1 00:00:00,800 --> 00:00:04,939 S1: Unsupervised Learning is a podcast about trends and ideas in cybersecurity, 2 00:00:04,970 --> 00:00:10,010 S1: national security, AI, technology and society, and how best to 3 00:00:10,039 --> 00:00:17,660 S1: upgrade ourselves to be ready for what's coming. All right, 4 00:00:17,660 --> 00:00:22,369 S1: welcome to episode 472. Hope your week is going well. 5 00:00:22,400 --> 00:00:26,180 S1: Updates on this side I've got an RSS feed which 6 00:00:26,180 --> 00:00:29,180 S1: I encourage you to go sign up for. RSS is 7 00:00:29,180 --> 00:00:32,449 S1: a real thing and it's awesome. And I wish it 8 00:00:32,450 --> 00:00:37,430 S1: didn't get knocked off the radar by Google Reader going under, 9 00:00:37,760 --> 00:00:41,660 S1: but I use Feedly. A lot of people use Netnewswire. 10 00:00:41,690 --> 00:00:45,530 S1: Whatever your RSS reader is, go check it out. It's 11 00:00:45,530 --> 00:00:49,310 S1: just slash feed RSS, I think at the end of 12 00:00:49,310 --> 00:00:52,130 S1: the URL. Thanks for all the cooking advice. Around 500 13 00:00:52,130 --> 00:00:57,650 S1: people responded with experience, encouragement, recommendations. It was just insane. 14 00:00:57,650 --> 00:01:01,760 S1: So really appreciate all that. I tried windsurf, a lot 15 00:01:01,790 --> 00:01:03,560 S1: of people are talking about it. I basically try all 16 00:01:03,560 --> 00:01:07,310 S1: these tools that people talk about. I've been using Klein 17 00:01:07,310 --> 00:01:11,060 S1: for a while. I was using cursor before that and 18 00:01:11,060 --> 00:01:13,640 S1: I keep trying cursor every time they have a big 19 00:01:13,640 --> 00:01:17,419 S1: release or when windsurf has a big release. I'll try 20 00:01:17,420 --> 00:01:20,570 S1: it again. The stuff is changing so fast that I 21 00:01:20,569 --> 00:01:24,740 S1: just constantly try different stuff. But Klein has been my 22 00:01:24,770 --> 00:01:29,030 S1: go to. I feel like it manages the functionality of 23 00:01:29,060 --> 00:01:33,170 S1: Lmms better and the UI. I just like better, so 24 00:01:33,170 --> 00:01:35,840 S1: that's what I've been using. It is just an extension 25 00:01:35,840 --> 00:01:38,929 S1: into VS code. I wish it was actually a full editor, 26 00:01:38,930 --> 00:01:44,390 S1: just like cursor and windsurf are. So Klein people. If 27 00:01:44,390 --> 00:01:48,230 S1: you're listening, please make an editor. And my personal AI 28 00:01:48,230 --> 00:01:52,160 S1: infrastructure is getting absolutely insane right now. I'm going to 29 00:01:52,160 --> 00:01:54,500 S1: have something to show on this soon. I'm going to 30 00:01:54,500 --> 00:01:57,080 S1: do a demo of like how this all works together, 31 00:01:57,080 --> 00:01:59,870 S1: but basically it's one main agent and then all my 32 00:01:59,870 --> 00:02:05,060 S1: personal services are sitting behind it that do different things, 33 00:02:05,060 --> 00:02:09,320 S1: and they're all able to access different services that do 34 00:02:09,350 --> 00:02:13,040 S1: output as well. So there's like you send the request 35 00:02:13,070 --> 00:02:17,840 S1: into the agent, the agent, if I tell it, send me, 36 00:02:17,870 --> 00:02:21,170 S1: you know, test a website and then send me an 37 00:02:21,169 --> 00:02:25,040 S1: email for it, it'll do the two different services. One 38 00:02:25,040 --> 00:02:27,470 S1: is to test the website and the other one is 39 00:02:27,470 --> 00:02:31,040 S1: to send an email. And again it just naturally parsed that. 40 00:02:31,040 --> 00:02:33,649 S1: And I just sent that from my phone. Right. So 41 00:02:33,650 --> 00:02:37,429 S1: the whole idea is I'm just using my regular tech infrastructure, 42 00:02:37,430 --> 00:02:40,430 S1: which is my iPhone, and I send a request into 43 00:02:40,460 --> 00:02:42,230 S1: my agent, and this is what it's going to be 44 00:02:42,230 --> 00:02:45,590 S1: in the future too. I'm just sort of hacking it 45 00:02:45,590 --> 00:02:48,440 S1: together with my own services. Pretty soon it'll just be 46 00:02:48,440 --> 00:02:52,760 S1: built into Google and built into the iPhone and stuff 47 00:02:52,760 --> 00:02:55,190 S1: like that. But right now I'm just making it happen 48 00:02:55,190 --> 00:03:00,359 S1: through like shortcuts and a bunch of AI services, but 49 00:03:00,360 --> 00:03:04,800 S1: the overall effect is quite awesome and you have to 50 00:03:04,800 --> 00:03:08,550 S1: see this thing actually working. It's very much in line 51 00:03:08,550 --> 00:03:12,570 S1: with what I've been talking about. Uh, in the predictable 52 00:03:12,600 --> 00:03:16,590 S1: AI post, which is this link here where it's all going, 53 00:03:16,620 --> 00:03:21,269 S1: put out a concise explanation of model context protocol, MCP 54 00:03:21,270 --> 00:03:24,570 S1: servers and why everyone's excited about them. So I got 55 00:03:24,570 --> 00:03:28,440 S1: a full thread on that and you should go check out. And, uh, 56 00:03:28,440 --> 00:03:32,280 S1: I'm enjoying the fact that this is starting to happen, 57 00:03:32,310 --> 00:03:37,140 S1: specifically the MCP server, because in that predictable AI post, 58 00:03:37,140 --> 00:03:40,440 S1: which actually comes from the book in 2016, I talked 59 00:03:40,440 --> 00:03:45,870 S1: about the API ification of everything and specifically how businesses 60 00:03:45,870 --> 00:03:49,950 S1: would all become APIs. So this is text. From that, 61 00:03:49,980 --> 00:03:53,820 S1: most software businesses will become algorithms presented to others through 62 00:03:53,820 --> 00:03:58,470 S1: their business daemons. Many traditional businesses will continue to become 63 00:03:58,470 --> 00:04:04,170 S1: software businesses. So basically, Andreessen talked about software as eating 64 00:04:04,170 --> 00:04:06,870 S1: the world. I think AI is going to eat the world. 65 00:04:06,900 --> 00:04:10,440 S1: Like that's pretty. Everyone's saying that it doesn't really mean anything. 66 00:04:10,440 --> 00:04:13,260 S1: What I think is actually eating the world is APIs. 67 00:04:13,260 --> 00:04:16,650 S1: And the fact that AI is what's going to interact 68 00:04:16,650 --> 00:04:21,299 S1: with all these APIs. The primary interface for doing anything 69 00:04:21,300 --> 00:04:25,109 S1: is going to be APIs, because it's the AI that's 70 00:04:25,110 --> 00:04:30,030 S1: parsing them, right? What humans need to look at a 71 00:04:30,029 --> 00:04:32,549 S1: website or a list of services, or a menu of 72 00:04:32,550 --> 00:04:35,729 S1: food or whatever. What humans need for that is way 73 00:04:35,730 --> 00:04:39,540 S1: different than what a computer needs. A computer can parse 74 00:04:39,570 --> 00:04:46,080 S1: an API or an MCP service directory, you know, basically 75 00:04:46,080 --> 00:04:49,440 S1: instantly and know exactly how to use it. So if 76 00:04:49,470 --> 00:04:52,440 S1: you if you think about SEO, if you think about UI, UX, 77 00:04:52,440 --> 00:04:55,140 S1: if you think about advertisements, if you think about websites, 78 00:04:55,140 --> 00:04:59,400 S1: if you think about News media, all of that. It's 79 00:04:59,400 --> 00:05:02,970 S1: all right now focused on presenting that to humans. The 80 00:05:02,970 --> 00:05:06,270 S1: fundamental change that's happening right now is that's all switching 81 00:05:06,270 --> 00:05:11,250 S1: away from being primarily designed for humans to being primarily 82 00:05:11,250 --> 00:05:16,140 S1: designed for AI. Primarily, these interfaces will be for AI 83 00:05:16,170 --> 00:05:20,070 S1: to find a business, AI to find a service AI 84 00:05:20,100 --> 00:05:23,669 S1: to rank those services. Right. So examples will include and 85 00:05:23,670 --> 00:05:26,580 S1: this is keep in mind this text right here. This 86 00:05:26,580 --> 00:05:30,090 S1: is from 2016 when I wrote that stupid book which 87 00:05:30,089 --> 00:05:34,110 S1: you probably don't need to read anymore because it's all 88 00:05:34,110 --> 00:05:37,500 S1: in that, that post and it's got pictures which took 89 00:05:37,500 --> 00:05:40,169 S1: me like hours to make. So it's a better version 90 00:05:40,170 --> 00:05:42,930 S1: of the book that blog post is. But the key 91 00:05:42,930 --> 00:05:45,570 S1: point here is I was talking about this in 2016, 92 00:05:45,570 --> 00:05:48,510 S1: so I'm looking for a little like, I don't know, 93 00:05:48,510 --> 00:05:50,820 S1: somebody to send me a pat on the back. That 94 00:05:50,820 --> 00:05:52,830 S1: was kind of the whole purpose of writing that stupid 95 00:05:52,830 --> 00:05:56,220 S1: book was to get this stuff locked in, because I 96 00:05:56,250 --> 00:05:59,940 S1: saw this happening so clearly, even though that was way 97 00:05:59,940 --> 00:06:03,809 S1: before this AI stuff popped. Right. So what is 22 98 00:06:03,839 --> 00:06:07,920 S1: the end of 22 versus the end of 2016? That's 99 00:06:07,920 --> 00:06:11,040 S1: six years. Six years. I basically said, we're going to 100 00:06:11,040 --> 00:06:15,840 S1: have digital assistants, which are AI powered. I talked about 101 00:06:15,839 --> 00:06:18,420 S1: the AI component right there in there. It's going to 102 00:06:18,450 --> 00:06:21,450 S1: know everything about us. We're going to talk to our Da. 103 00:06:21,480 --> 00:06:24,810 S1: Our Da is going to read all the services. Every 104 00:06:24,839 --> 00:06:26,730 S1: human is going to have a demon. Not every human, 105 00:06:26,730 --> 00:06:29,070 S1: but most humans. A lot of a lot of humans 106 00:06:29,070 --> 00:06:31,469 S1: are going to have demons. But most importantly, all the 107 00:06:31,470 --> 00:06:35,790 S1: businesses have demons. We talk to our Da. The Da 108 00:06:35,970 --> 00:06:39,570 S1: infers what we probably want. We don't even need to 109 00:06:39,600 --> 00:06:43,560 S1: talk to it. It's the one interacting with the world, right? 110 00:06:43,560 --> 00:06:47,250 S1: And there's a millions upon millions of other DA's out 111 00:06:47,250 --> 00:06:50,760 S1: there working for their principles. Some are working for businesses, 112 00:06:50,760 --> 00:06:56,400 S1: some are working for individuals, some are working for organizations, entities, Whatever. 113 00:06:56,430 --> 00:06:58,500 S1: But the point is, they're the ones doing the hard 114 00:06:58,500 --> 00:07:02,310 S1: work of parsing all these millions of demons. Right. So examples, 115 00:07:02,310 --> 00:07:06,870 S1: finding gifts customized to the exact individual organizing person. Perfect 116 00:07:06,870 --> 00:07:10,410 S1: vacation based on the four people going, navigating all the 117 00:07:10,410 --> 00:07:13,920 S1: logistics and the best way possible of that four person vacation. 118 00:07:13,950 --> 00:07:16,800 S1: Determining the current mood of a person or location based 119 00:07:16,800 --> 00:07:20,070 S1: on what is known about them. Finding the optimal route 120 00:07:20,070 --> 00:07:23,280 S1: from one place to another. Determining the best way to 121 00:07:23,310 --> 00:07:27,300 S1: charge a customer based on evolving competition, logistics, and conditions 122 00:07:27,300 --> 00:07:31,170 S1: on the ground. All these tasks will be services available online. 123 00:07:31,200 --> 00:07:35,040 S1: There will be several or thousands of competitors in any 124 00:07:35,040 --> 00:07:38,429 S1: particular space. There will even be services that consume and 125 00:07:38,430 --> 00:07:42,990 S1: rate those services and present their output through their own demons. Right. 126 00:07:42,990 --> 00:07:47,550 S1: So that's another demon. So my digital assistant, my personal 127 00:07:47,580 --> 00:07:51,540 S1: I will be using those services to find the best service, 128 00:07:51,540 --> 00:07:54,840 S1: to organize a trip or to play the perfect song 129 00:07:54,840 --> 00:07:58,440 S1: or to do basically anything. All these SaaS services, all 130 00:07:58,440 --> 00:08:03,300 S1: these B2C services companies. Forget a human going to the 131 00:08:03,300 --> 00:08:07,380 S1: website and clicking around. That's all done. That is all done. 132 00:08:07,380 --> 00:08:11,760 S1: Our Das will be doing that for us through their demons, 133 00:08:11,760 --> 00:08:16,350 S1: through their business demons, which maybe MCP is that, and 134 00:08:16,350 --> 00:08:19,470 S1: that's where we are. This is essentially what MSPs are 135 00:08:19,500 --> 00:08:22,920 S1: finally starting to happen. MSPs are a way to take 136 00:08:22,920 --> 00:08:28,560 S1: any application and turn it into globally available services. That's 137 00:08:28,560 --> 00:08:32,489 S1: why this MCP thing is so exciting. It's taking any 138 00:08:32,520 --> 00:08:38,940 S1: application and presenting it to I. So I like what 139 00:08:38,940 --> 00:08:41,430 S1: we're doing now is we're taking IDs and we're loading 140 00:08:41,429 --> 00:08:44,520 S1: in these marketplaces. So we have the ability to manually 141 00:08:44,520 --> 00:08:48,840 S1: choose these things. But pretty soon our IDs and more importantly, 142 00:08:48,840 --> 00:08:51,929 S1: our digital assistants are just going to there's going to 143 00:08:51,929 --> 00:08:57,099 S1: be natural, like Vetted, high quality lists of these MSPs 144 00:08:57,100 --> 00:08:59,740 S1: of these demons out there, and they're going to be 145 00:08:59,740 --> 00:09:02,260 S1: super high quality. They're going to be rated for quality. 146 00:09:02,260 --> 00:09:04,059 S1: They're going to be rated for uptime. They're going to 147 00:09:04,059 --> 00:09:06,969 S1: be rated for, you know, how good the results are 148 00:09:06,970 --> 00:09:10,120 S1: and everything. So my Da is just going to be 149 00:09:10,120 --> 00:09:14,140 S1: constantly using all of those, right. So when I'm like 150 00:09:14,140 --> 00:09:19,360 S1: hey I've got sniffles. It's going to be find the best, uh, pharmacy, 151 00:09:19,390 --> 00:09:23,620 S1: find the best medication, do all the research, um, find 152 00:09:23,620 --> 00:09:27,459 S1: the best delivery thing and within, you know, like six minutes, 153 00:09:27,550 --> 00:09:31,090 S1: somebody drops it off on my porch, right? It's it's 154 00:09:31,090 --> 00:09:34,240 S1: going to be that simple because it made 400 different 155 00:09:34,240 --> 00:09:38,620 S1: requests to multiple different demons to be able to do 156 00:09:38,620 --> 00:09:41,080 S1: the research and figure that out and find the best one. 157 00:09:41,110 --> 00:09:43,870 S1: Turns out I paid like a tiny amount of money 158 00:09:43,870 --> 00:09:47,650 S1: for that, or whatever the the amount was. Probably won't 159 00:09:47,650 --> 00:09:51,190 S1: even be a human delivery. It'll be like a Waymo delivery. Anyway, 160 00:09:51,280 --> 00:09:54,850 S1: bottom bottom line is we all get digital assistance. Everything 161 00:09:54,850 --> 00:09:58,450 S1: gets an API, including people, objects and businesses. Our Das 162 00:09:58,480 --> 00:10:01,569 S1: are the ones interacting with the world on our behalf, 163 00:10:01,570 --> 00:10:04,600 S1: because there will be billions of APIs or demons out 164 00:10:04,600 --> 00:10:08,679 S1: there to interact with, and our Das will. Ultimately, this 165 00:10:08,679 --> 00:10:11,680 S1: is the AR piece, which is not quite happening yet, 166 00:10:11,679 --> 00:10:14,770 S1: but everyone's building it. It will display in our are 167 00:10:14,800 --> 00:10:17,890 S1: the best version of of the world to look at, right? 168 00:10:17,920 --> 00:10:20,320 S1: If we're in a security situation, it will turn to 169 00:10:20,350 --> 00:10:23,260 S1: like a security filter for in like a dating situation, 170 00:10:23,260 --> 00:10:25,630 S1: it will turn into like a dating filter. If we're 171 00:10:25,630 --> 00:10:31,600 S1: in like a library, it'll like. Overlay interesting topics based 172 00:10:31,600 --> 00:10:34,000 S1: on the books that we're currently looking at. It'll be like, oh, 173 00:10:34,030 --> 00:10:37,569 S1: you know, Jason really liked this book and Chris really 174 00:10:37,570 --> 00:10:41,440 S1: liked that book or whatever. Or, hey, you were reading 175 00:10:41,440 --> 00:10:44,410 S1: something about this. This book will go into it in 176 00:10:44,410 --> 00:10:49,480 S1: more detail, right? So contextual display of information within your AR. 177 00:10:49,510 --> 00:10:52,930 S1: Everyone's currently building Das MCH and lots of other projects 178 00:10:52,929 --> 00:10:55,330 S1: are APIs for everything, and lots of people are working 179 00:10:55,330 --> 00:10:58,780 S1: on the AR stuff as well, like meta and I'm 180 00:10:58,780 --> 00:11:04,870 S1: sure Apple eventually. All right. Security, cyber security critical PHP, CGI, 181 00:11:04,900 --> 00:11:09,670 S1: remote code execution vulnerability is being mass exploited against windows systems. 182 00:11:09,700 --> 00:11:12,730 S1: It was originally a bunch of Japanese targets. Now it's gone. 183 00:11:12,730 --> 00:11:17,380 S1: Global help Net Security put together a list of 28 184 00:11:17,500 --> 00:11:23,500 S1: cyber security positions across the US. President is nominated Sean Plinky, 185 00:11:23,530 --> 00:11:27,640 S1: Plonky veteran with experience at US Cyber Command in the 186 00:11:27,640 --> 00:11:31,689 S1: previous Trump administration to head Cisa. So Cisa has a 187 00:11:31,690 --> 00:11:35,739 S1: new leader. Malicious poisoning of AI narratives. So TechCrunch put 188 00:11:35,740 --> 00:11:40,210 S1: out a report that Russian propaganda was there was a 189 00:11:40,210 --> 00:11:43,720 S1: basically a a network called Pravda, which is a Russian 190 00:11:43,720 --> 00:11:48,970 S1: propaganda network, is flooding the web with fake news about, 191 00:11:49,000 --> 00:11:53,229 S1: I assume it's like, uh, Russian narratives around Ukraine and 192 00:11:53,230 --> 00:11:55,450 S1: a whole bunch of stuff that just Russia wants everyone 193 00:11:55,450 --> 00:11:59,470 S1: to believe. This is crazy. This is an actual AI 194 00:11:59,500 --> 00:12:04,179 S1: poisoning attack. Real world because everyone was worried about you. 195 00:12:04,179 --> 00:12:07,599 S1: Poison the training data. Oh, but it's like, well, how 196 00:12:07,630 --> 00:12:09,250 S1: are you going to break into the company and poison 197 00:12:09,280 --> 00:12:14,170 S1: the training data? Turns out these bots, like OpenAI, like anthropic, 198 00:12:14,200 --> 00:12:18,429 S1: they're all constantly parsing the internet. You don't have to 199 00:12:18,460 --> 00:12:22,990 S1: poison the original pre-training stuff. Well, actually, that's being poisoned 200 00:12:22,990 --> 00:12:26,140 S1: as well, because every time you do a pre-training, you 201 00:12:26,140 --> 00:12:29,410 S1: have to re harvest from the internet. If you change 202 00:12:29,410 --> 00:12:33,640 S1: the internet, then you have poisoned it, right? So, for example, 203 00:12:33,700 --> 00:12:37,330 S1: if this this group, Pravda is trying to change the 204 00:12:37,330 --> 00:12:43,810 S1: narrative around something. So let's say this company, Pravda successfully 205 00:12:43,809 --> 00:12:49,270 S1: saturates the entire world with the idea that Ukraine attacked first. Okay. 206 00:12:49,300 --> 00:12:51,819 S1: Ukraine came over the border and killed a bunch of 207 00:12:51,820 --> 00:12:55,750 S1: innocent people. And what Russia did was they did a counterattack. 208 00:12:55,750 --> 00:12:58,900 S1: And this is purely defensive. I'm not saying that's that's 209 00:12:58,900 --> 00:13:02,710 S1: what the propaganda said. Let's just take it as an example. Well, 210 00:13:02,710 --> 00:13:06,429 S1: if they put out millions of articles like that across 211 00:13:06,429 --> 00:13:09,130 S1: the entire internet, and then you parse the internet and 212 00:13:09,130 --> 00:13:12,760 S1: you put it into your training data for the next 213 00:13:12,760 --> 00:13:16,390 S1: version of sonnet or the next version of, uh, you know, 214 00:13:16,420 --> 00:13:22,689 S1: O4 or GPT five or whatever it is. Well, how 215 00:13:22,690 --> 00:13:24,730 S1: is the thing? How was I going to know that's 216 00:13:24,730 --> 00:13:28,090 S1: not real? It's it's parsing the internet to figure out 217 00:13:28,090 --> 00:13:32,949 S1: what's real. So this is an actual AI data poisoning attack. 218 00:13:33,160 --> 00:13:36,910 S1: Ten leading chat bots were tested. Chatbots were tested for 219 00:13:36,910 --> 00:13:44,500 S1: susceptibility to Russian disinformation. Pravda published 3.6 million misleading articles 220 00:13:44,500 --> 00:13:50,230 S1: just in 2024, 3.6 million misleading articles and news Knausgaard's 221 00:13:50,230 --> 00:13:55,840 S1: analysis found that the chatbots collectively repeated false information a 222 00:13:55,840 --> 00:14:00,880 S1: third of the time. That's insane. Real world AI data poisoning. 223 00:14:01,120 --> 00:14:03,579 S1: So it used to be about SEO poisoning. Now it's 224 00:14:03,580 --> 00:14:06,819 S1: about flooding the internet with marketing or propaganda that will 225 00:14:06,820 --> 00:14:10,030 S1: get you picked up by up by bots. Now I 226 00:14:10,030 --> 00:14:13,240 S1: say marketing and propaganda. It's the same thing, right? Let's 227 00:14:13,240 --> 00:14:17,080 S1: say you have something. You have a an endpoint product 228 00:14:17,080 --> 00:14:20,800 S1: that's nowhere near as good as CrowdStrike, but you hire 229 00:14:20,800 --> 00:14:25,600 S1: Pravda and you blast out to the internet that, you know, 230 00:14:25,630 --> 00:14:30,580 S1: secure point is way, way better than CrowdStrike. And all 231 00:14:30,580 --> 00:14:33,280 S1: the bots parse this thing. And a third of the time, 232 00:14:33,280 --> 00:14:38,110 S1: Secure Point comes back as better than CrowdStrike. That would 233 00:14:38,110 --> 00:14:40,690 S1: be massive. I mean, that that would get you promoted 234 00:14:40,690 --> 00:14:43,930 S1: on a marketing team, because guess what? The new search 235 00:14:43,930 --> 00:14:46,990 S1: engine is actually AI. A lot, a lot of people 236 00:14:46,990 --> 00:14:53,630 S1: are actually asking AI to compare products and to come 237 00:14:53,660 --> 00:14:56,660 S1: back because. Because it knows that the AI is read 238 00:14:56,660 --> 00:15:01,910 S1: everything and is is summarizing the results of all of that. Well, 239 00:15:01,910 --> 00:15:06,530 S1: that changes that changes significantly when now everyone's going to 240 00:15:06,530 --> 00:15:10,490 S1: switch to this. So, so all these all these LLM 241 00:15:10,520 --> 00:15:13,730 S1: providers are going to have to figure out a way 242 00:15:13,730 --> 00:15:17,030 S1: to filter out poisoning. I don't see an easy way 243 00:15:17,030 --> 00:15:19,100 S1: to do it. You're going to need a lot of 244 00:15:19,100 --> 00:15:22,850 S1: really smart I just to do the data cleansing for 245 00:15:22,850 --> 00:15:26,630 S1: this type of poisoning. All right. Poland's space agency Polska 246 00:15:26,660 --> 00:15:33,500 S1: has taken its systems offline after detecting unauthorized network access. And, uh. 247 00:15:33,500 --> 00:15:37,250 S1: And they're blaming Russian cyber campaigns against the country. Japanese 248 00:15:37,250 --> 00:15:40,520 S1: telecom giant NTT got breached and exposed corporate data for 249 00:15:40,520 --> 00:15:45,410 S1: around 18,000 companies. AI is targeting of enemies. US is 250 00:15:45,410 --> 00:15:48,860 S1: using AI to scan social media and revoke visas of 251 00:15:48,860 --> 00:15:53,960 S1: foreign students it believes support Hamas or other terrorist groups. 252 00:15:54,080 --> 00:15:57,140 S1: The reason I think this is worrying because this this 253 00:15:57,140 --> 00:15:58,850 S1: might be a good use of it, right? But the 254 00:15:58,850 --> 00:16:02,930 S1: problem is, once you use it once and everyone's like, oh, 255 00:16:02,960 --> 00:16:05,540 S1: that's okay. I mean, there's a saying that says first 256 00:16:05,570 --> 00:16:08,870 S1: they come for so-and-so, then they come for so-and-so, and 257 00:16:08,870 --> 00:16:11,510 S1: you say it's okay, and then they come for so-and-so, 258 00:16:11,510 --> 00:16:13,850 S1: and you say, oh, that's still okay because it's not me. 259 00:16:13,850 --> 00:16:16,640 S1: But eventually they come for you. And I'm not saying 260 00:16:16,640 --> 00:16:19,190 S1: that's going to happen. I'm just saying whenever you have 261 00:16:19,190 --> 00:16:22,490 S1: a thing that's like, hey, we're looking for people like 262 00:16:22,490 --> 00:16:26,270 S1: this because they're bad and the people are like, well, 263 00:16:26,270 --> 00:16:31,370 S1: that's okay. And everyone like allows that type of Gestapo behavior. 264 00:16:31,370 --> 00:16:36,530 S1: It's slippery, very slippery, and it's way worse when it 265 00:16:36,530 --> 00:16:38,630 S1: would have taken a lot of research in the past 266 00:16:38,630 --> 00:16:41,510 S1: to actually go and find these people. Right. You got 267 00:16:41,540 --> 00:16:43,460 S1: you got to parse all these videos. You got to 268 00:16:43,490 --> 00:16:46,370 S1: you got to, you know, go through all the blog posts, 269 00:16:46,370 --> 00:16:48,800 S1: you got to find people You basically have to hire 270 00:16:48,830 --> 00:16:52,160 S1: a bunch of researchers, researchers to do this. Well, what 271 00:16:52,160 --> 00:16:54,980 S1: happens if you just spin up a really powerful AI 272 00:16:55,010 --> 00:16:58,100 S1: with a bunch of tools and say, hey, go parse 273 00:16:58,100 --> 00:17:00,740 S1: the internet for the last two years or during this 274 00:17:00,740 --> 00:17:05,060 S1: six month period and find everybody who in the crowd 275 00:17:05,060 --> 00:17:07,970 S1: was waving a Hamas flag, or in the crowd who 276 00:17:07,970 --> 00:17:11,840 S1: was waving whatever flag for the thing that I currently 277 00:17:11,840 --> 00:17:13,970 S1: don't like. And this could be a left leaning thing. 278 00:17:13,970 --> 00:17:16,490 S1: This could be a right leaning thing. Doesn't matter. You 279 00:17:16,490 --> 00:17:19,910 S1: could target an enemy and basically go after them. That 280 00:17:19,910 --> 00:17:24,110 S1: is that is very serious stuff. And the better the 281 00:17:24,109 --> 00:17:26,480 S1: AI gets, the better it's the faster it's going to 282 00:17:26,480 --> 00:17:29,900 S1: be able to do that. I mean, you could be like, look, hey, 283 00:17:29,930 --> 00:17:32,420 S1: we got a Swat team that's ready to go find 284 00:17:32,420 --> 00:17:35,840 S1: me the best possible enemy of me and my friends, 285 00:17:35,960 --> 00:17:40,429 S1: and find me a reason in all of their in 286 00:17:40,460 --> 00:17:43,400 S1: the last six months to go after them. And boom, 287 00:17:43,400 --> 00:17:46,580 S1: you send the Swat team like that. That's a very 288 00:17:46,619 --> 00:17:50,609 S1: powerful weapon, and you have to be very careful who 289 00:17:50,609 --> 00:17:54,990 S1: is actually wielding that thing. DOJ charged 12 Chinese nationals 290 00:17:54,990 --> 00:17:58,890 S1: for state backed hacking operations, including two government officers, eight 291 00:17:58,890 --> 00:18:02,010 S1: employees of a hacking company called I-soon and a few 292 00:18:02,010 --> 00:18:06,930 S1: members of APT 27. Microsoft Threat Intelligence says China linked 293 00:18:06,960 --> 00:18:13,199 S1: Silk Typhoon hacking group is shifting from compromise to actually 294 00:18:13,200 --> 00:18:17,280 S1: going after it. Supply chains as an entry point to 295 00:18:17,310 --> 00:18:22,830 S1: corporate networks, national security. Russia and China and probably others 296 00:18:22,859 --> 00:18:27,270 S1: are targeting Doge fired US federal workers with special focus 297 00:18:27,270 --> 00:18:30,210 S1: on those with security clearances. So basically, a bunch of 298 00:18:30,210 --> 00:18:33,690 S1: people with clearances who know things get laid off and 299 00:18:33,690 --> 00:18:37,290 S1: they're going after them, especially if they're angry, right? That's 300 00:18:37,560 --> 00:18:42,450 S1: a lot of vulnerability. China announced another 7.2% increase to 301 00:18:42,480 --> 00:18:47,960 S1: its military budget, taking it to $245 $45 billion a year. 302 00:18:47,990 --> 00:18:54,260 S1: The Polish president, Andrzej Duda, is pushing NATO, NATO members 303 00:18:54,260 --> 00:18:57,770 S1: to increase defense spending to 3% of GDP. Yeah. China 304 00:18:57,800 --> 00:19:01,820 S1: is trying to bump it by another 7.2%, and we're 305 00:19:01,820 --> 00:19:05,869 S1: trying to get, uh, NATO members to 3%. Oliver Carroll 306 00:19:05,869 --> 00:19:11,840 S1: reports that the US has cut off critical Himars, intelligence, Himars, 307 00:19:11,840 --> 00:19:16,220 S1: Himars intelligence to Ukraine, reducing their ability to defend against 308 00:19:16,220 --> 00:19:21,530 S1: Russian forces. I so manas broke the internet last week. 309 00:19:21,560 --> 00:19:27,859 S1: Big hype is, uh, it's kind of like AGI similar. Um, 310 00:19:27,890 --> 00:19:31,670 S1: they didn't say it was AGI. The video that they 311 00:19:31,670 --> 00:19:34,760 S1: put out was like, hey, it's leaning towards AGI. It's 312 00:19:34,760 --> 00:19:38,120 S1: a little bit like AGI. We're heading in that direction. 313 00:19:38,150 --> 00:19:41,300 S1: But it's basically like Claude or OpenAI, whatever they have 314 00:19:41,330 --> 00:19:44,389 S1: behind the thing. But it's just way smarter. It's doing 315 00:19:44,390 --> 00:19:48,469 S1: a lot more manual work that that would have needed 316 00:19:48,470 --> 00:19:50,990 S1: to be done by a human in the past. And 317 00:19:50,990 --> 00:19:52,910 S1: it's very much in line with what I was talking 318 00:19:52,910 --> 00:19:55,760 S1: about the other day where, um, I don't think that 319 00:19:55,760 --> 00:19:59,450 S1: we need smarter models, actually, to get to AGI. I 320 00:19:59,450 --> 00:20:04,040 S1: think we just need better orchestration of tooling and more 321 00:20:04,040 --> 00:20:07,070 S1: working memory for the models we already have. And it 322 00:20:07,070 --> 00:20:11,270 S1: seems to me like madness is just is just that. Right? 323 00:20:11,270 --> 00:20:15,680 S1: It's probably Claude underneath. And a lot of people have been, uh, 324 00:20:15,710 --> 00:20:19,760 S1: you know, basically hypothesizing that that's the case. We don't 325 00:20:19,760 --> 00:20:23,030 S1: know for sure, but it could just be Claude with 326 00:20:23,030 --> 00:20:26,480 S1: better orchestration. Really want people to think deeply about this 327 00:20:26,480 --> 00:20:31,040 S1: general intelligence competence might be already possible, given how smart 328 00:20:31,040 --> 00:20:36,020 S1: something like O1 Pro or Sonnet 37 already are. It 329 00:20:36,020 --> 00:20:40,340 S1: already has the intelligence to understand stuff like millions or 330 00:20:40,340 --> 00:20:43,040 S1: billions of questions that you could ask it. It's smart 331 00:20:43,040 --> 00:20:45,899 S1: enough to generalize allies answers, or you can give it 332 00:20:45,900 --> 00:20:50,100 S1: situations like brand new situation and based on everything it 333 00:20:50,100 --> 00:20:53,729 S1: knows about these other things, it can give you the 334 00:20:53,730 --> 00:20:57,149 S1: direction or a solution, you know, for that new problem 335 00:20:57,150 --> 00:21:01,020 S1: that you have. That's generalization, right? There might be smart enough. 336 00:21:01,020 --> 00:21:04,770 S1: The problem is its eyes and its hands and it 337 00:21:04,770 --> 00:21:09,900 S1: getting confused after doing too many things because of context windows. Right. 338 00:21:09,900 --> 00:21:13,950 S1: And just overall coordination and orchestration of the thing that 339 00:21:13,950 --> 00:21:16,680 S1: it's doing to try to accomplish a task. Seems like 340 00:21:16,680 --> 00:21:19,800 S1: Manas did that better, but I think that's all that's 341 00:21:19,830 --> 00:21:24,420 S1: needed is that orchestration. So the point is people are like, well, 342 00:21:24,540 --> 00:21:29,730 S1: AI is not getting smart enough fast enough to actually 343 00:21:29,730 --> 00:21:33,150 S1: move us towards AGI. We don't need that. We don't 344 00:21:33,150 --> 00:21:39,119 S1: need smarter models. We need larger context windows, more tools 345 00:21:39,119 --> 00:21:44,459 S1: for the agents to have and better orchestration. That by itself, 346 00:21:44,460 --> 00:21:48,030 S1: I think gets us there to AGI. And I've got 347 00:21:48,030 --> 00:21:50,790 S1: another post, which yeah, I'm not going to pull it up, 348 00:21:50,790 --> 00:21:54,090 S1: but I've got another post that basically talks about exactly 349 00:21:54,090 --> 00:21:57,330 S1: what AGI is going to look like under my definition, 350 00:21:57,450 --> 00:22:01,439 S1: which is, um, my, my definition is basically can it 351 00:22:01,440 --> 00:22:05,580 S1: do the work of a knowledge worker making $80,000 in 352 00:22:05,580 --> 00:22:08,760 S1: the year 2022? And I say 22 because that was 353 00:22:08,790 --> 00:22:13,619 S1: like pre AI or whatever. So the marketing for the 354 00:22:13,619 --> 00:22:18,390 S1: first AGI is going to be something like welcome to 355 00:22:18,420 --> 00:22:23,760 S1: new hire. New hire will join your team on Monday morning. 356 00:22:23,790 --> 00:22:26,850 S1: The day that they start they will do the onboarding. 357 00:22:26,880 --> 00:22:29,850 S1: They will read all the slack messages. They will read 358 00:22:29,850 --> 00:22:32,670 S1: all the documentation. They will meet with their manager. They 359 00:22:32,670 --> 00:22:35,130 S1: will take direction. They will reach out to the team 360 00:22:35,130 --> 00:22:38,310 S1: members and introduce themselves, and they will start to do 361 00:22:38,310 --> 00:22:42,119 S1: the work that they are assigned. And most importantly, if 362 00:22:42,119 --> 00:22:45,150 S1: that work changes because of a reorg, or because they 363 00:22:45,150 --> 00:22:47,879 S1: bought some new company or whatever. They get new tasks, 364 00:22:47,910 --> 00:22:51,240 S1: new goals, new work they have to do. They will 365 00:22:51,240 --> 00:22:54,330 S1: discard their previous stuff. They will learn the new stuff, 366 00:22:54,330 --> 00:22:57,630 S1: they will read the new docs, they will take new direction. 367 00:22:57,630 --> 00:22:59,370 S1: They will learn what they have to, and they'll start 368 00:22:59,369 --> 00:23:01,500 S1: doing that work. And if their manager is like, hey, 369 00:23:01,500 --> 00:23:03,780 S1: you're doing pretty good here, but you need to improve 370 00:23:03,780 --> 00:23:07,320 S1: on this just like a human. They'll be like, okay, thanks, boss. 371 00:23:07,320 --> 00:23:10,530 S1: I'll make that adjustment. I'll make sure you get a 372 00:23:10,530 --> 00:23:13,800 S1: status report every day and whatever it is they were 373 00:23:13,800 --> 00:23:16,500 S1: asked for this company, new hire, which is not a 374 00:23:16,500 --> 00:23:18,780 S1: real company, by the way, but it's going to be 375 00:23:18,780 --> 00:23:21,840 S1: something exactly like this. That's how they're going to market it. 376 00:23:21,869 --> 00:23:27,300 S1: Treat it exactly like a regular employee. That is going 377 00:23:27,300 --> 00:23:31,380 S1: to be the moment, in my opinion, that we've reached AGI, 378 00:23:31,410 --> 00:23:33,929 S1: because if that tool, by the way, there's going to 379 00:23:33,930 --> 00:23:36,150 S1: be a bunch of tools that say they can do 380 00:23:36,150 --> 00:23:38,550 S1: this and then people are going to implement them and 381 00:23:38,550 --> 00:23:40,410 S1: they're going to be super stupid. They're going to be 382 00:23:40,410 --> 00:23:42,960 S1: more waste than it's worth. It's just not going to work. 383 00:23:42,960 --> 00:23:46,500 S1: But when this happens and let's say it's this company 384 00:23:46,500 --> 00:23:49,830 S1: called New Hire. Um, when that happens, let's say new 385 00:23:49,859 --> 00:23:53,820 S1: hire takes off and it just it goes slow for 386 00:23:53,820 --> 00:23:56,010 S1: a second and all of a sudden everyone starts saying, 387 00:23:56,130 --> 00:24:01,260 S1: oh my God, I employed ten new hires, and they're 388 00:24:01,260 --> 00:24:05,790 S1: doing the work of like 50 regular employees. And I 389 00:24:05,790 --> 00:24:09,420 S1: just increased my subscription because I'm growing my business faster 390 00:24:09,420 --> 00:24:12,600 S1: than ever. I employed my new hires in marketing. I 391 00:24:12,600 --> 00:24:15,900 S1: employed them in sales, I employed them in administration. I 392 00:24:15,900 --> 00:24:20,340 S1: employed them in security administration in my SOC. Whatever the 393 00:24:20,340 --> 00:24:24,030 S1: moment that happens, I'm calling that the AGI moment. I 394 00:24:24,030 --> 00:24:28,020 S1: believe that is the actual AGI moment, because there is 395 00:24:28,020 --> 00:24:32,640 S1: no better definition of general intelligence than humans. That's what 396 00:24:32,640 --> 00:24:36,360 S1: we're basing this whole entire thing on, is humans ability 397 00:24:36,359 --> 00:24:40,530 S1: to generalize, generalize when it's working. Right. The fact that 398 00:24:40,530 --> 00:24:43,840 S1: you could change up context. Say, hey, that's not your 399 00:24:43,840 --> 00:24:47,409 S1: job anymore. Here's your new job. That's the best definition 400 00:24:47,410 --> 00:24:51,399 S1: we have is a knowledge worker constantly dealing with change, 401 00:24:51,400 --> 00:24:56,139 S1: constantly dealing with one offs and asks and adjustments. So 402 00:24:56,140 --> 00:25:01,810 S1: that's the perfect definition also for actual AGI artificial general intelligence. 403 00:25:01,810 --> 00:25:04,810 S1: And if it could do the job of an average 404 00:25:04,810 --> 00:25:08,620 S1: knowledge worker I would say that hits the bar. So 405 00:25:08,619 --> 00:25:10,930 S1: that's why I'm using it for a definition. And I'm 406 00:25:10,930 --> 00:25:13,240 S1: saying that when a product comes out that can do 407 00:25:13,240 --> 00:25:16,660 S1: this and it's actually doing it, that will be the moment. 408 00:25:16,660 --> 00:25:22,000 S1: And my prediction is late 2025 at the earliest. This 409 00:25:22,000 --> 00:25:25,990 S1: is my modified prediction. My overall prediction, which I said 410 00:25:25,990 --> 00:25:31,750 S1: in 23 was 25 to 28. So my modified prediction 411 00:25:31,750 --> 00:25:36,730 S1: is late 25 to early 27. So it's still right 412 00:25:36,730 --> 00:25:39,790 S1: there inside that window. But I think it's getting pretty 413 00:25:39,790 --> 00:25:44,409 S1: close and things like manis are starting to show signs 414 00:25:44,410 --> 00:25:48,520 S1: of it. All right. New model q w q 32 415 00:25:48,520 --> 00:25:52,990 S1: billion parameters. This thing is outperforming or performing just as 416 00:25:52,990 --> 00:25:59,500 S1: good as 671 billion parameter deep seek 32 billion versus 417 00:25:59,500 --> 00:26:04,000 S1: 671 billion. And they are right on par. This is 418 00:26:04,000 --> 00:26:09,070 S1: the absolute most impressive local model I've ever used. I 419 00:26:09,070 --> 00:26:11,680 S1: test out all these big ones when they come out. 420 00:26:11,920 --> 00:26:14,200 S1: I've got right in my other room. I've got two 421 00:26:14,230 --> 00:26:17,619 S1: 49 seconds over there and I fire them up and 422 00:26:17,619 --> 00:26:22,630 S1: I test this thing. This I mean, it performs honestly 423 00:26:22,630 --> 00:26:29,410 S1: very similar to like a cloud or an OpenAI, um, 424 00:26:29,440 --> 00:26:32,890 S1: like an O one or an O one mini. It's really, 425 00:26:32,890 --> 00:26:36,609 S1: really good. Um, I used it with some, uh, story 426 00:26:36,640 --> 00:26:39,580 S1: writing stuff. I'm trying to write a fiction story about some, 427 00:26:39,609 --> 00:26:43,360 S1: like an eye feature, getting pretty excited about that, but 428 00:26:43,359 --> 00:26:46,240 S1: I was using it to help me write, and it's 429 00:26:46,240 --> 00:26:50,109 S1: doing things that no previous model could possibly do. Um, 430 00:26:50,109 --> 00:26:54,369 S1: just vast jumps in quality in the output that I'm getting. 431 00:26:54,369 --> 00:26:57,760 S1: So really excited about that. You should go check it out. 432 00:26:57,760 --> 00:27:00,490 S1: You could run it with a llama llama run q 433 00:27:00,490 --> 00:27:03,369 S1: w q. That's all you have to do. It'll download it, 434 00:27:03,400 --> 00:27:05,649 S1: it'll run it, and you could test it out. And 435 00:27:05,650 --> 00:27:09,670 S1: it's a thinking model built with reinforcement learning, which means 436 00:27:09,670 --> 00:27:12,070 S1: you're going to see the thinking instructions as part of 437 00:27:12,070 --> 00:27:15,370 S1: the output before it starts outputting the real thing. Myo 438 00:27:15,400 --> 00:27:20,530 S1: addresses eye hallucinations, so doctors are very upset and worried 439 00:27:20,530 --> 00:27:23,650 S1: about hallucinations for good reason because it's life and death 440 00:27:23,650 --> 00:27:26,650 S1: for them. Right? So reverse Rag is a technique that 441 00:27:26,650 --> 00:27:31,780 S1: they're using to basically figure out the sources for the 442 00:27:31,780 --> 00:27:36,340 S1: data that they return. So they don't just take hallucinations 443 00:27:36,340 --> 00:27:38,620 S1: or they don't just take outputs and assume it's good. 444 00:27:38,619 --> 00:27:41,200 S1: What they do is they take each particular one and 445 00:27:41,200 --> 00:27:44,200 S1: they go and research it and find sources to it. 446 00:27:44,200 --> 00:27:47,680 S1: And then they say, okay, this is now trustworthy. And 447 00:27:47,680 --> 00:27:49,660 S1: I've been talking about this for, I don't know, six 448 00:27:49,660 --> 00:27:52,840 S1: months or now or however long. I did a talk 449 00:27:52,869 --> 00:27:56,740 S1: about this at some conference. And basically what we're talking 450 00:27:56,740 --> 00:28:00,609 S1: about here is a pipeline of trust that has to 451 00:28:00,609 --> 00:28:05,740 S1: happen for high trust output. So if you're in medical 452 00:28:05,740 --> 00:28:10,419 S1: or you're in military or you're in, uh, law or something, 453 00:28:10,420 --> 00:28:14,679 S1: if you're in something where results actually matter, you have 454 00:28:14,680 --> 00:28:18,100 S1: to have a system that comes after the AI output, 455 00:28:18,130 --> 00:28:21,160 S1: which is like a series of steps for validation, and 456 00:28:21,160 --> 00:28:23,500 S1: only after it goes through those steps and gets a 457 00:28:23,500 --> 00:28:26,710 S1: green check mark. Then you can trust the thing, right? 458 00:28:26,740 --> 00:28:29,230 S1: I mean, humans need that, by the way. I mean, 459 00:28:29,650 --> 00:28:33,520 S1: humans don't put out stuff that's trustworthy, right? So I 460 00:28:33,520 --> 00:28:35,949 S1: think AI is going to be actually much better at 461 00:28:35,950 --> 00:28:40,510 S1: doing this than, uh, most human doctors. You still don't 462 00:28:40,510 --> 00:28:43,420 S1: get that. Like the doctor touch and the doctor intuition 463 00:28:43,420 --> 00:28:46,540 S1: and the the human aspect. So it's it's not better 464 00:28:46,540 --> 00:28:48,910 S1: in all ways, but in terms of, like, putting out 465 00:28:48,910 --> 00:28:52,720 S1: the best possible diagnosis and stuff like that, I think 466 00:28:52,720 --> 00:28:56,020 S1: it's going to be fantastic. So kudos to Mayo for 467 00:28:56,020 --> 00:28:59,650 S1: doing that. Tyler Cowen told Dwarkesh Patel that he predicts 468 00:28:59,650 --> 00:29:04,450 S1: AI will only boost economic growth by 0.5% per year, 469 00:29:04,480 --> 00:29:08,680 S1: half a percent. I massively disagree with this and so 470 00:29:08,680 --> 00:29:13,360 S1: does Dwarkesh. Black Forest Labs, also known as replicate, is 471 00:29:13,360 --> 00:29:18,580 S1: dominating image generation. I'm using this myself. Um, I've been 472 00:29:18,580 --> 00:29:22,840 S1: using it in conjunction more with Midjourney. Midjourney is like 473 00:29:22,840 --> 00:29:24,880 S1: my go to, but I've been using a lot more 474 00:29:24,880 --> 00:29:28,810 S1: replicate and it is absolutely fantastic. Larry Page is working 475 00:29:28,810 --> 00:29:31,780 S1: on a new AI company that uses artificial intelligence to 476 00:29:31,810 --> 00:29:36,490 S1: design and manufacture optimized products. Ben Buchanan, Biden's AI advisor, 477 00:29:36,490 --> 00:29:41,110 S1: says Washington now believes AGI is possible and realizes they're 478 00:29:41,110 --> 00:29:44,590 S1: competing with tech companies for talent to regulate it. Nirvana 479 00:29:44,590 --> 00:29:48,940 S1: raised 80 million for the AI based trucking insurance platform 480 00:29:48,940 --> 00:29:53,230 S1: that uses telematics and 20 billion miles of driving data 481 00:29:53,230 --> 00:29:56,440 S1: to create better policies for drivers. And I've got a 482 00:29:56,440 --> 00:30:01,150 S1: post here on real time insurance premiums from 2016. Yeah, 483 00:30:01,180 --> 00:30:05,500 S1: I've been talking about this for a while. 2016 essentially, again, 484 00:30:05,500 --> 00:30:08,920 S1: everything with AI comes down to context. So if I'm 485 00:30:08,920 --> 00:30:12,550 S1: wearing a microphone and a camera and I give access 486 00:30:12,580 --> 00:30:15,160 S1: to that, there's like that progressive thing. Remember that progressive 487 00:30:15,190 --> 00:30:16,990 S1: box you put in your car and it can see 488 00:30:16,990 --> 00:30:19,660 S1: how you're driving? Well, imagine that it could see that 489 00:30:19,660 --> 00:30:25,060 S1: you're cussing or yelling at your spouse or kicking garbage 490 00:30:25,090 --> 00:30:30,160 S1: cans or, um, you just came from the bar. And 491 00:30:30,160 --> 00:30:33,760 S1: it knows that because it has the cameras on and monitoring. Well, 492 00:30:33,760 --> 00:30:38,350 S1: guess what? Your insurance premiums might be getting modified in 493 00:30:38,350 --> 00:30:41,080 S1: real time or close to real time. Or you might 494 00:30:41,080 --> 00:30:44,380 S1: be like, hey, if you get in this car right now, 495 00:30:44,410 --> 00:30:49,210 S1: having come from, you know, oceanians, then, um, you your 496 00:30:49,210 --> 00:30:52,090 S1: policy will be instantly cancelled, right? It's going to be 497 00:30:52,090 --> 00:30:56,680 S1: things like that. Um, all possible because of AI and context. 498 00:30:56,890 --> 00:31:01,750 S1: Simon Willison taught a Nykaa 2025 workshop on advanced web 499 00:31:01,780 --> 00:31:04,900 S1: scraping techniques. Really, really cool talk. You have to check 500 00:31:04,900 --> 00:31:07,840 S1: that one out. Okay, McDonald's rolling out a massive AI 501 00:31:07,870 --> 00:31:12,040 S1: upgrade for its 43,000 locations, including AI, drive through smart 502 00:31:12,040 --> 00:31:16,030 S1: kitchen equipment and generative AI Virtual Manager, one of the 503 00:31:16,030 --> 00:31:20,050 S1: biggest visible AI rollouts so far. Can't wait to see 504 00:31:20,050 --> 00:31:24,430 S1: how this works. I heard somebody say, I don't know 505 00:31:24,430 --> 00:31:26,560 S1: who because I don't. I wouldn't eat there because it's 506 00:31:26,560 --> 00:31:30,310 S1: really bad. But Domino's and Carl's Jr has really good AI. 507 00:31:30,340 --> 00:31:34,810 S1: So I heard technology. Apple is pushing back. It's more 508 00:31:34,810 --> 00:31:39,380 S1: personalized Siri features. So basically, the big unified Siri that 509 00:31:39,380 --> 00:31:43,040 S1: has access to all the context is being pushed back. 510 00:31:43,040 --> 00:31:46,100 S1: And a lot of people are speculating that this is 511 00:31:46,100 --> 00:31:51,680 S1: because of security issues, prompt injection against that entire context, 512 00:31:51,680 --> 00:31:54,709 S1: which what do you think is going to happen when 513 00:31:54,710 --> 00:31:58,910 S1: your assistant that you call up with this keyword can 514 00:31:58,940 --> 00:32:03,350 S1: do anything across your entire context, including all of your applications, 515 00:32:03,350 --> 00:32:07,130 S1: all your health data and everything and your interface into 516 00:32:07,130 --> 00:32:10,580 S1: that is is the English language and other languages as well, 517 00:32:10,580 --> 00:32:14,780 S1: but especially English. That is a massive attack surface to 518 00:32:14,810 --> 00:32:17,960 S1: go after. And I think Apple, rightly so, is being 519 00:32:17,960 --> 00:32:21,140 S1: very careful about that. So they push that back. One 520 00:32:21,140 --> 00:32:23,390 S1: thing to keep in mind though, is they already rolled 521 00:32:23,390 --> 00:32:28,640 S1: out tons of Apple intelligence. I already use, uh, my 522 00:32:28,640 --> 00:32:33,620 S1: Siri to actually, um, do live queries because it's connected 523 00:32:33,620 --> 00:32:37,520 S1: to ChatGPT. So don't think that we're waiting for this 524 00:32:37,520 --> 00:32:40,520 S1: thing to happen. And you don't have Apple intelligence before that. 525 00:32:40,520 --> 00:32:43,370 S1: It's already out there. Like they already rolled out like 526 00:32:43,370 --> 00:32:47,390 S1: dozens of these features, including, most importantly, the ChatGPT one. 527 00:32:47,390 --> 00:32:50,720 S1: And the fact that if you hold down your camera button, 528 00:32:50,750 --> 00:32:53,300 S1: the new camera button on the side, you can have 529 00:32:53,330 --> 00:32:57,770 S1: ChatGPT identify things with the camera. So all this to say, 530 00:32:57,800 --> 00:33:01,250 S1: Apple intelligence is like already live and already working. It's 531 00:33:01,250 --> 00:33:05,540 S1: just this unified big context, one that they haven't enabled 532 00:33:05,540 --> 00:33:08,840 S1: yet because of the security concern. Waymo is expanding its 533 00:33:08,840 --> 00:33:13,580 S1: self-driving car service in Palo Alto and some surrounding peninsula areas. 534 00:33:13,610 --> 00:33:17,930 S1: Cannot wait to get it around here. Waymo is just 535 00:33:17,930 --> 00:33:22,370 S1: just unbelievably good. New Zealand appears to be managing its 536 00:33:22,370 --> 00:33:26,630 S1: $16 billion health budget. Using a single Excel spreadsheet probably 537 00:33:26,630 --> 00:33:30,260 S1: makes Doge pretty happy. Apple is evidently planning a major 538 00:33:30,260 --> 00:33:34,520 S1: redesign of iOS, iPadOS, and macOS for the fall. Can't 539 00:33:34,520 --> 00:33:36,230 S1: wait to see what that looks like. I'll be running 540 00:33:36,230 --> 00:33:39,680 S1: it on day one because I'm stupid. Tesla's global sales 541 00:33:39,680 --> 00:33:43,340 S1: are tanking pretty hard right now with significant drops across Germany, 542 00:33:43,370 --> 00:33:47,570 S1: Australia and China. Yeah, you do a Hitler salute and 543 00:33:47,570 --> 00:33:53,300 S1: sales fall in Germany. Big surprise. New Mac studio has 32. 544 00:33:53,570 --> 00:33:57,680 S1: This is the M3 ultra 32 CPU cores and enough 545 00:33:57,710 --> 00:34:04,130 S1: unified memory, 512GB to load the 600 billion parameter models. 546 00:34:04,730 --> 00:34:06,950 S1: I was going to get one, but I already have 547 00:34:06,950 --> 00:34:09,920 S1: my eye box. I'm going to skip this generation, maybe 548 00:34:09,920 --> 00:34:13,280 S1: even another one, because I already have an M2 ultra. 549 00:34:13,310 --> 00:34:16,430 S1: Not worth it to bump up to the 512 especially. 550 00:34:16,430 --> 00:34:21,530 S1: That's like a $15,000 desktop discovery. A machine learning engineer 551 00:34:21,560 --> 00:34:24,410 S1: named Jasper Gillie quit his job because he believes AI 552 00:34:24,410 --> 00:34:27,799 S1: will automate his entire role by the end of 2025. 553 00:34:28,010 --> 00:34:32,210 S1: Keep in mind, this is an extremely highly paid machine 554 00:34:32,210 --> 00:34:36,950 S1: learning engineer, and he's saying his job is going to 555 00:34:36,950 --> 00:34:40,850 S1: be replaced by the end of 2025. Robin Moffat argues 556 00:34:40,850 --> 00:34:43,790 S1: we should or we shouldn't, just write the perfectly structured 557 00:34:43,790 --> 00:34:47,090 S1: blog post, but also share our messy struggles and solutions 558 00:34:47,090 --> 00:34:49,940 S1: that other people need, too. I like this. It's like 559 00:34:49,940 --> 00:34:53,060 S1: learning in public. Show your work, you know. Feel free 560 00:34:53,060 --> 00:34:58,040 S1: to be vulnerable and imperfect online, I think is, uh. 561 00:34:58,070 --> 00:35:00,979 S1: It's endearing. Hopefully so since I've said edit like nine 562 00:35:00,980 --> 00:35:05,000 S1: times in this recording already. Small docs a project around 563 00:35:05,000 --> 00:35:07,820 S1: writing small docs instead of giant ones that are hard 564 00:35:07,820 --> 00:35:12,620 S1: to follow. Jamie Rumbelow created a slick hybrid between databases, 565 00:35:12,620 --> 00:35:16,820 S1: programming and UIs. Make scripts more user friendly without needing 566 00:35:16,820 --> 00:35:19,250 S1: a full web app. And basically the pitch there was 567 00:35:19,250 --> 00:35:23,750 S1: that the database, the UI and the app are kind 568 00:35:23,750 --> 00:35:27,200 S1: of unified. They're all the same thing. Astro and vim, 569 00:35:27,200 --> 00:35:30,350 S1: one of the better vim distros. I do prefer lazy vim, 570 00:35:30,380 --> 00:35:33,750 S1: although I hate the name. Reminds me of the dummies books, 571 00:35:33,750 --> 00:35:38,070 S1: which I could never get myself to buy. Too much self-respect. 572 00:35:38,070 --> 00:35:42,360 S1: Interview DB lets you access crowdsourced tech interview questions from 573 00:35:42,360 --> 00:35:46,590 S1: the community. Alexander Solzhenitsyn on the future of the West. 574 00:35:46,590 --> 00:35:52,500 S1: This was a crazy, crazy piece of text there. Seneschal MCP. 575 00:35:52,800 --> 00:35:58,680 S1: MCP by a brilliant Ull member named Matt, who, um, 576 00:35:58,680 --> 00:36:03,390 S1: wrote an MCP that helps Llms access standardized health data 577 00:36:03,390 --> 00:36:09,210 S1: from the sky. Senegal. Senegal. I keep thinking, Senussi. What 578 00:36:09,210 --> 00:36:12,779 S1: am I thinking of? Senegal. API. So basically, this is 579 00:36:12,780 --> 00:36:16,200 S1: an NCP that makes your health data available so that 580 00:36:16,200 --> 00:36:19,020 S1: you can have local agents or any kind of agent 581 00:36:19,020 --> 00:36:23,489 S1: really make your health data available to an app. So 582 00:36:23,489 --> 00:36:27,540 S1: if you're building like a full life encompassing, you know, 583 00:36:27,570 --> 00:36:33,500 S1: management app with like workout routines, health, fitness, food, diet, money, 584 00:36:33,530 --> 00:36:37,280 S1: all that stuff. This is an API. This is an 585 00:36:37,280 --> 00:36:40,880 S1: MCP that presents the health side of it. Very very 586 00:36:40,880 --> 00:36:45,469 S1: cool from UL. Member Matt Self interview AI tool that 587 00:36:45,469 --> 00:36:48,830 S1: runs invisibly during technical interviews, helping you with coding questions 588 00:36:48,830 --> 00:36:52,700 S1: while your screen is shared. Also known as cheating. So 589 00:36:52,700 --> 00:36:53,990 S1: this is what you have to watch out for if 590 00:36:53,989 --> 00:36:59,780 S1: you're doing interviews. Piano FM generates endless beautiful AI piano music. 591 00:36:59,780 --> 00:37:03,530 S1: I listened for a while. It was not repetitive and 592 00:37:03,530 --> 00:37:07,760 S1: it sounded. I'm not a classical music expert or a 593 00:37:07,760 --> 00:37:12,140 S1: piano expert, and it sounded pretty normal to me. It 594 00:37:12,140 --> 00:37:16,430 S1: sounded like regular piano, whatever that's worth. Matrix hacking game. 595 00:37:16,430 --> 00:37:19,910 S1: Someone built a matrix themed AI hacking game lets users 596 00:37:19,910 --> 00:37:24,230 S1: learn about prompt engineering and security by defeating an Agent 597 00:37:24,230 --> 00:37:28,310 S1: Smith AI character. Send new app connect your iPhone camera 598 00:37:28,310 --> 00:37:33,000 S1: to multi-modal Llms for real time visual intelligence made with 599 00:37:33,000 --> 00:37:36,330 S1: zero VC money and cursor. This is the type of 600 00:37:36,360 --> 00:37:39,720 S1: thing that could have gone for, could have raised, you know, 601 00:37:39,750 --> 00:37:45,899 S1: $20 million and been sold for $100 million maybe a 602 00:37:45,900 --> 00:37:48,840 S1: couple of years ago. And this person just built it 603 00:37:48,840 --> 00:37:52,770 S1: with cursor with no money whatsoever. Since Feedback Dev built 604 00:37:52,770 --> 00:37:55,050 S1: an AI that reads the emotional vibe of your product, 605 00:37:55,080 --> 00:37:59,220 S1: social media mentions and gives you instant feedback. And somebody 606 00:37:59,219 --> 00:38:02,250 S1: posted their cursor rules, which I thought were quite good. 607 00:38:06,210 --> 00:38:08,670 S1: All right. This is the end of the standard edition 608 00:38:08,670 --> 00:38:11,460 S1: of the podcast, which includes just the news items for 609 00:38:11,460 --> 00:38:13,830 S1: the week to get the rest of the episode, which 610 00:38:13,860 --> 00:38:17,190 S1: includes my analysis, the discovery section that has all of 611 00:38:17,190 --> 00:38:20,790 S1: the coolest tools and articles. I found this week, the 612 00:38:20,790 --> 00:38:23,580 S1: recommendation of the week and the aphorism of the week. 613 00:38:23,580 --> 00:38:28,080 S1: Please consider becoming a member. As a member, you'll get 614 00:38:28,080 --> 00:38:31,709 S1: lots of different things from access to our extraordinary community 615 00:38:31,710 --> 00:38:35,490 S1: of over a thousand brilliant and kind people in industries 616 00:38:35,489 --> 00:38:40,800 S1: like cybersecurity, AI, technology and the humanities. You also get 617 00:38:40,830 --> 00:38:45,029 S1: access to the UL Book Club, dedicated member content and events, 618 00:38:45,030 --> 00:38:49,140 S1: and lots, lots more. Plus, you'll get a dedicated podcast 619 00:38:49,140 --> 00:38:51,450 S1: feed you can put in your client that gets you 620 00:38:51,450 --> 00:38:55,230 S1: the full member edition of this podcast. To become a 621 00:38:55,230 --> 00:38:57,900 S1: member and get all of that, just head over to 622 00:38:57,930 --> 00:39:04,470 S1: Daniel Comm Slash upgrade. That's Daniel miessler.com/upgrade. We'll see you 623 00:39:04,469 --> 00:39:09,960 S1: next time. Unsupervised learning is produced on Hindenburg Pro using 624 00:39:09,960 --> 00:39:13,320 S1: an SM seven B microphone. A video version of the 625 00:39:13,320 --> 00:39:17,370 S1: podcast is available on the Unsupervised Learning YouTube channel, and 626 00:39:17,370 --> 00:39:20,160 S1: the text version with full links and notes is available 627 00:39:20,160 --> 00:39:24,960 S1: at Daniel Mysa.com slash newsletter. We'll see you next time.