UL NO. 496: STANDARD EDITION | New Video on Building my Personal AI System, Anthropic Reveals One-person Hacking Company using Claude, Pentagon Says China Keeps Penetrating, and more...

0:20S1:All right. Welcome to episode 496. All right, so my video on building personal AI system using cloud code is out. This thing has been a blocker for me in like life and work and just it's taken so long to get this thing out. Recording it took like 5 or 7 takes or something. The blog post, the editing, everything about it has been difficult. I am so glad it is out and recommend you go check it out. Just upgraded to a Keychron Q11 keyboard, it is a canted and tilted. I can't remember like which words associate with which angle, but it's basically tilted inwards because that's the way your hands are when you type. And then it's also tilted upwards on the inside. So it's like an upside down V. Kind of not as extreme though. So if you imagine you just put your hands, your elbows by your side and you put your hands out, they're going to be like angled at like a, an upside down V shape, you know, slightly like a wide open upside down V. And they're also going to be moving well if you stick them straight out they will be straight out. But if oftentimes you'll want to slightly bent them inwards or, you know, uh tilt them inwards or

1:54S1:not tilt but um, angle them inwards. And if you do that then of course you can um, you could reach the keys on a regular keyboard that's lined up right now on a regular keyboard that's lined up just the way a traditional keyboard is. You actually have to tilt your your wrist wrists inward, and it's just not it's not good for you over time. I don't really have any pain that I'm dealing with and trying to address here. I just want it to be more natural.

2:28S1:I want to be able to type faster and I don't know. I also think it looks cool. It feels cool as well. So I like that. And then I'm experimenting with, um, the keyboard because the keyboard, basically it has um, it's two separate pieces. Okay. So the left is separated from the right. That's like the most important feature. Um, and that allows you to do this, this canting and everything completely separate for left and right. Um, obviously you want them to match. You want it to be symmetrical, basically.

3:03S1:But the idea is you can have them all the way across, like the distance for from your elbows, uh, one elbow to the next, which is what, you know, One, you know, two feet, maybe something like two feet apart. You could do that and then your hands would be just straight out. And a lot of people who do like the custom keyboards, they actually do that. I have mine probably at the bottom of the keyboard. It's probably separated by, I would say, um. Actually, I have a tape measure. Like five six inches is how far the bottom is and the top is only a couple of inches separated because again, I have it, you know, canted downwards in that way. And it's pretty cool. So the inside of the keyboard I have elevated. So it's in that upside down V shape. And yeah, it just feels really natural. Feels awesome. I've tried some other more customizable keyboards that are just like 60. They're called 40% keyboards and those you have to use layers to be able to get to a lot of the keys. Um, I didn't find that very fun. Um, the muscle memory there was not easy to grow, and I feel like there's not a reason to get rid of that many keys

4:41S1:and to force you to have to do layers. I think this is a better approach. It's a more traditional keyboard. So I have numbers above my fingers. I still have my spacebar on both sides. I still have number keys in the bottom right of my right hand. So it's like it's more classic, more traditional, but at the same time just oriented in a more ergonomic way. So again, it's a Keychron Q11 and I do watch too many keyboard videos on YouTube, so this might not be my last keyboard. I have way too many keyboards, unfortunately. Okay, um, found a major problem with doing member essays here in the newsletter, so I'm not actually going to do them anymore.

5:36S1:And the problem is very simple. It's like I'm what I'm trying to do is orient my work and my life around the central theme of ideas. Right? I've been trying to do this for years now, and the central theme of ideas is you try to have ideas first of all, and then you try to get them out into the world. And what I would do is I would be writing a newsletter and I would come up with a cool idea, and I would do a member essay on it, and I was going to do like one a month, but I ended up doing it like almost every month. But the problem is, they were so decent I thought they were halfway good a lot of times.

6:19S1:At least several of them were halfway decent. And I'm like, well, that just needs to be a regular blog post. Like, I can't, you know, just it doesn't make any sense to put the effort into it. Have people actually enjoy it, but it only goes out to like 1500 people or whatever the current member count is, right? So it's like it just it was not in harmony with me trying to get ideas out into the world as like a primary mission. So. So I'm not going to do that anymore. Um, basically I'm just going to turn those into blogs, right?

6:58S1:So you're still going to get them. They're still going to be in the newsletter. It's just you're not going to have to pay for them. And instead I'm going to basically raise the member perk that has been most asked for and just make that the primary perk. And basically a clearly stated explicit thing of 25 to 50% off all UL paid content. So when human 3.0 launches, when augmented online, the online version of augmented launches, which will be sometime this year, sometime soon, like you're just going to get a massive discount like hundreds of dollars or in some cases for courses or whatever. Or, you know, conceivably if I charge enough for a thing, it'll be thousands of dollars off for members, which is kind of ridiculous given that the membership is only like $100. So I don't know. I'll figure that out later. Uh, content is more important in the meantime. So yeah, you can subscribe at Daniel or subscribe. My friend Monica is running her Resilient Career Accelerator course again soon. I'd recommend you check it out. It's a week long course. How to build a career in security leadership. Run a security program

8:20S1:and generally make yourself more resilient career wise in this AI world. And she talks about that a lot. She's very bright, very passionate about these topics. And I'm also going to be a guest speaker at the course that is in November. Uh, very much in line with my recent Westerberg shares. You probably seen some of those videos from her. I'm loving this guy. His name is Mainstone, or at least that's his brand or whatever. And he's just like this. Extremely honest, like British guy. And, uh, yeah, he wears sunglasses and talks at the camera sometimes, but he also is a videographer, photographer, videographer. Anyway, he takes videos of like, terrain. And so his whole theme is like going back to basics. It's like, what are we actually doing here? It's a very existential, like, if I had to categorize this YouTube channel, it's an existential YouTube channel.

9:23S1:And he's just he's smart. He his takes on AI, I think are pretty clean. Like some of the cleanest that I've seen, which I find to be hilarious. Like this guy who is I don't know, I'm not even sure how technical he is. Um, computer wise, um, video wise, he's like, obviously extremely technical. The guy is. Yeah, he's he could probably be a professional, professional video guy if he isn't already, but I'm not sure if he's like a computer guy. I don't really know that. Um, he mentions being broke all the time and having like, $1,500 to his name and like, trying to find work. So I'm guessing he doesn't have or hasn't had like a tech career. That's. I'm guessing he doesn't have that to fall back on if he's talking about that. But in the economy right now, who knows if that's actually the case?

10:21S1:Could be that he just can't find a job there. But anyway he goes his one of his favorite things to do is just like to disappear and just like go far, far away. He's also really good at Lan Nev, by the way. Um, because I'm not that good at Lan Nev. Or at least I'd have to train up. Um, I think he does carry, like, a satellite device and, like, batteries and stuff. So he could, like, actually get signal if he needed to. I think he gets, like, weather reports, but he's going places like very few people have been.

10:56S1:And he, like, disappears, like in, like, random parts of Egypt or something, or like some random mountains in, like Alaska or whatever. Just like disappears for like seven days. And he's doing video the whole time. He's talking the whole time. And it's really cool to see him actually talk to the camera when he's feeling these extreme, like emotions of solitude and like the wisdom he's getting from it and everything. And his voiceovers, his narrations are great. Um, his honesty. Like, he's just like, look, I just can't figure this out. I don't know what I'm going to do. Like, he's just very plain and open. It's extremely refreshing, very authentic.

11:39S1:Highly recommend it. Especially if you're kind of thinking, you know, what is all this tech for? I want to go back to basics, you know, analog. It's got that sort of vibe to it. It's got the existential vibe to it, gorgeous scenery as well. Like some of these shots he's taking. Just absolutely amazing stuff. So anyway, highly recommend that. We read Epictetus, the Enchiridion for book club. And it was wonderful. Absolutely wonderful. Um, also reminding me so the Westerburg thing.

12:19S1:This guy mainstone, I don't know, all my stoic studies. A million different things have, like been bringing me and I don't know the attribution amounts, so I can't really properly give credit. But I've also been reading a bunch of, like, you know, dead tree books, uh, with no tech around me. And I feel like whenever I get an insight from a real book, an old book with text on a page, I feel like the inside just hits me harder than, you know, all the cool shit that I'm getting from YouTube, right?

12:55S1:And I'm getting great distilled, highly produced, great stuff from YouTube. And somehow the book stuff just hits me harder. And I'm trying to disentangle Like, is it because it's better content? Is it become because it's coming from a different angle where I'm not ready for it and I'm like surprised by it. But one of my possible frames for this is that it's simply like. One of the models I was thinking of is like, I heard about on Reddit somewhere, that some people use tea bags like 12 times and like they keep track with like a little counter of like how many times it's like, oh, 11th time. Well, we haven't got enough out of it. So they keep using it. And I'm like, I wonder if the content that I'm getting from YouTube and from like, all these, you know, modern sources or whatever, it's just like used tea bags like 12 times used tea bags, it's just it's good, but it's massively diluted. Just like, not the good stuff. And I wonder if the stuff that I'm getting from these books. I don't know, I can't tell if it's just a better tea leaf, or if it's just an undiluted tea leaf, or if it's just. I'm

14:15S1:not used to taking tea leaves by snorting them or shoving them into my ears like, I don't know if it's the medium. Like I'm still trying to figure that out. So if anyone has any ideas there. But bottom line is I am getting extremely excited about this concept of. Old sources. Classics, but not even classics like literature. Literature. I'm getting so much out of the biography of. Roosevelt. Teddy Roosevelt. I was thinking of the other one. I get them confused. But yeah. FDR is the later one.

15:03S1:He was the economy thing. FDR policies and I think New Deal and all that. But Teddy Roosevelt was much younger. He was doing things early 1900s or actually, no, 1800s. Yeah. Um, 18 like 76 ish or whatever. Uh, roughly around there. I think he was in his teens, like 100 years after the country was born or something like that. Garfield was assassinated, uh, during that time, something that just happened in the book. But anyway, uh, he was. Yeah. This biography was talking about his life, and he was talking about, like, um, so he was going through his journals. That's what the biography is largely source based on is his journals like Roosevelt wrote tons. And the guy was just absolutely insane, like, lie.

16:00S1:Completely insane. He would just. He was like a boxer, like a fisher, like dancing, but not very well. But he just spoke multiple languages and literature, like he just read everything. And recitations, evidently was a huge thing at Harvard, going from place to place, doing recitations, like, I don't know, is that people reading literature is that people reading memorized literature without the book. I'm not sure exactly what that is, but I need to go research that. But. I feel like this content I don't again, I don't know if it's the tea leaf itself or if it's the medium, or if it's just different and it's just a different style with different, you know, diction and and it spawns thoughts in the brain differently. But I just feel like people were so sharp back then. Um, I don't know.

17:03S1:I think there were extremely limited in other ways as well. So it's hard to, uh, you know, measure it on balance. But anyway, I'm getting extraordinary creative, uh, boosts from reading these older books. I'm reading a book right now about, um, rhetorical figures. It absolutely insane. It's like it's rhetoric, but specifically about something called figures in rhetoric. Uh, it's a book about writing, and I'm just loving every single page. I also just love how the page is written. I love the writing itself in this book. I think it's using like it's tricks that it's teaching me. It's using them on me while it's teaching me. Pretty sure it's doing that. But what I find is it's very similar.

17:59S1:I got to post somewhere like my favorite writers. So Hitchens, Scott Adams, the Dilbert guy before he went crazy, although he probably still writes well now, um, not so much Sam Harris. I like him for a different reason, but the writing style that I like the most, and that I use and probably emulate quite a bit is this alternate alteration alternation between long and short sentences? It's very powerful. I'm seeing this happen in this book here again. Um, Hitchens did it. It's it's in several books that I've read about good writing, so I'm not sure what the actual source is. I'm not sure if it's one of these figures, these rhetorical figures. But anyway, really, really powerful stuff. Um, and we are 700 minutes into this podcast. Um, let's get into it. That was just the intro. Cybersecurity. Anthropic says attackers are using models to automate full hacking operations.

19:09S1:So this is the thing I talked about in my Future of Hacking video, which you should definitely go look at if you haven't seen that. Um, and that's just a derivation of or derivative video, a hacking focused version of my unified entity context, which is like this unified context for the enterprise that everything runs off of, where the AI talks to it. And it's like, you know, the context is all in one place. Well. Anthropic just came out and basically said, look, we've we were tracking this actor, and the actor basically was acting like a giant team, like, like a team of multiple people. And this one person accomplished what normally takes large teams, you know, months to do. So, uh, 17 organizations in weeks using Claude.

20:08S1:So somehow they were getting Claude to do all this, by the way. So they were bypassing whatever restrictions Claude had healthcare, governments, defense contractors, even a church. It, uh, Claude analyzed stolen data, created customized extortion plans for each victim. North Korean agents use Claude to fake IT worker identities at companies. They asked Claude basic questions like what is a muffin to maintain cover? So this is one of the, uh, prompt injection bypass techniques, or at least, you know, to bypass the policy of, like. Hey, you seem to be suspect. Uh, fake employees can earn high salaries. Romance scammers use Claude as their emotionally intelligent conversation engine. That's diabolical. Chinese hackers use Claude to guide espionage against Vietnamese telecom companies. Criminals are building ransomware as a service. Businesses powered by AI.

21:09S1:So all of this is exactly what I've been talking about here, where it's like it's their AI system against our AI system, right? It's them defining goals and the AI basically using all its different tools, which is essentially what I just talked about in my personal AI video, where I'm talking about, like, you have a central system and you give it all these tools and you say, look, go do this, and it picks the tool to go do it, and it does it. And it looks like a system probably very similar to what I've built is being used by this attacker. And I'm sure many thousands of others, or at least hundreds of others soon to be thousands. And maybe, I don't know, maybe more than that later. I just think this is the way to go, right?

22:02S1:Everyone's going to have an automation stack. The question is what are they using it for? And so what do you have to do as a defender. You've got to have the same exact automation stack. That was my takeaway from this. You know future of hacking is context. Video is basically what's the answer. The answer is to have the same thing that the attackers have. And the framing that I have for all of this is you have to have a, a world model of your system and your company and your employees and your projects and your budget and everything. You have to have a world model of that whole thing, and you have to be able to audit that and understand its state at any given time.

22:47S1:And then at the bigger picture, the whole game is. Excuse me. Take a drink. At the bigger picture. The ultimate game is going from current state to desired state. And. Anyway, that's kind of a bigger topic, but the point is, attackers that I made before is attackers will have this world model of your company, and they will use that to use their automation system to continuously launch attacks against you. And basically what your logs are going to look like is thousands of AI automation systems just hitting you constantly.

23:42S1:Just constantly. And if you fart or you burp or you step out of line one second With like an S3 bucket or a misconfiguration or, you know, a DNS domain that you're not maintaining whatever it is that is going to get jumped on by a whole bunch of AI who does not sleep. Right. So your attacker is absolutely going to jump on that, and they're going to do it fast. And if not, one of their competitors is going to do it. So we're not going to have the time. So our only chance is we get there first. Our automated AI system sees that thing pop up first before them and goes and fixes it, or blocks it, or does whatever, you know, puts up a firewall rule or whatever. So this is the game. To me, this is the security story. This is the number one security story going forward. Everyone talks about, you know, AI and security. Hey, what are the future trends or whatever?

24:56S1:This is my number one answer. Number one answer. Unified entity context to get everything unified inside of a company. So the future of basically company management is the founders. The owners sit at a giant table with a bunch of other leaders who are also extremely business savvy, creative and and technical as well. And then there is a number of similar minded people that are basically like wizards, and that essentially is their IT staff. And their IT staff isn't really IT staff. It's more like business IT liaison.

25:45S1:So it's like they are experts at understanding exactly what the leaders at that table just said, because they're at the table as well. But they're kind of like the execution arm. And then what they do is they they're like, okay, I just understood that I am parsing that and bringing it into the unified entity context that is now just updated. Our goals. Our goals have now been updated. All our plans have been rewritten. Um, these additional details have been written into all future iterations of all of our products.

26:21S1:For all the products that you just mentioned building, we have now started specking those out and they're basically wielding these giant teams, these giant armies of. AI systems. And those AI systems are the planners and the speakers and the, you know, they're going to build marketing campaigns and they're going to build all this different stuff around that. But it has to start at that table with the people deciding actually what to do. Right now, of course, there's arguments that, okay, well, in the future you don't need them because, you know, the AI will just be smarter than all of them. Honestly, I believe that. Right? And that's that's the ACI story. Very possible. But that's not a security thing. I care about right now. It's just not tangible to me yet. Maybe in five years, maybe ten years, maybe 20 years. And who knows, it could be two years. But for the foreseeable future, and I think it'll be a good amount of time, even if ACI hits, it's going to take a while. It's going to be that leadership team in the middle, um, with an enforcement arm of these extremely bright people wielding giant

27:39S1:teams of thousands or millions of tiny little AIS doing execution. And in terms of security, really, it's about everything. It's continuous monitoring and continuous context management and understanding of the business. But especially for security, you have to continuously understand the state of what you have, which will previously have been called asset management. But really it's like state management of the business overall. And then once again, it's your AI systems against their AI systems. It's just that's just reality.

28:20S1:How fast are yours. How well do they understand the context? How small is the gap between a change and an understanding of the change that is the game. Okay. Um, anthropic shows how they detect sleeper agents hiding malicious behavior. Yeah. Catching AI agents that pretend to be helping while secretly planning harmful actions when deployed. Nasty FBI confirmed Salt Typhoon breached at least 200 American companies plus firms in 80 countries. NSA and NCIS trace the Salt Typhoon espionage campaigns to Sichuan Jianshui Beijing. How you knew? Why would I? Why would I choose to read these things? It's not going to help me. I don't want to butcher people's names. Um, but these are companies. Actually, not not people. So those are different companies providing Chinese company providing cyber tools to China's intelligence services. Uh, Claude will start training on personal account data unless you opt out. So this is an important one. They you probably seen the emails if you're a user. Uh, they emailed everyone and basically said you have until September 28th to opt out. I wouldn't opt it out immediately. Um, if you don't want your data

29:48S1:to be trained on and if you have like a custom like enterprise account or certain types of, you know, paid accounts, they won't do this by default. But if you just have a regular account, including like a pro account or a max account or whatever, um, they will turn this on. They will start using your content for training. Uh, if you do not opt out. So definitely go and do that if you want to. I'm also curious about why everyone's freaking out about this. Don't most other AI companies already do this? Like, I feel like this is just pretty standard at this point. Um, what I don't understand and what I wouldn't like is if they're like, we don't have the ability to opt out before September 28th.

30:34S1:Why would they say before September 28th? I guess that means. I guess that just means that they will turn it on. They will turn on the use after September 28th, but hopefully it also means that you can still opt out after that point. I'm not sure they made that clear in the email. I didn't see it in the email, but hopefully that would be the case. It would be super shitty if they were like, nope, we sent you the thing. I told you you had to do it before September 28th and if you didn't? Too bad. That would be weaksauce. And the more I think about it, that's so ridiculous. I don't think that's going to be the case. Researchers hide malicious prompts in images that only appear when AI systems downscale them. This is from Trail of Bits. Really cool attack. Uh, worth checking out. DoD uses node JS utility maintained by a Russian Yandex developer.

31:41S1:Supply chain Transparency. This is why I can't wait till there's like, you know, millions upon millions of AI eyes watching everything. Oh, here's a cool idea that, um, I'm going to be doing a panel with my buddy Sasha's dealer pretty soon. And by the way, sorry for my voice. I'm just realizing because I'm trying not to sniff while on the microphone because it's a sniffing sound. But if I'm trying to sniff, that means I have the sniffles, which means I probably sound a little bit nasal, so apologize for that. Not sick, but I'm like one quarter sick.

32:24S1:I just don't feel bad. Anyway, uh, the idea here is, you know how we were told many eyes sunlight is the best disinfectant and all this. And then we have the SSL thing, which was open source, and it was a massive, massive vulnerability across the whole internet. Well, what if AI is basically a revisiting of many eyes, right? Because Many Eyes turned out not to be true. Not to be real. Because many eyes implied that because people could look at the code that they would, that they actually would. Right. And that turned out not to be the case. Nobody was looking at all these open source projects.

33:10S1:Just because they're open source doesn't automatically make them secure, because that implies a second step, which is lots of people are actually taking the time to look. Why are they doing that? What time are they doing that in? How are they incentivized to do that? There's no answers to those questions, which means which is the reason that it wasn't being done. But what if we could just unleash like this army? Like the government can unleash an army of millions of agents, and their job is to literally just go and clean up all this stuff. And funny enough, that's exactly what the AI project was about.

33:54S1:Guess who? It was launched by DARPA. Yes, the creators of the interwebs. So they literally created a project that does this exact thing. It autonomously, autonomously, using agents can go and find bugs and fix them end to end without breaking the software. Isn't that insane? And this is what? Trilobites. They just won second place. This is what Michael Brown was talking about, which I haven't interviewed, which you should go check out. It's on YouTube. And yeah, this is exactly what the government set out to build. And this is like, this is many eyes. That is literally what many eyes was supposed to be. It was supposed to be humans. Turns out that doesn't really scale. Um, because open source turns out to be just one guy. I saw somebody say that recently. Actually, it wasn't a person.

34:55S1:I just saw it as a headline. Open source is just one guy. Maybe it's actually in the, uh, in the discovery section. Open source is just one guy. That is hilarious. Uh, anyway. FBI warns about three phase phantom hacker scam targeting seniors. Scammers use AI to analyze social media and personalize their attacks. Oh yeah? What do you know? First, they pose as tech support and gain remote computer access. Then they fake bank employees, convince victims their accounts are compromised overseas. Just so nasty, so nasty. Just going after seniors.

35:35S1:And it's just going to be done at scale now. national security. Pentagon security agency admits China keeps beating their defenses. Sad. AI Cloudflare radar tracks global AI bot traffic across the internet. This thing is really cool looking. Yeah, it shows you all the stats for all the crawlers and everything. Super super cool. AI is the natural step in making computers understand humans. AI browser agents fail at real work. Except for one unknown tool. Yeah, this was a really interesting one.

36:21S1:You should go check out that one in the newsletter. AI adoption linked to 13% job decline for young US workers. Yeah, I saw one of my smart buddies who was like, yeah, you know, there's no evidence that AI is affecting the job market. I'm like, what are you reading? Like, like the CEOs, so many founders are like, yeah, I'm not hiring anymore. I'm looking into how I could do more of this with AI. And it's just like dozens of these people coming out and saying this. The job stats are going down at the same time. And it doesn't mean like that's necessarily proof, right? You know, it's correlation, right.

37:04S1:But then you just see more and more people saying, yeah, I'm automated this, you know, I have automated this. I was told directly that the team was being laid off because of AI reasons. Like, there's just so much direct evidence that we have from people, you know, which is like anecdata you add those up and they turn into like almost data. So 13% I think is low. I think it's just just now starting. And yeah. And actually I have things queued for the next newsletter that, um, that backs that up. There's actually more studies coming out.

37:44S1:It's probably going to be way more than 13%. And yeah, only getting worse is my guess. Creators are automating their entire content workflow with AI. So you spot trending topics with AI agents, trained Claude on your style. Clone your voice with 11 labs, create your digital avatar with hey Jen, generate B-roll with Vo three and runway ML. Merge everything in Canva. Automate distribution across platforms. Complete the workflow. So that is basically the dream, right? This is like end to end. They're just doing everything or at least saying that it can be done. I love this. I think all anyone should have to do, like as a creator, is think about their thoughts and their ideas and like how to get them out there and like the format and the style that they want to get them out there.

38:41S1:But all the busy work that isn't fun, of course, obviously still do the fun stuff. All that stuff should just be automated away so you could focus on the actual creativity part. So yeah, interesting. AI coding makes me faster but kills my flow state. So this guy Praful is talking about how. He used to be able to just listen to music and code, and now it's like you're dictating or you're typing or you're going back and forth with this AI. Uh, for a lot of the code generation. And it's just like, it's not the same as just listening to music and coding. And I thought that was really, really interesting. I hadn't realized it until he said that. And it's definitely true for me. ChatGPT usage drops hard when summer vacation starts. Yeah. This is sad for me because it means they're not doing full life integration with with the stuff they're mostly using it for like homework and research and stuff like that, which is fine, but I would like to see it more widely adopted. Stop calling automation AI unless it actually learns something.

40:01S1:I included this just because I want to include counter opinions to my own. Um, yeah. I think it's fine that we want employees like it's obvious that we want employees who can learn, but that's not the main thing that we pay employees for to learn. We the main thing we pay employees to do is to. Execute the tasks. Right. So I think the learning and the executing of the tasks are two separate things, because if we had AI that could just execute the tasks, then the learning thing would not be as important. Um, because what's more likely to happen is that you would simply have, you know, separate systems that are like, okay, we detected a sufficient change in this thing. Therefore, you know, some retraining needs to, you know, be added, whether that's a, you know, a fine tuning of a model or whatever it is. And then, um.

41:09S1:Or just prompting or context or whatever it is, the method you're going to use to do the improvement and then you roll it back out. And the other thing is, as the stuff gets better, I think. A lot of that will generalize into being the same task and maybe not even require too much learning. Again, just keep in mind that just keep in mind how similar problems will start to look when eyes are seeing all the problems for all these different companies. How different really are the sales use cases and the marketing use cases, and the product management problems and the writing code problems. Right? It seems to me like the hardest part is the creating of.

42:05S1:Net new things. It is the true creativity. It's the true coming up with new ideas. And guess what? That's not what we're paying most people for. We're paying them to execute on these other tasks, which are the other 99.99% of knowledge work. So I think this argument about, you know, I'm paying you to learn. And A's can't learn on the fly. Really good. Right now. Therefore, it's not a threat to us. That is not a good argument. DSP turns prompt engineering from art into systematic optimization. And this is a towards data science article. Really cool. DSP is a really powerful framework. It's basically automation of intelligence tasks in pipeline type structure where it's repeatable and consistent and you can depend on it and kind of like an enterprisey kind of na to n type of vibe technology.

43:13S1:Three mystery customers account for over half of Nvidia's data center revenue. That's spooky to me as like an investor in Nvidia, which I currently don't hold any Nvidia, but if I were holding some, that would be scary. Like, who are those people and can they go away? And if so. What would happen to their stock? It's more likely that they'll gain five other unnamed ones and they'll go to the moon. But I don't know. It just seems weird to me if like so, such a high percentage of like they're good numbers are from these unknown people.

43:52S1:It's like one of them Satan. Like, I don't know. I just don't feel good about it. Venture capital in trouble. Based on SEC filing patterns, TJ Jefferson discovered that formed filings containing fund one peaked in Q3 2022 and have been dropping hard ever since. Pretty cool analysis there. CrowdStrike buys On'em for $290 million to build a genetic socks. Computer science grads are struggling to find new jobs despite demand. Entry level tech hiring drops 25%. Who knows where they're getting that stat from? There's a million different of these studies.

44:38S1:I wouldn't be surprised if it was 25%. I wouldn't be surprised if it was 65% drop. Yeah, like many applying to places like Chipotle while software engineer postings remain down. Oh, here we go. Software engineer postings remain down 70%. Cognitive load is what actually matters in software complexity. Good piece there. Humans. South Korea bans smartphones in schools starting 2026. Great job and great job to probably Jonathan Haidt. Therapists are secretly using ChatGPT and clients are upset about it. Middle aged are no longer the most miserable. Turns out it's now younger people. Younger people are now more miserable than middle aged.

45:32S1:It used to be a U-shape. With younger people happy and older, older people happy, now younger people are massively unhappy. US retail giants are raising prices based on Trump tariffs. Shocking everyone. A brain abscess gave Eric Markowitz the clearest moment of consciousness he's ever experienced. Describes how facing death from a cerebral abscess made him truly conscious for the first time. Trade school enrollment surges as Gen Z ditches degrees for HVAC and welding. I got an idea here.

46:21S1:Why go to school right now? Why go to college? I still think you should. So I'm of two minds of this. I'm very strong. Like polar minds here. You should still go because of connections. You should still go because you do learn some decent fundamentals. If you weren't to get. If you weren't going to get them some other way. Right. It's a rigid structure. You have to show up to class. You have to get the grades right. Your parents are usually pushing you. So these are like all good reasons to go. But holy crap, compared to having your AI build you a custom curriculum of like the best YouTube videos to teach you computer science. And I'm going to be talking about this a whole bunch actually, in the next few months. I've got this whole concept of slack in the rope and actually what a real curriculum could look like. I'm actually going to be building curriculums into substrate curriculums for computer science and like all these different, um, all these different degrees and stuff. And I'm actually going to go find some experts on this to help build the degrees, to help build the content that needs to be in there. And I'm also going to,

47:38S1:of course, look at like all the classic ones, I'm going to look at Oxford, I'm going to look at Stanford, I'm going to look at like all these different schools, um, including looking at what they used to teach, look at what the Romans used to teach, like the trivium. Right. And basically build like these custom bespoke things. But the point is, I bet you it's pretty easy to put together a set of YouTube videos, um, and maybe some custom content that is like, just gets to the heart of it, right? Imagine learning this stuff from like, Richard Feynman and like the best teachers you could find. And maybe certain teachers are better for certain types of learners, right?

48:20S1:And that could be stacked in there. But there's just no possible way that learning computer science in college is anywhere close, anywhere remotely close to like optimizing the cost function. Like nowhere close to ideal. Uh, I know I use I learned computer science in college in a traditional four year college, and I've seen it taught in many, many others. And I know what the curriculum looks like. Anyone could go look it up. They're following the same books, like most of them are reading the same books. Most of the lecture styles are very similar. The homework is very similar.

49:09S1:Like the algorithm is just so ancient for doing this. And. Right now, at this particular moment, we're like four years from now is like 14.5 years from now In the sense of like how much progress is going to be made. What is your opportunity cost for sitting in a class for a semester, trying to learn computer science, versus having an AI build you a custom thing and actually have conversations with you about it? And how about this? It's like, hey, write me some code that does this, and you're writing the code and it's like, yeah, that's not quite right.

49:47S1:It's more kind of like this. And it dynamically changes your code for you and you're like, oh, okay, cool. Or it's like, hey, I'm not going to help you. You go ahead and do this and see if it compiles. Um, go ahead and do this and, you know, test out the website. Okay. Why is it not rendering? Why is my thing off center? Why do my code not compile or whatever? And it's like, I don't know, what do you think it is like a Socratic sort of tutor type thing, except for it never gets tired and you don't pay extra money to ask the question again. Like, how does that even compare To like this, this older system that I just described, it's completely ridiculous. So there's only a matter of time before this is going to become open source. Like I'm going to make it open source.

50:33S1:I'm going to make these curriculums open source. And the part that I don't have that I don't think I can quickly build, and I'm sure other people will and they'll do a great job is building the tutor that can then use that curriculum. Right. Like I have Kai here, but Kai's not specifically trained for being a tutor. Kai is specifically trained right now because he lives inside of cloud code to actually be a, um. Developer, right? Kai is currently built because of his current body and brain structure, to be an executor of tasks, to be a writer of code, and that sort of thing. Um, but, you know, I could reskin him as a tutor. In fact, I'm getting excited about it. Just thinking about it. Like having a tutor mode or a set of commands which are tutor based. Okay, never mind, I'm excited. I'm actually writing this down right now. I'm making some custom commands for tutoring.

51:39S1:I'm really curious about this right now, pairing that together with substrate. Oh my goodness. Tutoring commands. And then open sourcing the whole thing free on substrate and the tutoring to go along with it. Question is how do you make that packageable like as a mobile app? Because it needs to go out to lots of people who need it. Not like the 18 people or the even the 1800 or 18,000 people who see it on YouTube or whatever, and they use it. That's barely making a dent. And it's helping people who already maybe don't need the help as much. Right. So anyway, I digress. Opportunity cost for college.

52:24S1:It's a tough question. I still think in general better to have a degree. But right now, at this particular moment, I am really worried about telling somebody to go and get a four year degree and basically shutting down their mind the whole time. What are you doing when you're in four years of college? You're doing the coursework. You're trying to please that teacher. You're trying to please that curriculum. You're trying to please that old structure of college courses from whenever this structure started. Right. And that is an ancient, broken structure, extremely inefficient. I'm going to give it a 3% efficiency score compared to AI tutoring you based on like custom match content. Custom match YouTube content? Are you kidding me?

53:16S1:Like no comparison? Person. I think we could get like a less ideal is 100. I think we can get like a 60% efficiency up from like, okay, three is too low, up from like a 10%. I think we can get what is that, 5X5X better. And like how does that turn into reality. I don't know. I'm talking about full degrees. Full understanding. Here's what I'm talking about. Having a better grasp of computer science. A better grasp of history a better grasp of psychology, evolutionary biology, political science.

54:00S1:Getting a grasp of those things in one week's time that is better than 99.99% of all people on the planet in one week's time. You can do that with 20 really good YouTube videos combined with active tutoring by your AI in a Socratic, challenging way to quiz you to make sure you're able to move on to the next video. Think about that. That's not even counting. We're going to make you custom videos, right? Which is not here yet. And still coming. Right. That that's next level.

54:43S1:But I think we can get massive multiples of gains just from using the tech we already have. So like I said violently of two minds on this whole college thing. Yes. You should go. Holy crap. Don't turn off your brain. You need to be doing all this AI stuff, all this modern shit at the same time. If you're doing it to get the degree, go do it. Get the degree right. Especially if you're almost done. But I don't think the people being laid off are like, nope, we can't lay that person off. They have a four year degree.

55:20S1:I don't think anybody fucking cares right now, honestly. Um, some people do, okay? Some people do. If you're trying to go into some traditional field, I'm just telling you, the world might be very different in four years. Like, dramatically fucking different. So really, really think about opportunity cost. Best of both worlds is do all the shit that I'm talking about at the same time. Then you don't really take a hit by getting the degree right. Plus, ideally you're using AI to help you get your degree right.

55:55S1:So if it's easy, it's easy. Plus it's friendships. It's maybe sports, it's extracurricular activities. It's social. You're making a friend network. There's tons of benefits to college that aren't related to the curriculum. I'm mostly criticizing the curriculum stuff. Man can't get through these podcasts just talking. Everyone loves it when I do this though, and go into more detail and talk about things, but it does take a lot of time. Um, I got to figure out how to get it down into my old format of just like, read the title and boom, go through, be done in 15 minutes. Didn't I say I'm supposed to be done in 15 minutes? Isn't it supposed to be a 15 minute podcast? Haven't had one of those in two years.

56:43S1:All right. Brain abscess already did that one. The feeling your work creates matters more than the checkboxes. This is by Mitchell Hashimoto, who also made the ghosty terminal, which is my go to terminal. A really good article there. Physical technical books create memory palaces that websites can't match. I was just hanging out with my buddy Tim Leonard, who is a I met through UL and, um, Hanging out in Vegas, and he opens up his his logbook that he keeps with him. And he's got like these custom, like hand-drawn images. So, like, he takes a note or whatever talks about a plant or a, or a thing that he saw with his wife or whatever. He takes these notes. He had like little pressed flowers, like he pulled from a place. It was just brilliant, just absolutely brilliant analog. He's got a pen, he's got paper memories.

57:47S1:I think you can have that with technical books as well. And that's what this article is arguing about. It was it was really cool. MIT and Scripps researchers develop one shot vaccines for HIV and Covid. Really, really cool. All right. Discovery I'm going to try not to talk for 74 minutes about each one of these. I'm going to blast through. I'm going to try. AI continues computing's evolution towards understanding humans. A beginner friendly jiu jitsu tutorial for non git users.

58:24S1:This is an alternative to git. It's called jiu jitsu. It's supposed to be awesome, but I do not have the time for the switching cost right now. The opportunity cost of me trying to learn jiu jitsu right now. Um. And migrate everything. Get off of GitHub. Does GitHub even do jiu jitsu? I don't know. I'm guessing not. Git is in the name. FFmpeg pages simplifies video processing with visual commands. AI agents should use files and git instead of context windows. Sounds pretty familiar to me with like this whole UFC thing I just came up with. Um, That's when I actually didn't read yet. I need to open this tab. Because it's in the discovery section. So I'm just kind of like sometimes kicking off cool links. Um, that you should go check out. Single servers can handle millions of HTTP requests per section. Dot files don't belong in application support. Loved that one. That was so spot on.

59:38S1:Google lets you block AI overviews from your site. Nardin automates bug bounty enumeration via discord, training yourself to like things you hate. Klein runs completely offline with LM studio. New vibe coded web app from Kathy Helps. I don't see many of my women friends doing the AI coding entrepreneurship, so I'm posting this here. This is a woman who just built an app and is having super fun is talking about it, and she's just really enjoying the whole development process. And I just want to see a lot more of my women friends, uh, doing this whole entrepreneurship, like build your own business, you know, AI coding, AI hacking thing. So I'm going to be sending this to a bunch of my friends. Trace prompt provides tamper proof audit trails for LLM applications. Oh, a duplicate got a duplicate there. Claude gets a chrome extension. Physical textbooks become unchanging memory places talked about that one. Wire prompts don't belong in git. Uh so oops oops on that one.

1:00:54S1:Guile's or shares creative bash prompt designs and K-pop Demon Hunters is Netflix's most watched video ever, and I still have not watched this, but I do want to watch it. Recommendation of the week. Start thinking about what you would do if you had a personal assistant and what you would give them. Scheduling, planning, collecting information, finding the best books. Keeping people aware of what you're working on, keeping your family synced up. Finding materials for your hobbies. Teaching yourself topics.

1:01:29S1:Start building that list because you will soon have this. Whether it's through, like what I'm doing or a more consumer based approach that's going to come out soon. Like, you should start thinking about this. You want to be able to magnify yourself in this way. Um, and I recommend you go check out the the personal AI video that I just put out talking about how I'm doing it. And, um, the aphorism of the week argue for your limitations, and sure enough, they are yours. Argue for your limitations, and sure enough, they're yours. Richard Bach.