WEBVTT - UL NO. 482 | STANDARD EDITION: AI Finds an 0-Day!, Postman Leaking Secrets, High Agency Mental Model, My Unified Entity Context Video, Github MCP Leaks Private Repos, Google vs. OpenAI vs. Apple on AI Vision, and more... 0:00:18.667 --> 0:00:21.107 All right. Welcome to unsupervised learning. This is Daniel Miessler. 0:00:22.027 --> 0:00:25.666 And this is episode 482. And I want to mention 0:00:25.667 --> 0:00:29.587 real quick that there haven't been any regular podcast episodes 0:00:29.587 --> 0:00:33.266 recently on the podcast, as I'm sure you've noticed, and 0:00:33.307 --> 0:00:40.586 I'm basically rethinking how I'm approaching, uh, basically the podcast. Basically, 0:00:40.587 --> 0:00:42.626 the issue that I have is that a lot of 0:00:42.627 --> 0:00:45.507 the podcast episodes were just really long, and they're really 0:00:45.507 --> 0:00:48.187 difficult for me to actually read out, because I end 0:00:48.227 --> 0:00:51.467 up going into a whole bunch of like, extra exposition 0:00:51.467 --> 0:00:56.067 into them, and it's either too long for people or 0:00:56.667 --> 0:01:01.147 it's really exhausting for me. Well, it is too long 0:01:01.147 --> 0:01:04.547 for many people, and it's definitely exhausting for me. And 0:01:04.547 --> 0:01:07.267 the other problem is, a lot of the content that's 0:01:07.267 --> 0:01:09.787 in there is not interesting to a lot of people, right? 0:01:09.827 --> 0:01:11.667 It's a lot of detail about stuff that a lot 0:01:11.667 --> 0:01:14.307 of people don't care about. So what I've been doing 0:01:14.307 --> 0:01:17.757 is switching up to focusing on ideas that are shorter 0:01:18.117 --> 0:01:20.117 and more compact. And that's what you've been seeing in 0:01:20.117 --> 0:01:22.277 the podcast feed, and it looks like you've been enjoying 0:01:22.277 --> 0:01:25.157 those based on the number of downloads. So I'm going 0:01:25.197 --> 0:01:27.997 to continue to do that and especially focus on that 0:01:27.997 --> 0:01:31.557 because I think that's that's basically, you know, the the 0:01:31.597 --> 0:01:35.877 future of the podcast is to focus mostly on, you know, 0:01:35.917 --> 0:01:38.877 ideas and, you know, exciting things that are happening and 0:01:38.877 --> 0:01:41.397 trends and basically talking through them in depth, which is 0:01:41.397 --> 0:01:43.517 the thing that the people like from the podcast anyway. 0:01:43.517 --> 0:01:45.676 So think of it this way, the podcast will still 0:01:45.677 --> 0:01:47.637 be the same, the podcast will still be coming out, 0:01:47.637 --> 0:01:50.517 but what you'll be getting is where I would go 0:01:50.517 --> 0:01:54.277 into a topic before inside of a podcast episode. Those 0:01:54.277 --> 0:01:56.437 are now getting pulled out and turned into their own 0:01:56.437 --> 0:01:59.517 little mini segments, and those are going into the the 0:01:59.557 --> 0:02:02.317 top of the podcast at the feed level, right? And 0:02:02.317 --> 0:02:04.997 they'll have their own names and everything. The question is, 0:02:04.997 --> 0:02:08.117 do I still do this news one which I'm doing now, right? 0:02:08.157 --> 0:02:09.556 And a lot of people still want to get the 0:02:09.556 --> 0:02:10.957 news one. So I think what I'm going to do 0:02:10.957 --> 0:02:15.117 is continue to release the specific episode based ones. You know, 0:02:15.157 --> 0:02:17.637 the idea ones I just talked about, but do the 0:02:17.637 --> 0:02:21.677 audio version of this one here, where it's labeled episodes 0:02:21.677 --> 0:02:24.756 and it's actually news and it's, you know, it gives 0:02:24.757 --> 0:02:26.637 you coverage of what happened in the week before, like 0:02:26.677 --> 0:02:28.637 the podcast has always been. So I'm going to try 0:02:28.677 --> 0:02:32.076 not to go into too much depth on various ideas 0:02:32.077 --> 0:02:33.717 and topics. I probably won't be able to hold to 0:02:33.716 --> 0:02:38.037 that because I get excited, but ideally, any depth content 0:02:38.357 --> 0:02:40.837 inside of that episode should be broken out into its 0:02:40.837 --> 0:02:42.957 own episode, which will be a mini segment inside of 0:02:42.957 --> 0:02:45.157 the feed. So this is why you haven't seen the 0:02:45.156 --> 0:02:50.397 same number of specific UL labeled episodes. Is just because 0:02:50.397 --> 0:02:53.637 I'm a little bit less interested in the news stuff, 0:02:53.637 --> 0:02:56.637 and more interested in the ideas and the analysis, which is, 0:02:57.036 --> 0:02:58.756 you know, where the main push is going to be. 0:02:58.756 --> 0:03:01.117 But again, enough people are saying they want the news. 0:03:01.117 --> 0:03:03.957 So I'm going to continue to put that out in 0:03:03.957 --> 0:03:06.956 audio only right here at the same place on the podcast. 0:03:09.436 --> 0:03:12.476 All right. So with that out of the way, let's 0:03:12.477 --> 0:03:17.327 go into the episode. All right. So if you still 0:03:17.327 --> 0:03:19.927 doubt that eyes are able to reason, you need to 0:03:19.927 --> 0:03:23.806 watch this segment of two anthropic researchers talking with Dwarkesh Patel. 0:03:24.406 --> 0:03:26.967 So these are two guys that he's brought on before, 0:03:26.966 --> 0:03:28.607 and they talked about the future of AI. There was 0:03:28.647 --> 0:03:30.966 like almost a year ago, seems like five years ago 0:03:30.966 --> 0:03:34.246 or whatever, but it was really good. But this particular 0:03:34.246 --> 0:03:38.167 segment that I highlighted in the newsletter, and by the way, 0:03:38.167 --> 0:03:41.167 you won't have the newsletter handy when I'm talking when 0:03:41.167 --> 0:03:43.527 you're listening to this audio, ideally, if you could scroll 0:03:43.527 --> 0:03:47.487 through the newsletter or watch the video, pause it, listen, whatever. Um, 0:03:47.927 --> 0:03:50.447 they are definitely sisters of each other, right? The podcast 0:03:50.447 --> 0:03:52.687 versus the newsletter. It's all the same stuff. It's just 0:03:52.687 --> 0:03:55.127 the text version, but more importantly, it has the links. 0:03:55.287 --> 0:03:58.407 So I have a link in there specifically going to 0:03:58.447 --> 0:04:01.167 that part in the video where they talk about this. 0:04:01.167 --> 0:04:03.967 And the thing they're talking about is this segment research 0:04:03.967 --> 0:04:08.367 that they did. So what they did was the interpretability 0:04:08.407 --> 0:04:11.607 team from anthropic released a paper saying that they could 0:04:11.607 --> 0:04:15.487 light up. They could when they ask, uh, Claude a question, 0:04:16.167 --> 0:04:19.217 what would happen is different sections. They could see them 0:04:19.217 --> 0:04:22.857 light up. And this is parameters. The circuit that they're 0:04:22.857 --> 0:04:28.657 talking about is parameters across multiple layers of the model. Okay. 0:04:28.697 --> 0:04:30.777 So you've got like you know I don't know how 0:04:30.777 --> 0:04:33.777 many layers current models have, but it's many, many, many, 0:04:33.777 --> 0:04:38.097 many layers deep right. From left to right. Now inside 0:04:38.097 --> 0:04:41.337 of each layer there's so many parameters, right. You know, 0:04:41.377 --> 0:04:44.617 millions of parameters or whatever it is. But what they 0:04:44.617 --> 0:04:48.257 could see is what each of those lit up areas 0:04:48.337 --> 0:04:51.497 actually correspond to in terms of concepts. So they can 0:04:51.497 --> 0:04:54.457 now see that when they start to prompt somebody, uh, 0:04:54.617 --> 0:04:58.417 a model for like write a sonnet or something like that, 0:04:58.697 --> 0:05:01.897 you could see it planning, you could see it planning 0:05:01.897 --> 0:05:04.377 early in the process, you could see it thinking about 0:05:04.377 --> 0:05:07.697 where it wants to go and moving through that. The 0:05:07.697 --> 0:05:10.257 other example that he gave that was absolutely insane is 0:05:10.257 --> 0:05:14.017 there's this really rare symptom of a problem that has 0:05:14.017 --> 0:05:16.987 to do with pregnancy. And what they did was they 0:05:16.987 --> 0:05:21.427 gave the problem to Claude, and Claude immediately took the 0:05:21.427 --> 0:05:26.547 symptoms and immediately started mapping like cause and effect across 0:05:26.547 --> 0:05:31.187 multiple possible diseases. And it immediately lit up the answer, actually. 0:05:31.347 --> 0:05:33.827 But it started thinking about what would the symptoms be, 0:05:33.867 --> 0:05:37.227 which would could exclude it. Right. So it started thinking 0:05:37.227 --> 0:05:39.347 very much like a human. But but the thing that 0:05:39.347 --> 0:05:41.587 was different about it is they released a paper showing 0:05:41.587 --> 0:05:45.587 how these sections are actually lighting up. They're getting visibility 0:05:45.587 --> 0:05:48.987 into how the eye is actually thinking and how it's 0:05:48.987 --> 0:05:52.907 thinking about cause and effect and logic and conclusion. And 0:05:52.907 --> 0:05:56.227 basically like the thing that a lot of people say, oh, 0:05:56.267 --> 0:05:59.427 it's just next token prediction. This pretty much puts that 0:05:59.427 --> 0:06:03.187 to sleep. It really does because obviously it's still only 0:06:03.187 --> 0:06:07.307 doing next token prediction. You know, air quotes only. But 0:06:07.307 --> 0:06:09.587 if it's lighting up these sections and we could see 0:06:09.787 --> 0:06:12.827 it's lighting up the same sort of sections, there's probably 0:06:12.827 --> 0:06:15.467 an analog like this in our own brains. We are 0:06:15.587 --> 0:06:19.947 mapping concepts, we are mapping cause and effect in the world. 0:06:19.947 --> 0:06:21.147 And by the way, it came up with the right 0:06:21.147 --> 0:06:24.427 solution for this medical thing. And I said this to 0:06:24.467 --> 0:06:27.747 a friend of mine named Jonathan who's a cardiologist. He's like, 0:06:27.747 --> 0:06:31.467 this is exactly how I think about problems, you know, 0:06:31.507 --> 0:06:34.707 related to my industry, which is he's an active cardiologist. 0:06:34.707 --> 0:06:37.747 He's in, you know, his his office, in his practice 0:06:37.747 --> 0:06:43.267 all the time. And I just want to emphasize how 0:06:44.227 --> 0:06:47.507 incredible this is, okay. And I want you to be 0:06:47.507 --> 0:06:50.107 able to send this link and send this clip to 0:06:50.147 --> 0:06:54.387 people who keep saying, I don't really understand. And I've 0:06:54.387 --> 0:06:57.307 also got another link related to this of Ilya talking 0:06:57.307 --> 0:07:01.507 about explaining how next token prediction and true understanding can 0:07:01.507 --> 0:07:03.587 be the same and in fact are the same. And 0:07:03.587 --> 0:07:05.987 basically what he says is it turns out if you 0:07:05.987 --> 0:07:09.387 want to do proper next token prediction, you actually have 0:07:09.387 --> 0:07:12.187 to understand the world that you're talking about, okay. And 0:07:12.187 --> 0:07:16.317 that very closely maps and tracks with this segment research, 0:07:16.837 --> 0:07:19.517 where you're lighting up different parts of the model, which 0:07:19.517 --> 0:07:24.357 are interconnected, and related concepts which show a clear understanding 0:07:24.357 --> 0:07:28.597 of cause and effect. Right. So this is just really 0:07:28.597 --> 0:07:32.637 important to understand that just the fact that they're doing 0:07:32.637 --> 0:07:36.317 next token prediction doesn't mean they're not understanding next token 0:07:36.317 --> 0:07:39.877 prediction is just the output. It's just the mechanism. It's 0:07:39.877 --> 0:07:43.837 like saying to going to a poet, watching a poet, 0:07:43.877 --> 0:07:48.077 watching Shakespeare speak words on a on a stage, and 0:07:48.077 --> 0:07:50.557 then looking over at somebody and saying, he's just moving 0:07:50.557 --> 0:07:53.997 his mouth. And what happens is when his mouth moves, 0:07:54.237 --> 0:07:57.077 it vibrates sound. And the sound comes to our ears 0:07:57.077 --> 0:07:59.157 and it vibrates something in our ears. And then we 0:07:59.197 --> 0:08:02.717 hear his words like, he's not doing anything special. He's 0:08:02.717 --> 0:08:07.477 just moving his mouth and vibrating air. Yeah, you're right. 0:08:07.477 --> 0:08:10.117 You 100% got me. He's just moving his mouth and 0:08:10.117 --> 0:08:14.277 just vibrating air. That has nothing to do with what 0:08:14.317 --> 0:08:18.047 underlies it? Right. And it is the same with AI 0:08:18.087 --> 0:08:24.167 understanding and next token prediction. All right. See, you see 0:08:24.167 --> 0:08:26.767 how I get diverted here. All of that. We're on 0:08:26.807 --> 0:08:29.687 the number one bullet before we even start the newsletter. Okay. 0:08:29.727 --> 0:08:31.887 Next one. If you're part of the 23 Andme breach, 0:08:31.887 --> 0:08:34.247 which I was, you can submit a claim to receive 0:08:34.287 --> 0:08:37.687 between $500 and $1500. But you have to do it 0:08:37.687 --> 0:08:41.367 before July 14th. And I have the link here. All right. 0:08:41.367 --> 0:08:43.967 I've got a new video out. It is called Unified 0:08:43.967 --> 0:08:47.967 Entity Context. The AI stack. Everyone is building without realizing it. 0:08:47.967 --> 0:08:51.087 You've got to go see this video and make sure 0:08:51.087 --> 0:08:56.367 you like and subscribe on the channel as well. And 0:08:56.367 --> 0:08:59.046 my friend Ali is looking for a new hybrid on 0:08:59.087 --> 0:09:01.686 site position, the New York area doing back end development. 0:09:01.687 --> 0:09:07.286 She's an absolute force MIT grad, a master's in engineering 0:09:07.886 --> 0:09:13.926 and bachelor's in electrical engineering and CS three years back. 0:09:13.926 --> 0:09:18.296 End engineer experience founder Devrel experience. Lots of Golang, really 0:09:18.296 --> 0:09:23.176 strong public speaking and presentation skills and she's on social media. 0:09:23.217 --> 0:09:26.377 She explains stuff. She does education for the community. She's 0:09:26.377 --> 0:09:30.816 just fantastic. And she's a like hardcore serious MIT developer. Like, 0:09:30.857 --> 0:09:32.657 don't get it twisted. A lot of people see her. 0:09:32.656 --> 0:09:35.177 They see that she's like, has good energy and she's 0:09:35.176 --> 0:09:37.497 on social media and they assume like she's some kind of, 0:09:37.536 --> 0:09:40.097 you know, influencer who's talking about coding. No, she's not 0:09:40.097 --> 0:09:45.057 talking about coding. She's doing it. Masters from MIT. Okay. 0:09:45.097 --> 0:09:47.977 So you want to hire her very quickly if she's 0:09:47.977 --> 0:09:55.176 not already snatched up. Cybersecurity. Palo Alto their research team 0:09:55.176 --> 0:09:58.697 simulated a full ransomware attack in 25 minutes using AI agents. 0:09:59.857 --> 0:10:02.937 I continue to like, help Net Security's job posting of 0:10:02.937 --> 0:10:05.537 different cybersecurity roles. I think they put this out every 0:10:05.857 --> 0:10:08.416 week or every couple of weeks. But, um, it's pretty good. 0:10:08.416 --> 0:10:10.896 And I got the list here to latest job listings. 0:10:11.416 --> 0:10:15.617 Postman looks like it's logging secrets and environment variables. So 0:10:15.977 --> 0:10:17.497 you want to be you want to go in and 0:10:17.497 --> 0:10:19.857 change this. This is actually on by default. A lot 0:10:19.896 --> 0:10:22.097 of people are really pissed about this. So, um, you 0:10:22.097 --> 0:10:24.097 want to make sure to turn that off or, you know, 0:10:24.097 --> 0:10:27.097 find another solution if it makes you that mad. Um, 0:10:27.097 --> 0:10:30.737 O3 actually found a remote Linux kernel zero day, so 0:10:30.776 --> 0:10:33.497 it might be the first instance, although I wouldn't be 0:10:33.497 --> 0:10:38.337 surprised if it weren't of an AI model actually finding 0:10:38.656 --> 0:10:43.576 a zero day vulnerability. Cloud and GitHub MCP will leak 0:10:43.577 --> 0:10:47.656 your private GitHub repositories. You got to watch out with 0:10:47.656 --> 0:10:51.537 these MCC. You really do. There's a lot of, um, 0:10:51.577 --> 0:10:54.656 hand-waving going on with them. They're very powerful. They're very awesome. 0:10:54.656 --> 0:10:56.857 But you got to watch out. So be careful with 0:10:56.857 --> 0:11:00.817 anything you're doing that's sensitive stuff. Signal blocks, windows, screenshots 0:11:00.817 --> 0:11:03.176 to counter recall features. So recall wants to take screenshots 0:11:03.176 --> 0:11:07.416 of everything. Signal is specifically blocking that because it goes 0:11:07.457 --> 0:11:11.896 against their ethos obviously police arrest third suspect in bitcoin 0:11:11.937 --> 0:11:14.906 torture case is this guy escaped after three weeks of 0:11:14.906 --> 0:11:16.906 torture by colleagues trying to force him to reveal his 0:11:16.906 --> 0:11:20.666 Bitcoin password. And evidently he did not. My understanding from 0:11:20.666 --> 0:11:23.427 the military is if someone's good at torturing, you are 0:11:23.426 --> 0:11:26.067 going to break. The only question is when. So the 0:11:26.067 --> 0:11:29.227 fact that it went three weeks, um, probably just because 0:11:29.227 --> 0:11:31.667 it's his colleagues and not really good at the whole, uh, 0:11:33.067 --> 0:11:37.466 I guess, uh, applying pressure in that way. US intelligence 0:11:37.506 --> 0:11:40.226 is creating a unified portal for buying your personal data. 0:11:40.227 --> 0:11:42.546 So they're building a centralized system to streamline how they 0:11:42.546 --> 0:11:46.826 purchase sensitive information from brokers. So they normally would have 0:11:46.827 --> 0:11:48.467 required a court order to get this stuff. But if 0:11:48.467 --> 0:11:51.506 they buy it legally from brokers, they can do that, 0:11:51.506 --> 0:11:54.707 which means they're just now building this giant AI portal 0:11:55.467 --> 0:11:57.747 AI data lake so they can ask questions and get 0:11:57.747 --> 0:12:01.307 the data back in aggregate. I mean, we definitely knew 0:12:01.307 --> 0:12:03.066 this was going to happen, but it's still pretty scary. 0:12:03.347 --> 0:12:07.787 Coinbase hit with insider data breach, refuses $20 million ransom. 0:12:08.107 --> 0:12:12.747 So here's the craziest their overseas support agents, which means 0:12:12.747 --> 0:12:15.077 you could pay them much less money to bribe them. 0:12:15.597 --> 0:12:17.317 So there's a really cool meme on this. I don't 0:12:17.317 --> 0:12:19.957 remember how they phrased it, but my phrasing is it's 0:12:19.957 --> 0:12:23.117 cheaper to hire overseas people, which means it's also cheaper 0:12:23.117 --> 0:12:26.877 to bribe overseas people. And definitely something to think about. 0:12:26.877 --> 0:12:31.676 Czech Republic accuses China of hacking foreign ministry AI hallucination 0:12:31.676 --> 0:12:34.837 cases database launched. So someone created a database of real 0:12:34.837 --> 0:12:40.677 world hallucinations caused by AI that you EU enters vulnerability 0:12:40.717 --> 0:12:43.756 tracking race as Nvda is getting really old. So they're 0:12:44.276 --> 0:12:47.997 bringing their own competitor in there. FBI warns of AI 0:12:48.036 --> 0:12:52.037 voice impersonation campaign targeting officials. So say to watch out 0:12:52.036 --> 0:12:53.957 for that. And they're actually coming after the people who 0:12:53.997 --> 0:12:56.597 are doing it if they can find them. Amazon has 0:12:56.597 --> 0:12:59.676 a serious problem with fake supplements. Really good thread here. 0:12:59.676 --> 0:13:02.717 Indian police are attempting to read suspects minds, so they're 0:13:02.717 --> 0:13:06.636 using brain scans to detect guilt. And pretty much all 0:13:06.636 --> 0:13:09.317 the experts are like, do not do this. It's not 0:13:09.357 --> 0:13:11.357 what you think it is, but they continue to do it. 0:13:11.796 --> 0:13:15.607 Snowflake CISO shifts from shared responsibility to shared destiny, which 0:13:15.607 --> 0:13:17.687 is really just more alignment with the business, which I 0:13:17.687 --> 0:13:19.967 think is the only way to survive as a security 0:13:19.967 --> 0:13:26.607 leader right now and going forward. Russian Apt28 breaches organizations 0:13:26.607 --> 0:13:28.687 to track aid routes. So they are hacking a whole 0:13:28.687 --> 0:13:31.806 bunch of different groups to figure out who's sending support 0:13:32.247 --> 0:13:35.166 to Ukraine, obviously, so they can try to shut that off. 0:13:35.567 --> 0:13:40.886 Regeneron to acquire 23 Andme and its customer data. National 0:13:40.886 --> 0:13:44.927 security Russia mandates location tracking app for foreigners in Moscow. 0:13:45.087 --> 0:13:46.807 Every foreigner they're going to have to they're going to 0:13:46.807 --> 0:13:49.807 track their location or they're going to try to. I'm 0:13:49.847 --> 0:13:53.127 not sure how far along it is. All right. I 0:13:53.487 --> 0:13:57.487 OpenAI's Jony Ive hardware Google I o in my current assessment. 0:13:57.487 --> 0:14:00.607 So they are building this little round necklace. It looks 0:14:00.607 --> 0:14:02.967 very it looks very pretty. It looks to have a 0:14:02.967 --> 0:14:06.007 camera and a microphone on it. And Apple actually bought 0:14:06.006 --> 0:14:09.646 his little company for over $6 billion. So he now 0:14:09.687 --> 0:14:13.687 works at OpenAI for OpenAI for Sam. And I think 0:14:13.687 --> 0:14:16.847 this is moving us towards the vision that I was 0:14:16.847 --> 0:14:19.007 talking about before or have been talking about for a 0:14:19.007 --> 0:14:25.127 long time, where basically, um, your digital assistant is using 0:14:25.127 --> 0:14:29.167 this thing that you're wearing to monitor the visual in 0:14:29.167 --> 0:14:30.847 front of you. Now, Apple is talking about doing this 0:14:30.847 --> 0:14:33.527 with AirPods and actually having cameras on the AirPods. I 0:14:33.527 --> 0:14:34.767 would love to be able to see in front of 0:14:34.767 --> 0:14:37.566 me and behind me. That would make me happy because 0:14:37.567 --> 0:14:44.607 my digital assistant, whose name is probably going to be Kai. Um, well, 0:14:44.647 --> 0:14:46.767 already is kind of Kai. But when the real one 0:14:46.767 --> 0:14:50.647 comes out, transfer his soul into the new location. But anyway, 0:14:50.687 --> 0:14:51.807 I want him to be able to see you in 0:14:51.807 --> 0:14:55.087 front of me. Obviously, here, obviously monitor all the APIs 0:14:55.087 --> 0:14:57.807 that we've been talking about for everything around me. And 0:14:57.847 --> 0:14:59.447 obviously you have an earpiece, so we could talk in 0:14:59.447 --> 0:15:02.167 my ear very quietly, like, this person is lying to you. 0:15:02.167 --> 0:15:04.367 Someone's sneaking up behind you. Someone is staring at your 0:15:04.367 --> 0:15:11.087 screen from your right rear oblique, you know, uh, 8:00 0:15:11.447 --> 0:15:15.297 to your 8:00. Someone's monitoring your screen, whatever. The camera 0:15:15.297 --> 0:15:19.017 is the biggest piece here. And, um. Yeah, I'm getting 0:15:19.017 --> 0:15:21.897 worried that Apple is falling behind here. Google's I o 0:15:22.017 --> 0:15:26.217 conference was unbelievably great. They announced so many different things. 0:15:26.217 --> 0:15:30.217 They are going they're basically the monster has awoken, the 0:15:30.377 --> 0:15:34.057 sleeping giant has woken up and they're moving quick. And 0:15:34.057 --> 0:15:37.897 Apple needs to do something cool hopefully in this week. Oh, 0:15:37.897 --> 0:15:40.897 and by the way, they're renaming their operating systems. So 0:15:40.897 --> 0:15:44.177 it's going to be iOS 26 coming out in the 0:15:44.177 --> 0:15:47.897 next couple of weeks instead of iOS 19. So they're 0:15:47.897 --> 0:15:50.617 naming it for the year coming. That's upcoming just like 0:15:50.657 --> 0:15:54.897 car models. So iOS 26 coming out. Uh anthropic's. Claude, 0:15:54.897 --> 0:15:58.257 opus four shatters limits with seven hour coding marathon. So 0:15:58.537 --> 0:16:03.177 the big thing here is that, um, the agent can 0:16:03.177 --> 0:16:06.257 maintain the context of the goal for seven hours and 0:16:06.257 --> 0:16:09.297 keep working towards that goal. That's the real big update here. 0:16:09.497 --> 0:16:11.696 I've been talking for a long time about this. Uh. 0:16:11.947 --> 0:16:16.627 Scaffolding matters more than the models. Context matters more than 0:16:16.627 --> 0:16:20.787 the models. Systems matter more than the models. Having a 0:16:20.787 --> 0:16:23.627 way to track what you're trying to do and understand 0:16:23.627 --> 0:16:26.306 what you're trying to do, and understand the various steps involved, 0:16:26.307 --> 0:16:29.147 and understand where you are in terms of doing all 0:16:29.147 --> 0:16:31.947 those steps. That is the thing that's going to super 0:16:31.987 --> 0:16:34.747 power all this AI stuff. Yeah, the models get better, 0:16:34.747 --> 0:16:37.387 they get smarter, whatever. But the more important thing is 0:16:37.387 --> 0:16:41.186 how big is our context window and how powerful is 0:16:41.187 --> 0:16:43.987 the system that they are using to try to accomplish 0:16:43.987 --> 0:16:46.227 their goals? And how good can they track the state 0:16:46.227 --> 0:16:49.427 of their progress towards those? And those are the things 0:16:49.427 --> 0:16:51.587 that they keep making more and more progress in across 0:16:51.587 --> 0:16:54.227 all these different companies. And that's why we're getting these 0:16:54.227 --> 0:16:58.826 giant leaps, not just because, like the intelligence of the models, 0:16:58.827 --> 0:17:03.907 but the scaffolding of everything around it. Authors accidentally leave 0:17:03.947 --> 0:17:06.627 AI prompts in published novels. So I think these are 0:17:06.667 --> 0:17:09.627 like romance novels. And I got some examples here. But yeah, 0:17:09.627 --> 0:17:12.877 it's got like it's got like the I talking back, okay, 0:17:12.877 --> 0:17:14.957 I did this in the voice of so and So author. 0:17:14.997 --> 0:17:17.157 Hope you like it. And it's like in the actual 0:17:17.157 --> 0:17:21.516 published book on Goodreads, researchers develop RL method that predicts 0:17:21.517 --> 0:17:28.436 future outcomes. So yeah, exciting. Scary. Um, which is everything 0:17:28.436 --> 0:17:32.557 in AI tech CEOs using AI avatars to replace themselves 0:17:32.557 --> 0:17:36.077 in earning calls. So Klarna and Zoom CEOs are actually 0:17:36.077 --> 0:17:39.237 using AI versions of themselves. That's I'm not sure that's 0:17:39.236 --> 0:17:43.357 a good precedent to set, but pretty cool technology. Apple 0:17:43.357 --> 0:17:46.956 races to build smart glasses after getting behind on AI wave. 0:17:47.597 --> 0:17:50.236 So yeah, Google is building these things. Meta is building 0:17:50.236 --> 0:17:53.637 these things. Apple was starting with Vision Pro, but that's 0:17:53.637 --> 0:17:55.557 not portable. So they got to go with the glasses 0:17:55.557 --> 0:17:58.517 model because that's where Google and Meta are going. And 0:17:58.557 --> 0:18:00.877 look they got to move fast. They got to move fast. 0:18:01.397 --> 0:18:05.277 Tesla makes big push to catch Waymo in Austin robotaxi race. 0:18:05.637 --> 0:18:08.277 I can't wait to see this play out. Um I 0:18:08.317 --> 0:18:12.037 want Elon to win here. I'm mad at Ellen, so 0:18:12.037 --> 0:18:14.957 I don't kind of personally want him to win all 0:18:14.956 --> 0:18:17.077 that much right now. But I want Tesla to win 0:18:17.077 --> 0:18:21.597 because I don't want BYD to win. Um, Google is American, 0:18:21.597 --> 0:18:24.837 so I'm happy that there's a competition between Waymo and Tesla, 0:18:25.157 --> 0:18:27.476 and that makes me happy. And I can't wait to 0:18:27.476 --> 0:18:31.316 see a lot fewer drunk drivers, hopefully. Uh, because when 0:18:31.317 --> 0:18:35.877 they get into this vehicle, um, maybe it won't drive 0:18:36.277 --> 0:18:39.236 if you're drunk. Maybe, um, they just get into a 0:18:39.236 --> 0:18:42.197 back of a free vehicle because some, uh, local community 0:18:42.196 --> 0:18:44.877 service pays for any drunk person to call, and it 0:18:44.877 --> 0:18:47.237 will take them home. Stuff like that I'm really looking 0:18:47.236 --> 0:18:51.277 forward to. Plus texting and driving. Uh, someone really wants 0:18:51.277 --> 0:18:53.436 to look at that TikTok, and they happen to be driving, 0:18:53.716 --> 0:18:55.677 and that is horribly dangerous and I can't wait for 0:18:55.677 --> 0:18:59.796 those accidents to go down. YouTube introduces peak points, AI 0:18:59.837 --> 0:19:04.277 ads after emotional moments. Google commits $150 million to develop 0:19:04.277 --> 0:19:06.917 AI glasses with Warby Parker. We just talked about that one. 0:19:07.997 --> 0:19:12.647 Walmart is cutting 1500 technology jobs trying to reduce costs. 0:19:14.127 --> 0:19:19.527 Greater Manchester NHS trust rejects Palantir's national data platform. They're 0:19:19.527 --> 0:19:22.806 going to build their own in-house system. And BYD is 0:19:22.847 --> 0:19:28.807 beating Tesla in Europe for the first time, registering 7231 0:19:28.807 --> 0:19:34.007 vehicles and Tesla fell to 11th place. Humans. Tech companies 0:19:34.007 --> 0:19:39.567 cut entry level hiring by half since 2019, GPT fake 0:19:39.567 --> 0:19:46.767 consciousness performance was disturbingly convincing, so Jesse Singal asked or 0:19:46.807 --> 0:19:50.486 single asked ChatGPT to pretend to be conscious and had 0:19:50.486 --> 0:19:55.367 a very unsettling conversation with it. Uh, yeah, it was convincing, 0:19:55.367 --> 0:20:01.407 in other words. Salesforce execs says, uh, that agents should 0:20:01.407 --> 0:20:03.726 make us rethink every job. I think he's correct, but 0:20:03.726 --> 0:20:06.327 it doesn't go far enough. The right path is to 0:20:06.367 --> 0:20:09.287 constantly ask, you know, if you're a business owner, what 0:20:09.287 --> 0:20:11.257 would I do today if I started start it over, 0:20:11.297 --> 0:20:13.736 it doesn't mean you could always do that. But you 0:20:13.736 --> 0:20:15.936 should always be thinking about that and not suffering from 0:20:15.936 --> 0:20:20.696 the sunk cost fallacy. And with AI and hiring specifically, 0:20:20.936 --> 0:20:22.736 this is going to this is going to be a 0:20:22.736 --> 0:20:24.496 big deal. And I think a lot of companies are 0:20:24.497 --> 0:20:25.976 doing that. And I think this is the bigger problem 0:20:25.976 --> 0:20:28.857 with the job market is companies are basically saying, look, 0:20:29.216 --> 0:20:31.337 what if we just started over? Why do we need 0:20:31.337 --> 0:20:33.777 all these thousands of people? What are they really doing 0:20:33.777 --> 0:20:37.257 for us? Let's get back to basics. Let's only look 0:20:37.257 --> 0:20:39.257 at the core programs that we have to have, the 0:20:39.257 --> 0:20:41.857 core products we have to have. Let's cut the products 0:20:41.857 --> 0:20:45.017 we don't need. Let's cut, you know, programs we don't need. 0:20:45.057 --> 0:20:48.296 Let's cut services we don't need. Basically, I think everyone's 0:20:48.297 --> 0:20:52.297 going to be like, let's get in super thin shape, 0:20:52.817 --> 0:20:57.417 reduce everything that's not necessary. And this doesn't just mean hiring. 0:20:57.417 --> 0:21:00.456 This means focus of the business, which happens to also 0:21:00.456 --> 0:21:02.936 mean hiring. And then once they have okay, these are 0:21:02.936 --> 0:21:04.737 the core things we're going to do. How would we 0:21:04.777 --> 0:21:07.857 hire for that now. And I'm telling you the answer 0:21:07.897 --> 0:21:12.467 is going to be find really really smart. Extremely ambitious. 0:21:13.107 --> 0:21:16.786 No work balance type. People who are experts with AI. 0:21:17.307 --> 0:21:19.387 That is who people are going to be hiring now. 0:21:19.747 --> 0:21:22.627 And of course, where you have like extreme expertise, like 0:21:22.667 --> 0:21:25.707 an AI data scientist or something, then obviously they're going 0:21:25.706 --> 0:21:28.107 to go with, you know, an extreme expert there. But 0:21:28.147 --> 0:21:32.026 in general, as employees, they're going to hire crazy people 0:21:32.027 --> 0:21:34.227 who are obsessed with the company, who will work crazy 0:21:34.226 --> 0:21:36.986 hours and not ask about work life balance, who are 0:21:36.986 --> 0:21:39.907 experts with AI and who will actually have multiple AIS 0:21:39.946 --> 0:21:42.707 working for them. And that is basically what the new 0:21:42.706 --> 0:21:46.306 employees look like. That is what new employment looks like. 0:21:46.347 --> 0:21:48.907 And if you are not one of those people, I 0:21:48.946 --> 0:21:52.787 think you're in extreme danger right now. Poland is about 0:21:52.787 --> 0:21:57.387 to overtake Japan's GDP. Llms outperform paid human persuaders in 0:21:57.387 --> 0:22:02.347 both truth and deception. Claude three five was significantly better 0:22:02.347 --> 0:22:06.107 at persuading people towards both correct and incorrect answers than 0:22:06.147 --> 0:22:10.867 humans who had financial incentives. Sleep apnea pill shows striking 0:22:10.867 --> 0:22:16.907 success in large clinical trial. Yeah. Cut breathing interruptions by 50%. 0:22:17.747 --> 0:22:22.827 Denmark raises retirement age to 70 by 2040. A Guardian 0:22:22.827 --> 0:22:27.267 writer downloaded years of Alexa data and basically found their 0:22:27.387 --> 0:22:32.587 his family's soul. So 15,000 recorded conversations included his daughter 0:22:32.587 --> 0:22:35.587 using Alexa as a therapist, asking about dating ages and 0:22:35.587 --> 0:22:40.946 sleep problems, which she didn't discuss with the parents. Earth's 0:22:40.986 --> 0:22:44.827 dual high tide phenomenon turns out there's actually two high tides, 0:22:44.827 --> 0:22:48.907 one on opposite sides of the Earth. Lost dog found 0:22:48.946 --> 0:22:51.507 thanks to dead AirTag, which someone put a new battery 0:22:51.507 --> 0:22:54.147 in AirTag. And it lit up and the owner came 0:22:54.147 --> 0:22:59.746 and got it. And, uh. Tim Hartfords four decade journey 0:22:59.747 --> 0:23:03.107 with Dungeons and Dragons. So a missing teenager story in 0:23:03.107 --> 0:23:06.946 1979 caused a moral panic, and it made D&D famous 0:23:07.347 --> 0:23:13.837 man 79. Wow. I didn't realize it was that old. Discovery. 0:23:14.476 --> 0:23:17.357 Reject jealousy and root for your friends. My buddy Joseph 0:23:17.357 --> 0:23:20.716 Thacker made a compelling case for rejecting jealousy and celebrating 0:23:20.716 --> 0:23:22.597 your friends wins. You got to go read this essay. 0:23:23.476 --> 0:23:26.677 What Sam Altman wishes someone had told him borrows from 0:23:26.677 --> 0:23:31.837 Anthropic's Claude team. I'm going to say that again. Borrows 0:23:31.837 --> 0:23:35.397 from Anthropic's Claude code team says context stuffing is way 0:23:35.397 --> 0:23:38.477 better than rag. It's funny because I've been using context 0:23:38.476 --> 0:23:42.797 stuffing as my main way of doing context. I don't 0:23:42.797 --> 0:23:46.637 like rags. I've always thought that they were janky, and 0:23:46.677 --> 0:23:50.677 it's great to hear Boris say that as well. Yeah, 0:23:50.716 --> 0:23:52.917 Ilya explaining that AI is actually do understand this. What 0:23:52.917 --> 0:23:55.837 I was talking about in the beginning inside the AI 0:23:55.877 --> 0:23:59.917 security arsenal that I've built, someone shows off 21 plus 0:24:00.236 --> 0:24:03.677 AI tools that they've created, from import table analysis to 0:24:03.716 --> 0:24:07.477 threat hunting frameworks. Paul Graham on what makes writing good. 0:24:08.077 --> 0:24:16.286 Internet Archive launches Live microfiche. Digitisation. Digitisation. Digitisation. Stream goodness. Digitisation. 0:24:16.767 --> 0:24:20.007 So basically this is a live YouTube stream of people 0:24:20.007 --> 0:24:24.167 digitizing microfiche. So they are just uploading things into the 0:24:24.167 --> 0:24:28.847 Internet Archive. And yeah, that's the stream. You're watching the 0:24:28.847 --> 0:24:34.006 internet be recorded into this giant archive. And yeah, um, 0:24:34.007 --> 0:24:40.047 it's accompanied by relaxing lo fi beats. And there's a 0:24:40.047 --> 0:24:43.967 Terraform MCP server integration tool between AI assistants. And Terraform 0:24:43.966 --> 0:24:48.446 helps you discover, explore and understand infrastructure code faster. Uh, 0:24:48.607 --> 0:24:52.167 got away here to use git as S3. Someone teaches 0:24:52.167 --> 0:24:55.087 data visualization with a bag of rocks. There's a new 0:24:55.087 --> 0:25:00.327 paradigm for psychology proposed here. Um, control systems as mental 0:25:00.367 --> 0:25:02.407 building blocks. I actually ordered this book. I can't wait 0:25:02.407 --> 0:25:05.566 to go read it. Um, the book is called The 0:25:05.567 --> 0:25:08.887 Mind and the wheel. Um. Oh, yeah, I gotta I 0:25:09.017 --> 0:25:11.617 got a go download this thing. I think it's already 0:25:11.617 --> 0:25:21.497 in my audible. Hold on one second. We're doing it live. No, 0:25:21.497 --> 0:25:30.577 there is no audible version. Okay, well, whatever. That's annoying. Um, anyway, 0:25:30.577 --> 0:25:32.777 I really want to learn this, uh, or check it out, 0:25:32.817 --> 0:25:36.177 so see if it's cool. Um, so it basically says 0:25:36.177 --> 0:25:38.497 the mind might function as a collection of control systems 0:25:38.497 --> 0:25:42.736 that regulate everything from hunger to loneliness. And, uh, yeah, 0:25:42.777 --> 0:25:45.456 I'm not sure if this is the dumbest thing ever 0:25:45.456 --> 0:25:47.897 or the coolest thing ever, but worth including for you 0:25:47.897 --> 0:25:51.456 to check out. Tycho is a new API that returns 0:25:51.456 --> 0:25:54.577 visual answers from structured data sources in response to natural 0:25:54.577 --> 0:25:59.577 language queries. So if data is trapped in databases or 0:25:59.577 --> 0:26:03.137 web crawlers, you can now go and basically get that 0:26:03.137 --> 0:26:07.857 data with natural language queries. A database of 1400 startup 0:26:07.857 --> 0:26:12.217 ideas from Hacker News and Reddit. Decibels are silly. That 0:26:12.216 --> 0:26:14.536 was a cool piece, basically talking about how the measurement 0:26:14.537 --> 0:26:18.736 doesn't make any sense whatsoever. Google Gemini advanced ads GitHub 0:26:18.736 --> 0:26:22.137 integration for code analysis, timeline views, and smart filtering on 0:26:22.137 --> 0:26:24.537 Hacker News. It's a new Hacker News interface with a 0:26:24.537 --> 0:26:30.016 lot of filtering options undetected. Tag stealth tracking for stolen items. 0:26:32.257 --> 0:26:36.817 And ideas entering into the Member edition stuff here. Ideas. 0:26:38.017 --> 0:26:40.817 So the idea of the season high agency and ideologies 0:26:40.817 --> 0:26:43.137 as tools versus values. So a lot of interest right 0:26:43.137 --> 0:26:46.536 now in what people are calling this high agency thing, which, uh, 0:26:47.657 --> 0:26:50.297 I think there was a yeah, George Mac came on 0:26:50.297 --> 0:26:54.617 Chris Williamson Modern Wisdom and I love the concept, and 0:26:54.617 --> 0:26:57.016 I think a lot of people need more of it. 0:26:57.017 --> 0:27:00.976 But the I would say most people, but not everyone, uh, 0:27:00.976 --> 0:27:03.816 we actually talked about this in book club and uh, 0:27:03.817 --> 0:27:06.137 someone in there was basically saying, look, this this might 0:27:06.137 --> 0:27:08.987 be really nasty. You know, it seems kind of elitist 0:27:08.986 --> 0:27:12.827 or like very talking about the NPC thing. And I 0:27:12.867 --> 0:27:16.067 think anything taken to an extreme can be bad. Like 0:27:16.107 --> 0:27:18.066 if you have too much capitalism, you need socialism, you 0:27:18.067 --> 0:27:20.907 have too much socialism. You might need some capitalism. Uh, 0:27:20.907 --> 0:27:23.547 too little self-esteem. You need more. But narcissism is bad. 0:27:23.587 --> 0:27:27.387 You know, Stalinism is bad. But maybe Nordic countries got 0:27:27.387 --> 0:27:30.067 something figured out. So one idea I put out there 0:27:30.067 --> 0:27:33.227 was basically a set of fundamentals, like I talked about, um, 0:27:33.347 --> 0:27:36.147 in my political post where it's like, these are my 0:27:36.186 --> 0:27:38.787 core values. This is what I believe, you know, humanity 0:27:38.787 --> 0:27:41.067 should look like. This is the ideal world that I 0:27:41.067 --> 0:27:44.547 wish we lived in. And that's one set of things. 0:27:44.547 --> 0:27:49.627 And then if someone says capitalism or whatever, some, uh, 0:27:50.387 --> 0:27:54.107 you know, Marxism or some other thing or some sort 0:27:54.107 --> 0:27:56.587 of ideology, think of the way I'm thinking of this 0:27:56.587 --> 0:28:00.346 is those are tools. They're separate from values. So are 0:28:00.347 --> 0:28:03.827 you a capitalist? No, no, what I am is a 0:28:03.867 --> 0:28:07.306 Star Trek liberal, right? I want to elevate humanity. I 0:28:07.307 --> 0:28:10.956 want humanity to be more, um, enabled. I want everyone 0:28:10.956 --> 0:28:14.597 to have equal opportunities. I want I want to raise education. 0:28:14.597 --> 0:28:17.357 I want to have everyone, you know, participating in creativity 0:28:17.357 --> 0:28:21.036 and exchanging creativity with other humans. I want technology to 0:28:21.077 --> 0:28:23.637 be in the background, not in the foreground. I want 0:28:23.677 --> 0:28:27.077 humans to be in the foreground. Right. So those are 0:28:27.077 --> 0:28:29.957 the core things that I think about kindness, helping others, 0:28:29.956 --> 0:28:34.517 exchanging value with others. Right. Uh, freedom. All these core 0:28:34.557 --> 0:28:38.717 core value things. Now, if someone hits me with, you know, 0:28:38.757 --> 0:28:43.877 high value or high agency or Marxism or whatever, I'm 0:28:43.877 --> 0:28:46.797 going to listen, right? Okay, cool. Sounds awesome. But that's 0:28:46.797 --> 0:28:49.957 a tool. That's not something that replaces one of my values, right? 0:28:49.997 --> 0:28:52.397 It's a method for getting there. So if I don't 0:28:52.397 --> 0:28:55.597 have enough socialism in my life to achieve my values, 0:28:56.437 --> 0:28:59.157 we should sprinkle some more on there. If someone sprinkles 0:28:59.157 --> 0:29:01.797 too much and I only taste socialism on my sandwich 0:29:01.797 --> 0:29:04.837 or my steak or whatever it is, well, that's too much. 0:29:04.837 --> 0:29:07.077 It went too far. And now you're actually conflicting with 0:29:07.077 --> 0:29:10.327 my values. We need some more capitalism. Cool. Let's do 0:29:10.327 --> 0:29:14.286 some more capitalism. It goes crazy, becomes like a crony capitalism. 0:29:14.287 --> 0:29:17.527 It becomes banana republic. It becomes all this stuff. Well 0:29:17.567 --> 0:29:19.606 guess what? We got too much capitalism there need to 0:29:19.607 --> 0:29:22.846 take some of that off. So the key thing is 0:29:22.847 --> 0:29:25.647 to be guided by these values and then use everything 0:29:25.647 --> 0:29:29.487 else as like these different slider bars of tools that 0:29:29.487 --> 0:29:31.887 you could use and adjust to how much you, you know, 0:29:31.927 --> 0:29:35.407 adjust to taste, with the guiding principle being not affecting 0:29:35.647 --> 0:29:39.207 the actual values of what you're trying to do. So 0:29:39.207 --> 0:29:42.007 that was the idea of the week. Um, got a 0:29:42.007 --> 0:29:45.967 simplified I security term breakdown, which, um, which I basically 0:29:45.967 --> 0:29:49.687 modified off of my buddy Joseph Thacker's list. So I 0:29:49.727 --> 0:29:54.687 alignment don't kill us. I safety don't help us do 0:29:54.687 --> 0:30:00.247 bad stuff. I security basically I appsec oh we already 0:30:00.247 --> 0:30:04.847 talked about discovery so we're good and now recommendation of 0:30:04.847 --> 0:30:08.086 the week embrace chaotic reading if that happens to you. 0:30:08.087 --> 0:30:11.727 I'm currently reading letters to a young poet, the Dharma Bums. 0:30:11.967 --> 0:30:15.286 Don't tell me I can't. Rilke's Book of Hours, the 0:30:15.287 --> 0:30:18.447 Pleasure of Finding Things out, The Peregrine, The Rise of 0:30:18.447 --> 0:30:25.007 Theodore Roosevelt, Accelerando Dominion birds, Sex and Beauty, the player 0:30:25.007 --> 0:30:27.927 of games. And they thought we were free. I'm usually 0:30:27.927 --> 0:30:29.367 reading like 2 to 3 books at a time, so 0:30:29.367 --> 0:30:32.487 this is way more than usual. Maybe more than ever. 0:30:32.527 --> 0:30:35.567 What is this? Is this, like 20 books? Like 15 books? 0:30:35.567 --> 0:30:38.407 Something like that. But I'm learning to be okay with this. 0:30:38.447 --> 0:30:41.247 And this is my recommendation. This is my ask for you. 0:30:42.207 --> 0:30:44.207 Be okay with it. If you're only reading one book, 0:30:44.247 --> 0:30:46.447 be okay with it. If you're reading 15 and you're 0:30:46.487 --> 0:30:49.326 switching between them, be okay with it. The one thing 0:30:49.367 --> 0:30:52.007 to not be okay with in my mind is not reading. 0:30:52.007 --> 0:30:54.647 So just read one book at a time or 17. 0:30:54.687 --> 0:30:59.407 Don't be frustrated with yourself, just read. And the aphorism 0:30:59.407 --> 0:31:01.967 of the week. The most important decision you make is 0:31:01.967 --> 0:31:05.047 to be in a good mood. The most important decision 0:31:05.047 --> 0:31:07.007 you make is to be in a good mood.