1 00:00:00,920 --> 00:00:05,040 S1: Unsupervised Learning is a podcast about trends and ideas in cybersecurity, 2 00:00:05,160 --> 00:00:10,000 S1: national security, AI, technology and society, and how best to 3 00:00:10,039 --> 00:00:17,840 S1: upgrade ourselves to be ready for what's coming. All right, 4 00:00:17,840 --> 00:00:21,320 S1: welcome to unsupervised learning. This is Daniel Miessler. All right. 5 00:00:21,320 --> 00:00:25,120 S1: So absolutely must see conversation. I'm going to open this 6 00:00:25,120 --> 00:00:25,800 S1: up real quick. 7 00:00:26,640 --> 00:00:27,400 S2: One of the top. 8 00:00:28,200 --> 00:00:34,360 S1: Yeah. This these guys here. Unbelievable conversation about AI specifically. 9 00:00:34,840 --> 00:00:39,840 S1: It is extraordinarily good on the aspect of what just 10 00:00:39,840 --> 00:00:44,120 S1: happened with deep seek and actually competition around chips and 11 00:00:44,120 --> 00:00:48,800 S1: AI and Nvidia and all that stuff. And basically how 12 00:00:48,800 --> 00:00:53,320 S1: competition is happening between all the pinnacle makers, open versus 13 00:00:53,320 --> 00:00:57,040 S1: closed source, basically everything. They're just doing so well in 14 00:00:57,080 --> 00:01:01,900 S1: that conversation. So highly, highly recommend that we made the 15 00:01:01,900 --> 00:01:06,179 S1: list of san's top security newsletters and podcasts. That was cool. 16 00:01:06,220 --> 00:01:08,820 S1: Got a quick thing on how to use O1 and 17 00:01:08,819 --> 00:01:13,340 S1: O3 on fabric? It's just a PP is basically the 18 00:01:13,340 --> 00:01:17,740 S1: flag you want to use. You can't use s um, 19 00:01:17,780 --> 00:01:20,100 S1: you have to use R, which is raw, because you 20 00:01:20,100 --> 00:01:24,259 S1: cannot send a temperature to O1 or O3 where you 21 00:01:24,300 --> 00:01:27,900 S1: can with most of the other models. And uh, after 22 00:01:27,900 --> 00:01:30,500 S1: listening to that conversation I just told you about, I 23 00:01:30,500 --> 00:01:33,620 S1: ended up tightening up my definition of AGI, which we're 24 00:01:33,620 --> 00:01:37,780 S1: going to talk about in a second. But it's essentially, uh, 25 00:01:37,780 --> 00:01:39,580 S1: we're going to go into more depth here in a second, 26 00:01:39,580 --> 00:01:42,300 S1: but it's the ability for an AI, and that could 27 00:01:42,300 --> 00:01:45,060 S1: be a model, a product or a system to perform 28 00:01:45,060 --> 00:01:51,300 S1: the work of an average US based knowledge worker from 2022. And, uh, yeah, 29 00:01:51,300 --> 00:01:53,500 S1: I talk about how this is the most debated term, 30 00:01:53,700 --> 00:01:55,660 S1: and I want to spend a lot of time on it. 31 00:01:55,780 --> 00:02:00,330 S1: This is in, um, raid, uh, real world AI definitions. 32 00:02:00,930 --> 00:02:03,450 S1: And I've got two criteria for whether or not something 33 00:02:03,450 --> 00:02:08,290 S1: is a good definition of AGI. Regular people should understand 34 00:02:08,290 --> 00:02:11,649 S1: if we reached it or not. And attaining it should 35 00:02:11,650 --> 00:02:15,770 S1: be significant to society. So who cares if nobody cares 36 00:02:15,770 --> 00:02:18,610 S1: about it? And also, who cares if you can't figure 37 00:02:18,610 --> 00:02:22,050 S1: out like what somebody is talking about? You know, they 38 00:02:22,050 --> 00:02:25,050 S1: start using big words or something. I feel like that's 39 00:02:25,050 --> 00:02:28,210 S1: not going to stick. It's not a useful definition. All right. Security, 40 00:02:28,250 --> 00:02:32,410 S1: deep sky exposed customer data and unprotected database. So basically 41 00:02:32,410 --> 00:02:36,890 S1: as soon as they launched and all this press starts happening, 42 00:02:36,889 --> 00:02:40,850 S1: Nvidia's stock starts falling. Some people a lot of people 43 00:02:40,850 --> 00:02:45,050 S1: and definitely some people, researchers at Wiz found an open 44 00:02:45,050 --> 00:02:50,010 S1: database full of nasty stuff chat logs, API keys. It's 45 00:02:50,010 --> 00:02:52,730 S1: in Chinese, so I'm not sure if there's any. This 46 00:02:52,730 --> 00:02:55,730 S1: is kind of what I was wondering. Does it reveal 47 00:02:55,770 --> 00:02:59,430 S1: any evidence that they used OpenAI for training? because that 48 00:02:59,430 --> 00:03:02,389 S1: has been one of the things that people have been alleging, 49 00:03:02,630 --> 00:03:04,550 S1: that the only reason they got so good is because 50 00:03:04,550 --> 00:03:08,430 S1: they were using OpenAI to actually train on. And I'm 51 00:03:08,430 --> 00:03:13,230 S1: curious if that's talked about in the logs or, um, or, 52 00:03:13,270 --> 00:03:16,470 S1: you know, if there's any evidence of that happening inside 53 00:03:16,470 --> 00:03:19,550 S1: of the thing inside of the leak. Cisa found that 54 00:03:19,550 --> 00:03:24,510 S1: contact patient monitors have been secretly been sending patient data 55 00:03:24,510 --> 00:03:29,550 S1: to China, and you can download and execute files remotely. Evidently, 56 00:03:29,550 --> 00:03:35,710 S1: there's some sort of, uh, yeah, RCI type deal there. Yeah. Evidently, 57 00:03:36,510 --> 00:03:39,350 S1: when you actually fixed it, it still didn't fix it. 58 00:03:39,350 --> 00:03:42,790 S1: It just disabled the network interface, which the backdoor re-enables. 59 00:03:43,590 --> 00:03:46,390 S1: So that was kind of a mess. Sonicwall says there's 60 00:03:46,390 --> 00:03:53,070 S1: a massive exploit campaign against authentication bypass. And, uh, got 61 00:03:53,070 --> 00:03:56,510 S1: some details here. It's 84. 43 is the port about 62 00:03:56,510 --> 00:04:00,540 S1: 2000 of these things vulnerable on showdown right now. Major 63 00:04:00,540 --> 00:04:06,100 S1: hacking forums seized in international operations. Law enforcement took down 64 00:04:06,100 --> 00:04:09,660 S1: some of the biggest hacking forums included Cracked and Nulled, 65 00:04:09,940 --> 00:04:13,220 S1: which had over 10 million users combined. A new startup 66 00:04:13,220 --> 00:04:16,780 S1: called backline raised $9 million to use AI agents that 67 00:04:16,779 --> 00:04:21,140 S1: automatically fix vulnerabilities. This this is interesting, right? Everyone's talking 68 00:04:21,140 --> 00:04:25,580 S1: about using AI to find vulnerabilities. What about fixing the vulnerabilities? 69 00:04:26,260 --> 00:04:31,860 S1: So I'm really looking forward to vulnerability management remediation services 70 00:04:31,860 --> 00:04:34,860 S1: that use AI. I personally think that this is going 71 00:04:34,860 --> 00:04:37,860 S1: to require a whole lot of understanding of the company. 72 00:04:37,860 --> 00:04:41,420 S1: It's going to require a lot of asset management. You're 73 00:04:41,420 --> 00:04:43,100 S1: going to need to know people. You're going to need 74 00:04:43,100 --> 00:04:46,060 S1: to know organizations. You're going to need to know how 75 00:04:46,060 --> 00:04:49,500 S1: different engineering groups actually push code. So there's a lot 76 00:04:49,500 --> 00:04:51,620 S1: of context you're going to have to get from something 77 00:04:51,620 --> 00:04:54,420 S1: like an asset management system to be able to do this. 78 00:04:54,460 --> 00:04:57,520 S1: Tulsi Gabbard is facing a lot of pushback back in 79 00:04:57,520 --> 00:05:01,960 S1: her thing. I'm not sure if she's already been passed through. Uh, 80 00:05:02,000 --> 00:05:05,080 S1: I'm trying not to pay attention to the news right now. 81 00:05:05,200 --> 00:05:09,360 S1: It's a bit depressing, honestly. But, um. Yeah. Want to 82 00:05:09,360 --> 00:05:11,440 S1: make a little bit of a statement here? I think 83 00:05:11,440 --> 00:05:14,120 S1: if you dump top secret documents to the internet, this 84 00:05:14,120 --> 00:05:15,720 S1: is one of the things she was being asked about 85 00:05:15,760 --> 00:05:19,560 S1: was Snowden. Or if you break into the Capitol building 86 00:05:19,640 --> 00:05:24,840 S1: on verification day, on results verification day, because you want 87 00:05:24,839 --> 00:05:29,279 S1: to hopefully change the results, you are an actual criminal. 88 00:05:29,320 --> 00:05:31,440 S1: That's that's what I think. At one point I saw 89 00:05:31,440 --> 00:05:35,359 S1: Snowden as a whistleblower, too. I think he actually technically 90 00:05:35,360 --> 00:05:38,000 S1: is one, and I think he might have been trying 91 00:05:38,000 --> 00:05:40,799 S1: to do some good at some point, but I think 92 00:05:40,839 --> 00:05:42,880 S1: it kind of got all messed up. And I haven't 93 00:05:42,880 --> 00:05:45,960 S1: really been on that boat for many years now. Bedfordshire 94 00:05:46,000 --> 00:05:49,600 S1: Police just became the first UK force to deploy Palantir's 95 00:05:49,640 --> 00:05:53,520 S1: AI system, and they got some good results. 123 at 96 00:05:53,560 --> 00:05:57,470 S1: risk kids were found in just one day argument that 97 00:05:57,470 --> 00:05:59,950 S1: we're getting passkeys all wrong, and that they should be 98 00:05:59,950 --> 00:06:04,510 S1: used alongside magic links, not as a complete replacement for 99 00:06:04,510 --> 00:06:08,110 S1: other auth methods. I've been thinking about this a lot. 100 00:06:08,470 --> 00:06:12,589 S1: I really do love Passkeys, but I do worry that 101 00:06:12,589 --> 00:06:15,390 S1: it's I don't know, it feels like we might be 102 00:06:15,390 --> 00:06:17,910 S1: being a bit sloppy with them. All right. This is 103 00:06:17,910 --> 00:06:21,029 S1: why I think you should care about us reaching AGI. 104 00:06:21,070 --> 00:06:23,390 S1: Wanted to say more about the AGI thing. I think 105 00:06:23,390 --> 00:06:27,190 S1: it's the most important topic in AI. Actually, tons of 106 00:06:27,190 --> 00:06:32,590 S1: very smart people don't know why they should care about it. Like, 107 00:06:32,589 --> 00:06:37,150 S1: who cares if AGI hits this or that benchmark. And 108 00:06:37,150 --> 00:06:41,190 S1: I think there's only one good reason why you should care. 109 00:06:41,190 --> 00:06:45,710 S1: And here's how I talk through that. So think I 110 00:06:45,750 --> 00:06:50,030 S1: coworkers right. Like actual coworkers that you have in person. 111 00:06:50,029 --> 00:06:52,349 S1: But imagine your team at work. You've got like five 112 00:06:52,350 --> 00:06:56,409 S1: coworkers or 20 or 35 whatever, however big your team is. 113 00:06:56,570 --> 00:07:01,409 S1: Now imagine it's like 10,000 co-workers instead. Like overnight. So 114 00:07:01,410 --> 00:07:04,170 S1: it was seven people. You show up and now it's 115 00:07:04,170 --> 00:07:08,490 S1: a zoom call with 150 people or 10,000 people instead 116 00:07:08,490 --> 00:07:13,290 S1: of seven. And you're watching them work. You're seeing them work. 117 00:07:13,890 --> 00:07:17,130 S1: A couple of them, you see physically work you see 118 00:07:17,170 --> 00:07:20,930 S1: them the results of their code, you see it come in. 119 00:07:20,930 --> 00:07:23,170 S1: You get other signals telling you how good of a 120 00:07:23,170 --> 00:07:26,770 S1: job they're doing or whatever, but you find out they're 121 00:07:26,770 --> 00:07:30,650 S1: not perfect. They make mistakes just like anyone else. Someone 122 00:07:30,650 --> 00:07:33,210 S1: still has to review their work. They still get lost. 123 00:07:33,210 --> 00:07:36,130 S1: Sometimes they're like, what are we doing? Like, I'm not sure. 124 00:07:36,170 --> 00:07:37,770 S1: You know, I'm not sure if I should be working 125 00:07:37,770 --> 00:07:40,610 S1: on this project. Does this look correct? Or they mess 126 00:07:40,610 --> 00:07:44,050 S1: up the code base or whatever. In some ways, they're way, 127 00:07:44,090 --> 00:07:48,210 S1: way better than your human counterparts, and in some ways 128 00:07:48,210 --> 00:07:51,930 S1: they're just way more stupid, right? But what they do 129 00:07:52,370 --> 00:07:56,350 S1: is they make steady progress. They show up for video calls. 130 00:07:56,510 --> 00:08:00,390 S1: They can read docs. They can read code. They're talking 131 00:08:00,430 --> 00:08:03,830 S1: to you in slack. They're asking questions. When you ask 132 00:08:03,830 --> 00:08:06,270 S1: them a question, they give you a response about the 133 00:08:06,270 --> 00:08:09,230 S1: work that they're doing, or they help you find a document, whatever. 134 00:08:09,430 --> 00:08:13,630 S1: And they can also readjust based on a coworker or 135 00:08:13,630 --> 00:08:15,710 S1: their manager telling them, hey, that's not the way you 136 00:08:15,710 --> 00:08:17,710 S1: should be doing this, or we got a new directive 137 00:08:17,710 --> 00:08:21,150 S1: from above. You know, the goal is slightly shifted or whatever. 138 00:08:21,230 --> 00:08:23,390 S1: They can change what they're doing. But the key here 139 00:08:23,390 --> 00:08:28,310 S1: is there's actually 10,000 of these people, right? There's 10,000 140 00:08:28,350 --> 00:08:32,230 S1: of these agents instead of ten, or there's 100,000 of 141 00:08:32,230 --> 00:08:36,110 S1: these instead of your team of 100, and they work 142 00:08:36,110 --> 00:08:38,830 S1: 24 over seven. They put in the work, they put 143 00:08:38,830 --> 00:08:42,110 S1: in the effort, and they do not get tired. And 144 00:08:42,110 --> 00:08:45,069 S1: they also constantly improve. So if they start off with 145 00:08:45,070 --> 00:08:49,030 S1: like 105 IQ, which is, you know, decently smart or 146 00:08:49,030 --> 00:08:52,830 S1: two years of knowledge in security or programming or intern 147 00:08:52,830 --> 00:08:55,740 S1: level or whatever. After a year or two, they've had 148 00:08:55,780 --> 00:08:59,700 S1: like 20 updates, which were silent updates or whatever. And 149 00:08:59,700 --> 00:09:03,220 S1: assuming none of those updates went horribly wrong, uh, now 150 00:09:03,220 --> 00:09:07,100 S1: they got 110 IQ. Now they got 120 IQ. Right. 151 00:09:07,140 --> 00:09:12,580 S1: Or whatever. The AGI alternative to uh, or analog to 152 00:09:12,620 --> 00:09:15,460 S1: IQ is. It's not quite IQ, but call it raw 153 00:09:15,460 --> 00:09:19,540 S1: fluid intelligence. Right. So they're getting those constant updates, but 154 00:09:19,540 --> 00:09:22,220 S1: they're also learning to be better coders, not just by 155 00:09:22,220 --> 00:09:24,820 S1: the work that they're doing, but actually by just better 156 00:09:24,820 --> 00:09:28,980 S1: and better models. So now instead of 105 IQ with 157 00:09:28,980 --> 00:09:34,939 S1: two years work experience, now it's 120 IQ with 20 158 00:09:34,980 --> 00:09:38,780 S1: years of experience. So it's like a staff engineer, but 159 00:09:38,780 --> 00:09:42,620 S1: you still have 100,000 of them, right? This is why, 160 00:09:43,340 --> 00:09:45,860 S1: in my opinion, AGI is a big deal. And I 161 00:09:45,860 --> 00:09:48,740 S1: think we're getting really close to this. I've got another 162 00:09:48,900 --> 00:09:50,940 S1: thing to read about. This is what I just posted 163 00:09:50,940 --> 00:09:54,280 S1: earlier on one of the Socials, but I think we're 164 00:09:54,280 --> 00:09:57,960 S1: getting close to this. It's not one component, right? This 165 00:09:57,960 --> 00:10:00,000 S1: is not like one model that's going to release and 166 00:10:00,000 --> 00:10:02,800 S1: do this. It will be a system. It will behave 167 00:10:02,800 --> 00:10:05,480 S1: like one person or one thing or whatever you want 168 00:10:05,480 --> 00:10:08,800 S1: to call it, but it'll be this composite that lets 169 00:10:08,800 --> 00:10:10,960 S1: it behave in a cohesive way, which is actually the 170 00:10:10,960 --> 00:10:14,400 S1: way our brains work as well. My guess was in 171 00:10:14,400 --> 00:10:17,600 S1: 2023 that it was going to be 25 to 28. 172 00:10:17,720 --> 00:10:19,360 S1: That was my thing. And I put that in like 173 00:10:19,360 --> 00:10:23,280 S1: two years ago. Okay. Like March of 23, I think 174 00:10:23,280 --> 00:10:26,959 S1: I made this prediction. 25 to 28 is when we're 175 00:10:26,960 --> 00:10:30,199 S1: going to have this. And my definition for this is 176 00:10:30,240 --> 00:10:35,400 S1: basically somebody who can replace a human knowledge worker, a 177 00:10:35,400 --> 00:10:40,320 S1: decent human knowledge worker working at some remote job. Right. 178 00:10:40,360 --> 00:10:42,160 S1: So I don't think they're going to be able to like, 179 00:10:42,200 --> 00:10:45,560 S1: bring you a coffee because that's robotics. But working a 180 00:10:45,559 --> 00:10:53,750 S1: remote job, pushing code, doing organization, project management. Um, editing documents, 181 00:10:54,590 --> 00:10:58,230 S1: that kind of stuff. 25 to 28. I think that's 182 00:10:58,270 --> 00:11:02,550 S1: what is when we're going to actually hit this. So 183 00:11:02,550 --> 00:11:06,630 S1: my current estimate for this is late 25 or maybe 184 00:11:06,630 --> 00:11:10,550 S1: sometime in 26, and it could actually go into 27. 185 00:11:10,550 --> 00:11:13,750 S1: Because this type of progress that we're seeing in AI, 186 00:11:13,750 --> 00:11:17,350 S1: it's big jumps. But then it kind of plateaus for 187 00:11:17,550 --> 00:11:20,750 S1: weeks or months. But that plateau could actually go for 188 00:11:20,750 --> 00:11:23,030 S1: a year. I don't think it will. I think it's 189 00:11:23,030 --> 00:11:26,390 S1: going to continue to jump kind of violently, but it 190 00:11:26,390 --> 00:11:31,869 S1: could plateau out and then maybe it turns into more 2728. 191 00:11:32,630 --> 00:11:35,990 S1: I think the chances that it happens after 29 are 192 00:11:35,990 --> 00:11:40,590 S1: really low, maybe 5%, which is a lot. It's a lot, right? 193 00:11:40,630 --> 00:11:46,870 S1: So my guarantee for this right now, my high, it's like, uh, 194 00:11:46,950 --> 00:11:52,210 S1: what's the CIA, CIA range? I think the CIA range is, Um, 195 00:11:52,250 --> 00:11:56,929 S1: almost guaranteed or something like that. It's, um, extremely probable. 196 00:11:56,929 --> 00:12:00,330 S1: It's one of those. But it's only like 95%. And 197 00:12:00,330 --> 00:12:03,170 S1: that's where I'm at because I'm not really 100% on anything. 198 00:12:03,290 --> 00:12:04,969 S1: And when this does happen, I think it'll be the 199 00:12:04,970 --> 00:12:09,010 S1: single biggest impact on humanity from tech, by far bigger 200 00:12:09,010 --> 00:12:13,770 S1: than the internet and both in the negative and positive directions. 201 00:12:13,770 --> 00:12:16,250 S1: So that's why I think you should care about AGI. 202 00:12:16,410 --> 00:12:20,490 S1: And in perfect timing, MuleSoft reports that almost all IT 203 00:12:20,530 --> 00:12:24,690 S1: leaders are planning to use autonomous AI agents within two years. 204 00:12:24,690 --> 00:12:28,930 S1: That's 93% and about half are already doing so. Sam 205 00:12:28,929 --> 00:12:31,450 S1: Altman basically says he feels like he was on the 206 00:12:31,450 --> 00:12:34,730 S1: wrong side of history with open source. He didn't say 207 00:12:34,730 --> 00:12:37,650 S1: he's going to do anything about it or what he's 208 00:12:37,650 --> 00:12:39,650 S1: going to do about it. It seems like kind of 209 00:12:39,690 --> 00:12:42,650 S1: obvious that he will do something. My guess is they 210 00:12:42,650 --> 00:12:47,130 S1: will release some nerfed versions of some of their models 211 00:12:47,130 --> 00:12:49,450 S1: just to be like, look, we are playing nice with 212 00:12:49,450 --> 00:12:52,429 S1: the open source model, but I don't think they're going 213 00:12:52,470 --> 00:12:55,750 S1: to completely pivot to like what meta is doing, for example. 214 00:12:55,910 --> 00:12:59,710 S1: And like I said, yeah. OpenAI claims Chinese rival deep 215 00:12:59,750 --> 00:13:03,190 S1: tech stole training data. Yeah, I don't actually think it's 216 00:13:03,190 --> 00:13:06,110 S1: kind of funny to be like, oh, OpenAI is worried 217 00:13:06,110 --> 00:13:08,830 S1: about their stuff being stolen when, you know, they stole 218 00:13:08,830 --> 00:13:12,670 S1: the whole internet to train on. I think. I mean, 219 00:13:12,830 --> 00:13:14,630 S1: anything you put on the internet is just going to 220 00:13:14,630 --> 00:13:17,390 S1: be on the internet, and that's free game. I think, 221 00:13:17,550 --> 00:13:20,829 S1: I think unless you have specific like licenses or something, 222 00:13:21,030 --> 00:13:23,589 S1: and obviously I don't want people to steal like paid 223 00:13:23,590 --> 00:13:28,150 S1: content that I'm producing. Right. But but I don't release 224 00:13:28,270 --> 00:13:33,190 S1: paid content live to the public, on social media or 225 00:13:33,190 --> 00:13:36,110 S1: on a blog post, and then get mad when someone 226 00:13:36,110 --> 00:13:39,110 S1: scrapes it, right? I consider all the stuff that I 227 00:13:39,110 --> 00:13:42,069 S1: put out since like 1999 or whatever, how long I've 228 00:13:42,070 --> 00:13:44,230 S1: been on the internet. You gave it to the world 229 00:13:44,230 --> 00:13:46,790 S1: when you posted it online. That's the way I see it. 230 00:13:46,790 --> 00:13:51,100 S1: I always have R0 or R1 zero Shows AI reasoning 231 00:13:51,100 --> 00:13:54,380 S1: without human training. This is really cool because like I 232 00:13:54,380 --> 00:13:57,940 S1: talk about here, it's like what happened in chess. So 233 00:13:57,980 --> 00:14:01,660 S1: at first they used they built a chess computer. Um, 234 00:14:01,660 --> 00:14:04,340 S1: I think this was Google that was doing this, and 235 00:14:04,340 --> 00:14:07,179 S1: it just got really good at learning the rules, and 236 00:14:07,179 --> 00:14:10,260 S1: it watched human players and it improved on that. And 237 00:14:10,300 --> 00:14:14,020 S1: of course, it could beat previous chess computers like Deep 238 00:14:14,020 --> 00:14:16,940 S1: Blue is nothing compared to these computers. They just crush it. 239 00:14:17,220 --> 00:14:21,500 S1: What the big breakthrough was with AlphaZero and also with 240 00:14:21,500 --> 00:14:27,540 S1: R1 zero is it's learning more actually, or at least 241 00:14:28,140 --> 00:14:34,340 S1: significantly more. It's learning significantly more from the reinforcement learning 242 00:14:35,140 --> 00:14:38,380 S1: from the world. It's learning from the world as opposed 243 00:14:38,380 --> 00:14:41,980 S1: to from the rules or watching other people interact with 244 00:14:41,980 --> 00:14:45,660 S1: the world. So it's just an accelerator because it could 245 00:14:45,660 --> 00:14:48,620 S1: do that by itself, right? Without a lot of effort 246 00:14:48,620 --> 00:14:52,320 S1: from other people and that's really, really powerful. A new 247 00:14:52,320 --> 00:14:55,640 S1: study shows deepfakes AI model refuses to answer the vast 248 00:14:55,640 --> 00:15:00,240 S1: majority of sensitive questions about China. Shocking, shocking. This is 249 00:15:00,240 --> 00:15:03,240 S1: why it matters where you get your models right. Because 250 00:15:03,440 --> 00:15:06,080 S1: according to them, Tiananmen Square didn't happen. It's all a 251 00:15:06,080 --> 00:15:10,320 S1: breakdown of how to properly evaluate ragtag systems and llms 252 00:15:10,320 --> 00:15:16,560 S1: in Practice by Salman Khan and Andrej Karpathy talks about 253 00:15:16,680 --> 00:15:19,480 S1: vibe coding, where you basically get in a flow state 254 00:15:19,640 --> 00:15:22,200 S1: and code like you're playing an instrument. It says the 255 00:15:22,200 --> 00:15:25,600 S1: key is to stop overthinking and just think and respond 256 00:15:25,600 --> 00:15:28,000 S1: and let the AI do most of the work. I 257 00:15:28,000 --> 00:15:30,840 S1: think this is spot on and a lot of people 258 00:15:30,840 --> 00:15:33,080 S1: are really mad. They're like, oh, I can't believe you know, 259 00:15:33,080 --> 00:15:35,000 S1: you're a hard core coder and you're just going to 260 00:15:35,000 --> 00:15:38,440 S1: let AI code for you. And again, he doesn't care. 261 00:15:38,520 --> 00:15:41,800 S1: I don't care. People like us, we don't care because 262 00:15:42,160 --> 00:15:45,360 S1: we could see what's actually coming, which is pretty soon. 263 00:15:45,360 --> 00:15:47,960 S1: You don't fire up an IDE at all. What you 264 00:15:47,960 --> 00:15:51,390 S1: do is you fire up a voice conversation and the 265 00:15:51,390 --> 00:15:55,870 S1: thing starts showing you what the product looks like on 266 00:15:55,870 --> 00:15:59,230 S1: one side, and it's showing you roughly like the code 267 00:15:59,230 --> 00:16:01,910 S1: or the diagram or whatever it is on the left, 268 00:16:02,550 --> 00:16:05,630 S1: and you're having a verbal back and forth like that 269 00:16:05,630 --> 00:16:09,390 S1: is coding within, you know, 1 to 3 years. So 270 00:16:09,390 --> 00:16:11,630 S1: this whole thing of like, oh, you're not doing the 271 00:16:11,630 --> 00:16:14,510 S1: real thing unless you're doing the real coding. Once again, 272 00:16:14,510 --> 00:16:17,350 S1: you have to take that all the way back. Um, 273 00:16:17,350 --> 00:16:20,870 S1: are you writing assembler by hand? No you're not. Okay, well, 274 00:16:21,550 --> 00:16:24,470 S1: you're cheating by using Python or whatever, right? So it 275 00:16:24,470 --> 00:16:27,590 S1: just keeps going farther and farther away, and eventually it's 276 00:16:27,590 --> 00:16:29,190 S1: going to be like, I wish I had a product 277 00:16:29,190 --> 00:16:32,830 S1: that looked like this, blah, blah, blah, and boom, it 278 00:16:32,830 --> 00:16:37,030 S1: just goes into the world. And what did you really contribute? 279 00:16:37,030 --> 00:16:41,270 S1: You contributed the raw base idea. But at some point, 280 00:16:41,270 --> 00:16:43,790 S1: the A's are good enough to contribute that as well. 281 00:16:43,790 --> 00:16:47,070 S1: And now you're competing with them on that. And that's 282 00:16:47,070 --> 00:16:50,290 S1: when it turns a little bit sad. Apple quietly added 283 00:16:50,290 --> 00:16:54,530 S1: Starlink satellite support to iPhones through a software update. And 284 00:16:54,530 --> 00:16:58,410 S1: this is a partnership with SpaceX and T-Mobile. So yeah, 285 00:16:58,450 --> 00:17:01,130 S1: SpaceX is how your iPhone is able to talk to 286 00:17:01,130 --> 00:17:05,570 S1: satellites when you have no signal. Humans massive new survey 287 00:17:06,010 --> 00:17:11,649 S1: revealed that 87% of astrobiologists think that extraterrestrial life exists 288 00:17:11,650 --> 00:17:15,970 S1: somewhere in the universe. 87%. I am one of those people, 289 00:17:15,970 --> 00:17:18,090 S1: not one of the astrobiologists, but one of the people 290 00:17:18,090 --> 00:17:20,689 S1: who agrees with that study. A Montana shows that drones 291 00:17:20,690 --> 00:17:24,330 S1: are way better at keeping grizzlies away from humans, so 292 00:17:24,330 --> 00:17:27,250 S1: they basically fly up to them and chase them and 293 00:17:27,250 --> 00:17:32,169 S1: scare them. Could it be cool to take drones with 294 00:17:32,170 --> 00:17:35,210 S1: you when you go into the forest? So you're worried 295 00:17:35,210 --> 00:17:38,369 S1: about bears? You release like three drones or whatever. That's 296 00:17:38,369 --> 00:17:41,250 S1: part of your panic response. You hit a button. It 297 00:17:41,250 --> 00:17:44,490 S1: makes a really loud sound, like some sort of horn 298 00:17:44,490 --> 00:17:47,280 S1: or something, and you could point the horn at the thing, 299 00:17:47,680 --> 00:17:50,960 S1: but it also launches the drones and the drones go 300 00:17:50,960 --> 00:17:55,119 S1: over and maybe they make noises. Maybe they have sounds, whatever, 301 00:17:55,240 --> 00:17:59,359 S1: but they just start buzzing around the the bear. And maybe, like, 302 00:17:59,400 --> 00:18:02,359 S1: you could sell that as a package is like bear defense. 303 00:18:02,440 --> 00:18:05,680 S1: Maybe they shoot, uh, bear spray as well. That would 304 00:18:05,680 --> 00:18:09,200 S1: be cool. All right. Swerving Broncos I was thinking about well, 305 00:18:09,200 --> 00:18:12,120 S1: I was basically driving on the 101 at like 11:00 306 00:18:12,119 --> 00:18:15,520 S1: at night or something just a couple days ago. And 307 00:18:16,359 --> 00:18:18,600 S1: I am watching the cars on the, on the road 308 00:18:18,600 --> 00:18:22,320 S1: in front of me. And they're just like, drifting, like 309 00:18:22,800 --> 00:18:25,600 S1: right over the line. The second wheel goes over the line. 310 00:18:25,600 --> 00:18:29,600 S1: They're just like drifting. Obvious that they're on their phones 311 00:18:29,600 --> 00:18:32,600 S1: or drunk or high. And then like, I get away 312 00:18:32,600 --> 00:18:35,520 S1: from that thing or it turns off the highway or whatever. 313 00:18:35,520 --> 00:18:38,240 S1: And like, here comes another giant truck, a giant like 314 00:18:38,240 --> 00:18:40,840 S1: Cadillac or a Bronco or something like that comes next 315 00:18:40,840 --> 00:18:43,560 S1: to me, and it just starts drifting into my lane. 316 00:18:43,560 --> 00:18:46,300 S1: And I'm like, what are you doing? And I And 317 00:18:46,300 --> 00:18:48,220 S1: I look at them, they're like looking down. They're not 318 00:18:48,220 --> 00:18:51,620 S1: even paying attention. And I see a third and this 319 00:18:51,619 --> 00:18:54,659 S1: is within like 30 miles or something. I see a 320 00:18:54,660 --> 00:18:57,700 S1: third one. They're just like drifting over. There's no cops around. 321 00:18:57,700 --> 00:19:01,260 S1: I hardly ever see cops anywhere. Like on the on 322 00:19:01,260 --> 00:19:05,180 S1: the highways anymore. There's no cops. There's no one stopping them. 323 00:19:05,180 --> 00:19:08,180 S1: They're just kind of like driving like crazy people. And 324 00:19:08,180 --> 00:19:12,820 S1: I'm thinking, you know what won't do? This is Waymo. 325 00:19:12,859 --> 00:19:17,060 S1: Waymo won't do this either. Will autonomous Tesla or an 326 00:19:17,060 --> 00:19:21,340 S1: autonomous like BYD device from China or whatever. Say what 327 00:19:21,340 --> 00:19:26,220 S1: you want about AI driving. They bicyclists love them because 328 00:19:26,220 --> 00:19:29,500 S1: they never crash into them. I mean, I think we 329 00:19:29,500 --> 00:19:33,460 S1: just have to realize how bad human drivers are before 330 00:19:33,580 --> 00:19:37,980 S1: we start criticizing how bad I drivers are. AI's big jump. 331 00:19:38,180 --> 00:19:43,020 S1: This one is very simple. I think AI from Apple 332 00:19:43,020 --> 00:19:46,040 S1: is like it's been the worst for a while and 333 00:19:46,040 --> 00:19:48,280 S1: it's about to be the best. That's the bottom line. 334 00:19:48,480 --> 00:19:54,200 S1: They are putting out stuff in 18.4 that basically finally 335 00:19:54,200 --> 00:19:58,840 S1: connects Siri to the context inside the phone. So your phone, 336 00:19:58,840 --> 00:20:01,720 S1: you have your exercise, you have your heartbeat, you have 337 00:20:01,720 --> 00:20:06,040 S1: your appointments and your calendar and your email. And it 338 00:20:06,040 --> 00:20:09,680 S1: could see all the different stuff. It could see your life. Right. 339 00:20:09,680 --> 00:20:13,480 S1: But it's never unified that stuff. Well, Apple just got really, 340 00:20:13,480 --> 00:20:16,760 S1: really serious about Siri or they're in the process of 341 00:20:16,760 --> 00:20:19,600 S1: doing that. They got a new leader over it. Like 342 00:20:19,600 --> 00:20:23,399 S1: it's just become super high priority. They're doing that at 343 00:20:23,400 --> 00:20:26,560 S1: the same time that they're doing the cloud secure enclave. 344 00:20:26,800 --> 00:20:29,440 S1: At the same time, they're doing the partnership with ChatGPT 345 00:20:30,000 --> 00:20:33,320 S1: at the same time that they're now giving Siri access 346 00:20:33,320 --> 00:20:36,560 S1: to all that context that they have, and access to 347 00:20:36,960 --> 00:20:41,240 S1: call apps and pass data between the different apps. So 348 00:20:41,240 --> 00:20:44,439 S1: they've been setting up this infrastructure for AI for the 349 00:20:44,440 --> 00:20:46,870 S1: last couple of years. They've built all this stuff. They've 350 00:20:46,869 --> 00:20:50,430 S1: spent all this R&D. And again, I used to work there. 351 00:20:50,470 --> 00:20:52,510 S1: This has nothing to do with anything. I've been gone 352 00:20:52,510 --> 00:20:56,190 S1: for a while. And also, I wouldn't be disclosing anything 353 00:20:56,190 --> 00:20:58,990 S1: if I actually worked on a project like that. But 354 00:20:58,990 --> 00:21:03,189 S1: I've been watching them since, whatever, 2007. And I'm telling 355 00:21:03,190 --> 00:21:06,229 S1: you that they are about to switch this thing on. 356 00:21:06,830 --> 00:21:09,389 S1: And I really do think that they're about to have 357 00:21:09,390 --> 00:21:15,430 S1: the best personal ecosystem integration AI type play. And it's 358 00:21:15,430 --> 00:21:18,310 S1: going to go from being like a are they even 359 00:21:18,310 --> 00:21:21,430 S1: doing anything to Holy crap, this is what they've been 360 00:21:21,430 --> 00:21:25,149 S1: working on. That is my prediction around that. And I 361 00:21:25,190 --> 00:21:27,790 S1: novels are coming. I actually have a YouTube video about this. 362 00:21:27,910 --> 00:21:30,550 S1: But yeah, I'll wait till that comes out. It might 363 00:21:30,590 --> 00:21:33,109 S1: actually be out soon, but yeah. Anyway, I'm going to 364 00:21:33,109 --> 00:21:35,629 S1: wait to talk about this one. Basically, it's going to 365 00:21:35,630 --> 00:21:39,990 S1: be possible to make a complete decent novel using AI. 366 00:21:40,030 --> 00:21:43,790 S1: I think within months or years. I've already got a 367 00:21:43,790 --> 00:21:45,689 S1: got a couple of emails actually, where they're like, what 368 00:21:45,690 --> 00:21:47,330 S1: are you talking about? I've been doing that for a 369 00:21:47,330 --> 00:21:51,410 S1: year and a half. So I think it was possible 370 00:21:51,410 --> 00:21:53,330 S1: to do a year and a half ago, and I've 371 00:21:53,330 --> 00:21:58,410 S1: messed around with some, uh, longer, uh, like, fiction stories, 372 00:21:58,530 --> 00:22:01,890 S1: and it was halfway decent. And that was like a year, 373 00:22:01,930 --> 00:22:05,050 S1: year and a half ago. But it was difficult. What 374 00:22:05,050 --> 00:22:07,050 S1: I'm trying to say is it's going to get easy 375 00:22:07,250 --> 00:22:09,850 S1: and lots of people are going to do it. Um, 376 00:22:09,890 --> 00:22:12,889 S1: ChatGPT new tasks feature. A lot of people are talking 377 00:22:12,890 --> 00:22:17,609 S1: about this. It's basically scheduled and automated, um, agents to 378 00:22:17,650 --> 00:22:20,129 S1: go and do things for you. I have not messed 379 00:22:20,130 --> 00:22:23,409 S1: with this one, but it looks really, really exciting. Really 380 00:22:23,410 --> 00:22:27,330 S1: clever way to use UV which is a fast Python 381 00:22:27,330 --> 00:22:30,690 S1: package manager. Look at this. You put this at the 382 00:22:30,690 --> 00:22:33,290 S1: top of oh man, I haven't messed with this yet. 383 00:22:33,290 --> 00:22:35,490 S1: I have to write this down. I ran it once, 384 00:22:35,490 --> 00:22:38,609 S1: but I haven't put it in like all my main scripts. 385 00:22:39,369 --> 00:22:43,600 S1: This allows you to run anything that has multiple dependencies 386 00:22:43,600 --> 00:22:49,800 S1: and normally breaks, and it will automatically instantiate you a 387 00:22:49,800 --> 00:22:53,800 S1: working environment to make your script run. Isn't that just beautiful? 388 00:22:53,840 --> 00:22:57,719 S1: It's absolutely beautiful. I mean, this is so much better 389 00:22:57,720 --> 00:23:01,679 S1: than regular Pip, because if you have multiple environments like 390 00:23:02,040 --> 00:23:04,950 S1: that's my problem. My regular Python has like 3 or 391 00:23:04,950 --> 00:23:10,439 S1: 4 different environments, different virtual environments, like nested spread all 392 00:23:10,440 --> 00:23:14,040 S1: over the file system. It's like a nightmare. Um, so 393 00:23:14,040 --> 00:23:18,080 S1: that's why I use UV now. And this is just. Yeah, 394 00:23:18,119 --> 00:23:20,640 S1: this is really great. I haven't seen a downside yet, 395 00:23:20,640 --> 00:23:22,840 S1: but I haven't messed with it that much. Uh, cool. 396 00:23:22,840 --> 00:23:28,120 S1: Developer created Napthine's malicious software that traps aggressive AI web 397 00:23:28,119 --> 00:23:31,119 S1: crawlers in infinite loops and feeds them garbage data to 398 00:23:31,160 --> 00:23:35,560 S1: poison their models. Smiling, not smiling OpenAI dropped new ChatGPT 399 00:23:35,560 --> 00:23:39,240 S1: capability called deep Research. Supposedly you only get like 100 400 00:23:39,280 --> 00:23:42,320 S1: a month because they're really expensive and it takes like 401 00:23:42,320 --> 00:23:46,260 S1: takes like ten minutes to finish, but I've heard it's 402 00:23:46,260 --> 00:23:51,300 S1: created some serious, serious research. Really good papers. Haven't checked 403 00:23:51,300 --> 00:23:53,820 S1: if I have access to that API yet. All right. 404 00:23:53,859 --> 00:23:58,979 S1: Justin McGurk explores how William Gibson novels perfectly capture our 405 00:23:58,980 --> 00:24:03,020 S1: obsession with commodifying everything unique and authentic until it loses 406 00:24:03,020 --> 00:24:06,739 S1: all meaning. This is why I'm doing human 3.0. I 407 00:24:06,740 --> 00:24:10,660 S1: do not want to be replaced by AI and technology. 408 00:24:10,780 --> 00:24:15,380 S1: No thank you. I want it to enhance humanity. Found 409 00:24:15,380 --> 00:24:19,380 S1: a really clean python script that lets you download YouTube videos. 410 00:24:19,420 --> 00:24:22,900 S1: This thing works really well, I just worried. I'm just 411 00:24:22,900 --> 00:24:25,900 S1: worried it's going to get blocked soon. Recommendation of the week. 412 00:24:25,900 --> 00:24:28,620 S1: If you ever get overwhelmed by what all this AI 413 00:24:28,660 --> 00:24:31,620 S1: stuff even means, or you want to explain to someone else, 414 00:24:31,740 --> 00:24:34,020 S1: I got a little quote here. So within the next 415 00:24:34,020 --> 00:24:37,580 S1: few years we might have something called AGI where I 416 00:24:37,619 --> 00:24:40,620 S1: can work as a full knowledge worker, like joining the 417 00:24:40,619 --> 00:24:46,490 S1: onboarding cohorts, reading documentation, participating on slack, submitting code, adjusting 418 00:24:46,490 --> 00:24:49,130 S1: their work based on the work of others. But instead 419 00:24:49,130 --> 00:24:52,810 S1: of 2 or 5 of these, imagine hundreds of them 420 00:24:52,810 --> 00:24:56,090 S1: for the cost of one human employee. I think that's 421 00:24:56,090 --> 00:24:59,730 S1: pretty cool way to kind of explain the possible impact 422 00:24:59,850 --> 00:25:02,850 S1: in a way that, like regular people who don't pay 423 00:25:02,850 --> 00:25:06,050 S1: attention to this all the time could understand. And the 424 00:25:06,050 --> 00:25:09,330 S1: aphorism of the week, some books leave us free and 425 00:25:09,330 --> 00:25:12,810 S1: some books make us free, some books leave us free, 426 00:25:13,250 --> 00:25:18,449 S1: and some books make us free. Ralph Waldo Emerson. Unsupervised 427 00:25:18,450 --> 00:25:21,770 S1: learning is produced on Hindenburg Pro using an SM seven 428 00:25:21,810 --> 00:25:25,450 S1: B microphone. A video version of the podcast is available 429 00:25:25,450 --> 00:25:29,090 S1: on the Unsupervised Learning YouTube channel, and the text version 430 00:25:29,090 --> 00:25:34,210 S1: with full links and notes is available at Daniel Miessler newsletter. 431 00:25:34,850 --> 00:25:35,810 S1: We'll see you next time.