good metric for you. Yeah, 157 00:08:04,140 --> 00:08:08,580 S1: it's called Left's Brexit security got a bad checkpoint. Von 158 00:08:08,580 --> 00:08:14,340 S1: massive botnet was busted up. Ukraine is using AI against jamming. 159 00:08:14,340 --> 00:08:19,800 S1: So yeah, it's basically like kill decision from Daniel Suarez where, um, 160 00:08:19,800 --> 00:08:24,720 S1: the drones were getting jammed, GPS jammed, and also just 161 00:08:24,720 --> 00:08:28,800 S1: any RF was getting jammed. So they made them autonomous 162 00:08:28,800 --> 00:08:33,600 S1: so they could just navigate by vision and looking at landmarks. 163 00:08:33,600 --> 00:08:36,540 S1: And then the idea was you have a drone swarm. 164 00:08:36,540 --> 00:08:39,780 S1: And this is from the book and probably also from reality, 165 00:08:39,780 --> 00:08:42,660 S1: also brought to you by the Pentagon, a drone swarm 166 00:08:42,660 --> 00:08:45,870 S1: with tiny little explosives on them. And they would just 167 00:08:45,870 --> 00:08:48,750 S1: all fly and navigate. They would have a target that 168 00:08:48,750 --> 00:08:51,570 S1: they were trying to kill, and they would navigate by 169 00:08:51,570 --> 00:08:54,420 S1: the landmarks. And when they find the person, they could 170 00:08:54,420 --> 00:08:57,150 S1: do facial recognition on board, don't need to call out 171 00:08:57,150 --> 00:09:01,980 S1: anywhere and then just crash into them, kamikaze and kill 172 00:09:01,980 --> 00:09:05,550 S1: the person. And jammers don't matter because they're not using 173 00:09:05,550 --> 00:09:10,890 S1: navigation based on RF smart home tech in warfare. So 174 00:09:10,890 --> 00:09:14,550 S1: Home Assistant is being used in Ukraine to warn people 175 00:09:14,550 --> 00:09:19,260 S1: of alerts of attacks incoming. Yeah. Incogni is another data 176 00:09:19,260 --> 00:09:22,560 S1: deletion service similar to delete me moves your info from 177 00:09:22,559 --> 00:09:26,339 S1: over 170 data brokers. They are not a sponsor. Otherwise 178 00:09:26,340 --> 00:09:29,370 S1: you would say sponsor on the top of it Ticketmaster. 179 00:09:29,370 --> 00:09:32,250 S1: So they got hit by a group called Shiny Hunters, 180 00:09:32,250 --> 00:09:36,300 S1: and it looks like around 560 million people's data has 181 00:09:36,300 --> 00:09:39,840 S1: been stolen. 560 million. I'm so glad they were getting 182 00:09:39,840 --> 00:09:42,270 S1: broken up, or at least we're trying to break them up. 183 00:09:42,270 --> 00:09:45,510 S1: And so glad that Taylor made that happen. I think 184 00:09:45,510 --> 00:09:49,320 S1: she did. A few people spread the most information about 185 00:09:49,320 --> 00:09:52,470 S1: fake news about vaccines and stuff like that, going, this 186 00:09:52,470 --> 00:09:57,449 S1: is back to like 2020 and 2021, but, uh, older 187 00:09:57,450 --> 00:10:03,210 S1: Republican women. 80% of the misinformation. Yeah, I wrote this myself. Small, 188 00:10:03,210 --> 00:10:07,230 S1: persistent groups can have a significant impact on propaganda. You 189 00:10:07,230 --> 00:10:09,959 S1: wouldn't think that. You would think that small groups like 190 00:10:09,960 --> 00:10:12,360 S1: that would just be like a glass of water in 191 00:10:12,360 --> 00:10:15,900 S1: the ocean, but turns out if you're persistent enough, it'll 192 00:10:15,900 --> 00:10:18,090 S1: get out there. NSA released a guide on how to 193 00:10:18,090 --> 00:10:20,460 S1: keep your phone from getting hacked, and they basically said 194 00:10:20,460 --> 00:10:23,640 S1: you should reboot your phone once a week. Frightening. What 195 00:10:23,640 --> 00:10:25,949 S1: do they know that I don't know? Or they're like, yeah, 196 00:10:25,950 --> 00:10:28,830 S1: you should reboot your phone every week. And I'm like, why? 197 00:10:28,860 --> 00:10:32,730 S1: Never mind. Don't want to know technology. All right. So grok, 198 00:10:32,730 --> 00:10:35,699 S1: I talked about this. How fast it is wickedly fast 199 00:10:35,700 --> 00:10:39,420 S1: this thing okay. So it can now do be lama 200 00:10:39,420 --> 00:10:43,950 S1: 3 to 1200 tokens per second. That's essentially like it 201 00:10:43,950 --> 00:10:46,920 S1: just fills out the page as if it just instant. 202 00:10:46,920 --> 00:10:49,860 S1: And so here's my question. Why doesn't Google or OpenAI 203 00:10:49,860 --> 00:10:53,130 S1: or Microsoft or someone just like, walk up like, you know, 204 00:10:53,130 --> 00:10:56,760 S1: a clearinghouse sweepstakes or something? And just like, here's a 205 00:10:56,760 --> 00:11:02,760 S1: box of $1 billion, please give us your technology and disappear. Right. 206 00:11:02,820 --> 00:11:07,050 S1: This is like Magic Beans proprietary chip tech that lets 207 00:11:07,050 --> 00:11:11,820 S1: him basically run AI. What, tens of times faster than 208 00:11:11,820 --> 00:11:15,780 S1: anyone else? Ridiculous. Looks like Satya Nadella is freaking out 209 00:11:15,780 --> 00:11:19,560 S1: about a potential partnership between OpenAI and Apple. I think 210 00:11:19,559 --> 00:11:21,420 S1: he's right to freak out, and he should make a 211 00:11:21,420 --> 00:11:25,050 S1: phone before it's too late. Or a personal assistant on 212 00:11:25,050 --> 00:11:28,949 S1: some kind of device. He needs a personal device. Anyone 213 00:11:28,950 --> 00:11:32,190 S1: who wants to play going forward need a personal device 214 00:11:32,190 --> 00:11:34,050 S1: and an OS, and it needs to be able to 215 00:11:34,050 --> 00:11:37,860 S1: see and hear and basically be integrated with all your context. 216 00:11:37,860 --> 00:11:40,650 S1: And anyone who's not doing this is screwed. And people 217 00:11:40,650 --> 00:11:43,709 S1: have a massive advantage. Are people who already have access 218 00:11:43,710 --> 00:11:50,010 S1: into a robust operating system, namely Android and Apple. Researchers 219 00:11:50,010 --> 00:11:55,949 S1: have new attention mechanisms that actually outperform standard multi-head attention. Cool. Well, 220 00:11:55,950 --> 00:11:59,490 S1: let me know when it fixes haystack performance. Need I 221 00:11:59,490 --> 00:12:02,070 S1: need haystack performance. I need to be able to give 222 00:12:02,070 --> 00:12:05,969 S1: you giant text documents and you don't miss a single thing. 223 00:12:05,970 --> 00:12:09,000 S1: So wake me up when that happens, all right? Recall 224 00:12:09,000 --> 00:12:12,810 S1: a privacy disaster. This is the Microsoft thing that records everything. 225 00:12:12,809 --> 00:12:15,780 S1: Stealing everything. This is what I wrote. Stealing everything you've 226 00:12:15,780 --> 00:12:17,940 S1: ever typed or viewed on your windows PC is now 227 00:12:17,940 --> 00:12:21,720 S1: a reality, thanks to a feature in Copilot plus recall 228 00:12:21,720 --> 00:12:25,110 S1: that is a privacy nightmare. Yeah, disagree with that take. 229 00:12:25,350 --> 00:12:28,110 S1: I think in ten years almost everyone will think computers 230 00:12:28,110 --> 00:12:31,080 S1: who don't have this are basically worthless. And in fact, 231 00:12:31,080 --> 00:12:35,040 S1: ten years most people won't be talking, will only be talking, 232 00:12:35,040 --> 00:12:38,579 S1: will mostly be talking and gesturing to their computers. Computer 233 00:12:38,580 --> 00:12:41,880 S1: will mostly be like monitors and HUD displays or whatever, 234 00:12:41,880 --> 00:12:45,300 S1: and your computers are basically doing the work. So the 235 00:12:45,300 --> 00:12:48,059 S1: idea that it's going to forget something you worked on 236 00:12:48,059 --> 00:12:50,190 S1: or forget something that was said to you, or that 237 00:12:50,190 --> 00:12:53,819 S1: you said to somebody else is completely asinine. Of course 238 00:12:53,820 --> 00:12:56,520 S1: your AI is going to remember everything. What else would 239 00:12:56,520 --> 00:13:00,480 S1: it do? All right, Llms aren't just internet simulators anymore. Yeah, 240 00:13:00,480 --> 00:13:05,010 S1: so increasingly llms are being trained on custom non-internet data 241 00:13:05,370 --> 00:13:08,070 S1: that's there's still a data wall there as well. But 242 00:13:08,070 --> 00:13:09,689 S1: I don't know, I think a lot of people like 243 00:13:09,690 --> 00:13:13,650 S1: social media companies. Meta is in great position for this, 244 00:13:13,650 --> 00:13:16,319 S1: for example. I mean, because they have Instagram, they have 245 00:13:16,320 --> 00:13:19,080 S1: all this thing. So basically whatever people are posting and 246 00:13:19,080 --> 00:13:22,980 S1: doing and talking about, that becomes a massive source of data. 247 00:13:22,980 --> 00:13:25,980 S1: Anyone who's like really close to the OS and especially 248 00:13:25,980 --> 00:13:29,460 S1: in messaging apps, social media apps, that is a massive 249 00:13:29,460 --> 00:13:32,309 S1: source of data that they can then sell to other 250 00:13:32,309 --> 00:13:35,250 S1: AI companies or keep it for themselves to have an advantage. 251 00:13:35,250 --> 00:13:38,219 S1: In the case of like Lama3, Nvidia is getting ready 252 00:13:38,220 --> 00:13:43,080 S1: to overtake Apple in market cap. That's insane. Completely insane. 253 00:13:43,080 --> 00:13:45,210 S1: But here's a way to make sense of this. Imagine 254 00:13:45,210 --> 00:13:48,180 S1: that there's a giant filter on how many people are 255 00:13:48,179 --> 00:13:51,270 S1: capable of creating a hit movie, writing a book, starting 256 00:13:51,270 --> 00:13:53,880 S1: a business, sharing their art with billions of people. In 257 00:13:53,880 --> 00:13:56,790 S1: other words, imagine the current number of major builders or 258 00:13:56,790 --> 00:14:06,210 S1: creators is something like 0.00000000000 17 of our 8 billion people. 259 00:14:06,210 --> 00:14:08,400 S1: I just made up that number. It's not a real number, 260 00:14:08,400 --> 00:14:11,130 S1: despite how precise it is. So just work with me. 261 00:14:11,130 --> 00:14:14,040 S1: So the reason Nvidia is rising so fast is because 262 00:14:14,040 --> 00:14:18,240 S1: one of the primary pieces of tech is the GPU. Okay? 263 00:14:18,240 --> 00:14:21,870 S1: It is like center mass of this entire thing. And 264 00:14:21,870 --> 00:14:25,260 S1: what we're talking about is removing 3 to 5 zeros 265 00:14:25,260 --> 00:14:29,100 S1: from this big zero number. We're talking about multiplying humanity's 266 00:14:29,100 --> 00:14:32,790 S1: creativity output by thousands in the next decade. And that 267 00:14:32,790 --> 00:14:36,000 S1: requires chips. And thank you for coming to my Ted 268 00:14:36,030 --> 00:14:39,990 S1: talk about Nvidia. That's essentially why they are skyrocketing. Google 269 00:14:39,990 --> 00:14:44,580 S1: released an SRE handbook. Basically saying simplicity is a core 270 00:14:44,580 --> 00:14:48,060 S1: principle for reliability makes total sense to me. They should 271 00:14:48,060 --> 00:14:50,850 S1: make one for product management. That's mean. But it's funny. 272 00:14:50,850 --> 00:14:56,970 S1: All right. Karpathy's GPT two 124 million parameters looks like 273 00:14:57,000 --> 00:15:02,280 S1: LMK and 90 minutes for. $20. Karpathy is an absolute beast. 274 00:15:02,310 --> 00:15:06,960 S1: Tiny Grad's latest update. This is from Geohot. Slashes Python 275 00:15:06,960 --> 00:15:11,970 S1: time and ditches external dependencies. Terminal animations. These look really cool. 276 00:15:11,970 --> 00:15:13,290 S1: I'm not sure I'm going to use them, but they 277 00:15:13,290 --> 00:15:16,920 S1: look really cool. And humans little known intelligence bureau that 278 00:15:16,920 --> 00:15:22,140 S1: nailed the outcomes in Vietnam, Iraq and Ukraine. I want 279 00:15:22,140 --> 00:15:26,220 S1: to build something that basically programmatically goes and gets these 280 00:15:26,220 --> 00:15:30,090 S1: predictions and incorporates them into my life model and how 281 00:15:30,090 --> 00:15:33,330 S1: I see the world, because it's essentially like these prediction 282 00:15:33,330 --> 00:15:37,680 S1: markets are the smartest people doing the best kind of predictions, 283 00:15:37,680 --> 00:15:42,510 S1: and they're doing it in a very transparent and logical way. 284 00:15:42,510 --> 00:15:45,690 S1: And I learned about this from the book called Superforecasters. 285 00:15:45,690 --> 00:15:48,060 S1: Is that Tetlock? Bill Tetlock, I think, might be the 286 00:15:48,060 --> 00:15:51,750 S1: name of the author or one of the authors. Unbelievable stuff. Really, 287 00:15:51,750 --> 00:15:55,170 S1: really good stuff. New study says sleep inequality in the 288 00:15:55,170 --> 00:15:58,350 S1: US is tied to economic stress. Makes sense to me. 289 00:15:58,350 --> 00:16:02,310 S1: Mice with PTSD like behavior showed significant improvement after having 290 00:16:02,310 --> 00:16:07,500 S1: access to a running wheel. Basically, exercise induced neurogenesis like 291 00:16:07,500 --> 00:16:10,590 S1: healing trauma with exercise. That was a bit of a stretch, 292 00:16:10,590 --> 00:16:13,650 S1: but that's where they're going with this Ozempic and wegovy 293 00:16:13,650 --> 00:16:17,940 S1: changing eating habits, fine tuning your taste buds, making sweets 294 00:16:17,940 --> 00:16:20,790 S1: taste sweeter. Mm. Yeah. I mean, it's definitely having an 295 00:16:20,790 --> 00:16:23,850 S1: effect on me. Like I had pizza. I've been having 296 00:16:23,850 --> 00:16:27,990 S1: pizza because of watching this damn pizza influencer who samples 297 00:16:27,990 --> 00:16:30,540 S1: pizza and is just making me want to eat pizza. Anyway, 298 00:16:30,540 --> 00:16:32,850 S1: I get a pizza, I eat, like three slices and 299 00:16:32,850 --> 00:16:34,950 S1: I'm just like, yeah, it was good, I guess, but 300 00:16:34,950 --> 00:16:38,250 S1: I'm full. Whereas before before I took Wegovy, I would 301 00:16:38,250 --> 00:16:40,500 S1: be like, I would eat three quarters of the pizza 302 00:16:40,500 --> 00:16:42,810 S1: and be like, I should not eat any more. And 303 00:16:42,810 --> 00:16:44,910 S1: then I would eat the rest and be like, what's 304 00:16:44,910 --> 00:16:49,380 S1: up with ice cream? John Carmack dives into bullshit jobs. Yeah, 305 00:16:49,380 --> 00:16:52,230 S1: really good book by David Graeber. Got to read that book. 306 00:16:52,230 --> 00:16:54,360 S1: I need to read it again. No, I don't I 307 00:16:54,360 --> 00:16:57,090 S1: should just summarize it. I have a fabric pattern for that. 308 00:16:57,090 --> 00:17:00,720 S1: Imagine navigating high school social maze without a smartphone, missing 309 00:17:00,720 --> 00:17:03,600 S1: out on chats and memes, beginning a unique perspective on 310 00:17:03,600 --> 00:17:06,630 S1: life and friendships, the social lives of teens who don't 311 00:17:06,630 --> 00:17:08,940 S1: have phones. I don't have to imagine this. I grew 312 00:17:08,940 --> 00:17:11,969 S1: up in the 80s. It was glorious. Yeah, very much 313 00:17:12,000 --> 00:17:17,010 S1: on the tip of Jonathan Haidt basically saying getting phones 314 00:17:17,010 --> 00:17:22,050 S1: out of schools highly support that effort. Comprehensive review finds 315 00:17:22,050 --> 00:17:26,880 S1: I exercises might actually help in preventing and controlling myopia. 316 00:17:26,880 --> 00:17:30,449 S1: This is actually an incorrect read. This was my fault. 317 00:17:30,450 --> 00:17:34,260 S1: This is an incorrect read. The study. There are a 318 00:17:34,260 --> 00:17:36,030 S1: whole bunch of studies saying it helps and a whole 319 00:17:36,030 --> 00:17:38,639 S1: bunch of studies saying it's not. But this particular study 320 00:17:38,640 --> 00:17:42,960 S1: actually says they couldn't find much evidence, strong evidence that 321 00:17:42,960 --> 00:17:45,959 S1: it did help. So that's that's boo on my part. 322 00:17:45,960 --> 00:17:49,440 S1: Lung cancer breakthrough unprotected. Oh yeah a lot of people 323 00:17:49,440 --> 00:17:51,750 S1: are talking about this. I don't get excited until it 324 00:17:51,750 --> 00:17:55,320 S1: comes to market and everyone loves it like Wegovy. But 325 00:17:55,320 --> 00:18:00,000 S1: lots of people are talking about how this thing is insane. 60% 326 00:18:00,000 --> 00:18:04,139 S1: of advanced stage patients progression free for five years. Marc 327 00:18:04,140 --> 00:18:06,959 S1: Andreessen says to stay in school. But then he says, 328 00:18:06,960 --> 00:18:09,390 S1: if you're the type of person to drop out, you 329 00:18:09,390 --> 00:18:12,930 S1: also won't listen to this advice. And I love this letter. 330 00:18:12,930 --> 00:18:15,660 S1: Love my Wife is dead letter from Richard Feynman to 331 00:18:15,660 --> 00:18:19,139 S1: his dead wife. Very touching ideas and analysis. My favorite 332 00:18:19,140 --> 00:18:22,020 S1: way to explain prompt engineering. Basically, don't think of it 333 00:18:22,020 --> 00:18:25,050 S1: like I think of it in this way. Instead, prompt 334 00:18:25,050 --> 00:18:28,409 S1: engineering is how to explain to a superior intelligence what 335 00:18:28,410 --> 00:18:30,810 S1: your problem is and what you would like to happen. 336 00:18:30,810 --> 00:18:33,930 S1: And if you can't do that well, you will lose 337 00:18:33,930 --> 00:18:36,480 S1: to people who can't. That's it. How to explain to 338 00:18:36,480 --> 00:18:39,419 S1: a superior intelligence what your problem is and what you 339 00:18:39,420 --> 00:18:42,690 S1: would like to happen. That's prompt engineering. All right. My 340 00:18:42,690 --> 00:18:45,270 S1: X thread on why it's so hard to find tech 341 00:18:45,270 --> 00:18:47,760 S1: jobs right now. All right. So this this is the 342 00:18:47,760 --> 00:18:50,369 S1: vibe here problem for the job market isn't that AI 343 00:18:50,369 --> 00:18:52,800 S1: is happening. The problem is that it's happening at the 344 00:18:52,800 --> 00:18:56,969 S1: exact same moment that most companies are figuring out that 80% 345 00:18:56,970 --> 00:19:00,030 S1: of their employees are worthless. We need to stop expecting 346 00:19:00,030 --> 00:19:02,190 S1: things to go back to the way they were. The 347 00:19:02,190 --> 00:19:07,050 S1: new reality is companies mostly hiring super ambitious, exceptional, proven 348 00:19:07,050 --> 00:19:10,710 S1: people who are total gods. With AI, this means most 349 00:19:10,710 --> 00:19:13,290 S1: formal education becomes a waste of time and money, because 350 00:19:13,290 --> 00:19:16,890 S1: a degree doesn't certify that you are super ambitious or 351 00:19:16,890 --> 00:19:20,460 S1: exceptional or proven, it doesn't certify any of those things. 352 00:19:20,460 --> 00:19:23,250 S1: So in this model, only elite schools will matter because 353 00:19:23,250 --> 00:19:26,910 S1: the filtering for being exceptional will have happened just by 354 00:19:26,910 --> 00:19:29,970 S1: being accepted into the school. That doesn't count for like 355 00:19:29,970 --> 00:19:33,960 S1: people who got in because their mom already graduated, right? 356 00:19:33,960 --> 00:19:36,959 S1: But for people who got in via merit to a 357 00:19:36,960 --> 00:19:39,570 S1: top school, that is a good filter. So really this 358 00:19:39,570 --> 00:19:42,510 S1: is two separate things. One, you're exceptional enough to be 359 00:19:42,510 --> 00:19:46,200 S1: accepted into the top school and two, you finish some classes. 360 00:19:46,200 --> 00:19:49,590 S1: Problem is, number two, you can get anywhere, right? The 361 00:19:49,590 --> 00:19:53,729 S1: actual training in school between one like a state school 362 00:19:53,730 --> 00:19:56,609 S1: and like Harvard is not that much different. Number one 363 00:19:56,609 --> 00:19:59,830 S1: is the thing that employers actually care. The filter to 364 00:19:59,830 --> 00:20:02,740 S1: be able to get into an elite institution. So the 365 00:20:02,740 --> 00:20:05,770 S1: result will be companies hiring the top 10 to 20% 366 00:20:05,770 --> 00:20:09,130 S1: of people in competence. And this will be filtered by 367 00:20:09,130 --> 00:20:13,120 S1: elite school attendance and proof of competence via something you've 368 00:20:13,119 --> 00:20:15,970 S1: put into the world, like on your website, YouTube as 369 00:20:15,970 --> 00:20:19,119 S1: a tool or a company you built, or an AI 370 00:20:19,119 --> 00:20:23,110 S1: algorithm rating you. This is going to be huge as well. 371 00:20:23,109 --> 00:20:26,620 S1: So the hyperbolic form of this is if you're not 372 00:20:26,619 --> 00:20:29,890 S1: 19 and at Harvard, or if you don't have your 373 00:20:29,890 --> 00:20:32,830 S1: own projects or companies you've built and talked about online, 374 00:20:32,830 --> 00:20:36,459 S1: you are not going to be interesting to employers. You 375 00:20:36,460 --> 00:20:40,030 S1: will be part of the 90% basically fighting for scraps 376 00:20:40,030 --> 00:20:43,090 S1: and the recommendation of the week for everyone you know 377 00:20:43,090 --> 00:20:45,550 S1: who's going to be thinking about their career in the future, 378 00:20:45,550 --> 00:20:49,179 S1: which is basically anyone who's not independently wealthy try to 379 00:20:49,180 --> 00:20:52,150 S1: get them to think about things in the following way. 380 00:20:52,180 --> 00:20:56,080 S1: AI is going to take out most executors of knowledge work. 381 00:20:56,080 --> 00:21:00,040 S1: The people who will survive are the people with ideas 382 00:21:00,040 --> 00:21:04,480 S1: who are actively building that thing, actively building that thing. 383 00:21:04,480 --> 00:21:07,090 S1: So it's not just ideas. So the safest thing to 384 00:21:07,090 --> 00:21:09,669 S1: be is a creator or a builder. That means a 385 00:21:09,670 --> 00:21:14,200 S1: founder or an entrepreneur, a programmer, somebody with ideas who 386 00:21:14,200 --> 00:21:18,010 S1: is productive and ambitious, like an artist like you could 387 00:21:18,010 --> 00:21:21,070 S1: be an artist. It doesn't have to be some hardcore engineer, 388 00:21:21,070 --> 00:21:23,709 S1: but it's got to be some kind of category like that. 389 00:21:23,710 --> 00:21:25,810 S1: And the trick is you have to be able to 390 00:21:25,810 --> 00:21:29,410 S1: have the ideas, be able to make it yourself, or 391 00:21:29,410 --> 00:21:32,560 S1: attract the talent and the AI and harness the AI 392 00:21:32,590 --> 00:21:34,330 S1: to be able to make it. And then you've got 393 00:21:34,330 --> 00:21:35,770 S1: to be able to market the hell out of it, 394 00:21:35,770 --> 00:21:38,229 S1: which is a lot of writing. So a lot of 395 00:21:38,230 --> 00:21:42,580 S1: like communication and persuasion. So bottom line, the winners in 396 00:21:42,580 --> 00:21:45,159 S1: this new game will be the people making new things 397 00:21:45,160 --> 00:21:47,800 S1: and bringing them into the world. I'm going to say 398 00:21:47,800 --> 00:21:50,530 S1: that again, the winners in this new game will be 399 00:21:50,530 --> 00:21:54,100 S1: the people making new things and bringing them into the world. 400 00:21:54,100 --> 00:21:57,220 S1: And the aphorism of the week. The most beautiful people 401 00:21:57,220 --> 00:22:00,700 S1: we have known are those who have known defeat, known suffering, 402 00:22:00,700 --> 00:22:04,119 S1: known struggle, known loss, and have found their way out 403 00:22:04,119 --> 00:22:06,940 S1: of the depths. Elisabeth Kubler-Ross.