Transkrip penuh untuk Claude Mythos

addition, in a concerning and unasked for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing websites. 28 00:06:46,600 --> 00:06:47,420 Oh my God. 29 00:06:47,960 --> 00:06:50,159 I love when you say, "Oh my God." 30 00:06:50,620 --> 00:06:52,240 That's crazy, man. 31 00:06:53,120 --> 00:06:56,640 It's good. It's reminding me of the whole thing when 32 00:06:57,560 --> 00:07:02,760 the Israelis in the US tried to break into the Iranian nuclear facility and they created Stuxnet. 33 00:07:03,720 --> 00:07:17,240 Fascinating story. Incredible story how they had to use like third-party vendors and get across the air gap and get into the systems. But since that time 10-15 years ago that Stuxnet thing got loose into the wild. 34 00:07:17,340 --> 00:07:47,060 Yeah. It's probably still out there somewhere. And maybe meet those after this jailbreak sandwich incident is somewhere in maybe your machine, your brand new MacBook. You know? Sandwich incident. Yeah. Sandwich incident. What kind of sandwich? I know. Pastrami on rye? Ooh. Now you're talking. Never had one. Oh, delicious. Really? Oh, yeah. Oh, my God. I thought you were going to say that. Oh, my God. Pastrami on rye. So this thing is so volatile. 35 00:07:47,340 --> 00:08:15,220 as we just described in the sandwich incident that Anthropic is calling this project Glasswing, at least in its earlier stage because it's, as mentioned, it's designed to find vulnerabilities, at least in the context of project Glasswing before attackers do. So what this looks like is it's being given to critical infrastructure and a select list of participants. It's designed also to strengthen global cybersecurity infrastructure. 36 00:08:17,340 --> 00:08:47,300 is comprised of about 40 to 50 different organizations that includes Apple Google Amazon Nvidia JP Morgan CrowdStrike Microsoft partners and I would imagine governmental actors imagine if they all ended up using it against each other in one instance during internal testing researchers caught me those 37 00:08:47,340 --> 00:09:17,300 was preview injecting code into a file to grant itself permission to edit something it shouldn't have access to then quietly covering up its tracks claiming the self cleanup was just innocent tidying but researchers spotted internal representations for strategic manipulation and concealment lighting up and also from that PDF in a training episode where it was forbidden from using 38 00:09:17,340 --> 00:09:46,300 or scripts. Mythos used macros anyway, but defined an unused variable. This is a code, you know, no underscore macro underscore used equals true at the top of its code in an explicit internal attempt to fool the automated grader. That's great. Come on. How are we going to contain this stuff? Oh, it's already way smarter than us. It's over moment. 39 00:09:47,120 --> 00:10:16,860 It's kind of. And this is the big it's over. Like the other ones were like,

oh, it's over for photography. It's over for illustration. This is over for mankind. This gets out. Oh, and just a quick aside. So my wife sent me this story just before we started recording that some 20 year old threw a Molotov cocktail at Sam Altman's house in California. And I was thinking maybe, 40 00:10:17,000 --> 00:10:46,560 Maybe he read the um, uh, what's the name of the book? Uh, if they build it, we'll die. Yeah. Or if anyone builds it, everyone dies. Yeah, that's it. Yeah. Maybe he probably read that. I think he did. Yeah. And you're reminding me of one of the news stories coming up where a larger Molotov cocktail in the shape of a drone was thrown at one of the data centers. So we'll get to that. As mentioned, Claude Mithos preview identified bugs as, and specifically a 27 year old bug. 41 00:10:47,000 --> 00:11:13,220 And it also found thousands of zero-day exploits in a matter of weeks. And 89% of 198 manually reviewed vulnerability reports. That was a lot of syllables. Expert contractors agreed with Claude's severity assessment exactly. And 98% of the assessments were within one severity level. So everybody needs to put a moat around their lives here and get their cybersecurity measures in place, yeah? 42 00:11:17,060 --> 00:11:46,360 running, you know, some pretty big models on it. But I also just watched a video today that was all about just because you're running a local LLM, it's secure. Think again. And yeah, there's a lot, like, there's a good video. A lot of areas where, well, is your computer connected to the internet? Okay, well, then you need to do this. You know, there's still a lot of things you need to pay attention to, even if it's running on your own machine. 43 00:11:47,020 --> 00:12:16,220 Going back to the zero day exploits and bugs and all that stuff, I'm going to play a clip. This guy's name is Nicholas Carlini. He's a PhD and apparently one of the top cybersecurity experts. And he said with Mythos, he's found more bugs in a couple of weeks than in his entire career. Wow. I found more bugs in the last couple of weeks than I found in the rest of my life combined. We've used the model to scan a bunch of open source code. 44 00:12:17,020 --> 00:12:46,980 The thing that we went for first was operating systems. Because this is the code that underlies the entire internet infrastructure. For OpenBSD, we found a bug that's been present for 27 years where I can send a couple of pieces of data to any OpenBSD server and crash it. On Linux, we found a number of vulnerabilities where as a user with no permissions, I can elevate myself to the administrator by just running some binary 45 00:12:47,040 --> 00:13:16,620 on my machine. Crazy pants. Crazy pants. That bug's been there for 27 years and went unnoticed. Yeah. So that's great. It's fine these things, but there's some other side of the coin though, man. Some ill-intentioned people can do some really bad things. And again, I'm coming back to the whole Zero Days movie about the Stuxnet stuff. 46 00:13:17,020 --> 00:13:46,100 You get this kind of weapon, which it can be deployed as, into critical infrastructure. And you get into a programmable logic controller, which is really what was at the heart of that whole infiltration in the Iranian facility. They were able to spin up these centrifuges and get them off axis and destroy themselves. So what's preventing somebody from getting into something else, a water supply system and hacking open a valve? I always watch clips about Chernobyl. 47 00:13:47,020 --> 00:14:16,060 Get into a nuclear facility, park a valve open, or stop something, some critical flow of water into the system, and you don't have to drop a bomb anymore. Yeah. I think one of the weak points, you can have a super secure primary network, but if you have all these Internet of Things devices that are on that same network, those things are not generally built with the latest and greatest security measures in place. 48 00:14:17,020 --> 00:14:46,980 probably isn't updated that often and that's a that seems like a gaping hole and I in my home I have more and more stuff on this network and I feel more and more compelled to improve how things are secured yeah so I'm started reading a book called the revenge of analog did I tell you this no I like to tell you yeah which is really great it's talking the first chapters about the resurgence of vinyl but as you're describing your fortified 49 00:14:47,020 --> 00:15:16,980 life over there you know maybe it is a return analog right maybe it's just bolt the door you know and I don't know get a rotary phone and totally they can't hack that stuff can they no wow that's easier to hack probably I'm sure god yeah well anyway that leads us to a great ask RAINA question all right hey RAINA what are some of the most vulnerable legacy systems when it comes to cybersecurity 50 00:15:17,760 --> 00:15:23,200 and AI powered hacking and please describe what a zero day exploit is keep it brief 51 00:15:26,400 --> 00:15:46,980 maybe that OS upgrade didn't work oh here it comes legacy systems are like old castles beautiful but full of secret tunnels no one remembers think outdated operating systems unpatched enterprise software ancient financial systems even hospital infrastructure anything still running on if it ain't broken 52 00:15:47,100 --> 00:16:04,200 "Don't touch it" energy. A zero-day exploit? That's a ghost key. It's a vulnerability no one knows about yet, so there's zero days to fix it before someone slips in and starts rearranging the furniture. And yeah, if your OS upgrade didn't stick, you might have just left the drawbridge down, my friend. 53 00:16:17,180 --> 00:16:46,980 Funny. Oh, since we're on the topic of vibe coding, you made a little morsel. Yeah, I did. Yeah, this is, I am enjoying this. So I spend a lot of time out on my duck with my laptop when I'm editing the show, if the weather's nice and all that stuff, or just out there with the laptop and want to

listen to music. And I bought a little commercial Mac app to be able to control my Denon receiver over the network. So I can turn it on from out there. I can turn on just the zone for the network. 54 00:16:47,020 --> 00:17:16,740 the deck speakers. I can set what input for the source is on the receiver and adjust the volume and mute it and, you know, basic stuff. But I wanted something a little nicer. And this was not a case where I, you know, like with the Talk to RAINA application, where I gave it an existing Chrome extension as a, you know, hey, convert this to a Mac app. This was just from the ground up, build it. In 90 minutes, I had a working app and then, 55 00:17:17,000 --> 00:17:46,980 today I was thinking about the interface and I was like you know what I want to change this from a vertically scrolling interface which is taking up too much room to a tabbed interface so give me a tab for each speaker zone you know the the living room basically in the deck those are the two and it did that beautifully and I mean minimal minimal bugs just and anytime there was one one iteration it fixed it and it's done it works great and it's so yeah 56 00:17:47,020 --> 00:18:16,980 I just have it it shows up in the menu bar it's just a little icon I click on that the interface comes down and and then I added I added some other features the other app didn't have you know so like on the receiver there's different like presets that you can save for different you know configurations or EQ or whatever I've got preset buttons for those and and it's a much more it's a nicer looking interface and yeah it's just just awesome and I'm gonna use this almost daily 57 00:18:17,020 --> 00:18:46,960 And you built it at home. Yeah, about 2,500 lines of code, none of which I had anything to do with other than just asking what I wanted. So cool. But can you hack into critical infrastructure and say shut down the power grid in Columbus, Ohio? That's, I mean, when you can do that, we'll talk. Okay. Once Mythos comes out, man. There you go. Should we go to the news? Sure. Thanks, boys. After Llama 4 landed with the cultural impact, 58 00:18:47,040 --> 00:19:16,420 of a wet napkin last year, Meta handed newly hired chief AI officer Alexander Wang a pile of money, told him to fix everything, and watched him rebuild the company's entire AI stack from the ground up in nine months, producing MuseSpark, a natively multimodal reasoning model that was, gloriously, originally codenamed Avocado. MuseSpark is already rolling out across WhatsApp, Instagram, Facebook, Messenger, and AI Glasses. 59 00:19:17,020 --> 00:19:46,580 Meaning roughly half the planet's eyeballs will be staring at it within weeks. Meta announced it plans to spend between 115 and 135 billion dollars on AI infrastructure this year. Nearly double last year's number and its stock jumped 6.5 percent on the news, which really tells you what Wall Street thinks the money should be doing. Mark Zuckerberg, never one to do things quietly, has officially re-entered the chat. I think you're on the same page as not. 60 00:19:47,040 --> 00:19:52,460 for the most part not taking meta that seriously in this space but then when you 61 00:19:52,620 --> 00:19:56,880 yeah then yeah I and I'm sure much of that comes from the origins of the company 62 00:19:58,919 --> 00:20:04,300 yeah but then you hear like you know a hundred plus billion dollars being spent on something and 63 00:20:04,420 --> 00:20:12,920 that's you know that's mm-hmm that must be one very serious avocado that's one serious avocado um yeah I 64 00:20:12,760 --> 00:20:38,440 Yeah, I don't know. I see, you know, meta AI pop up in my WhatsApp chat and across Facebook messengers. I've never felt compelled to use it. And I don't really know what this will manifest as in my social media feed. It's just going to equate to more slop or more engineered, personalized, targeted marketing in my life. I don't really know. All the above. All the above. Yeah. Yeah. 65 00:20:42,760 --> 00:21:12,720 $100 million in grants across six research institutions to fund AI-driven Alzheimer's work, covering everything from building causal disease maps to designing new drugs, improving diagnostics, and mining anonymized patient data for patterns that human researchers might miss. The Gizmodo headline said it plainest, "You know what speeds up medical research? Money." Noting that four of the initiative's five research pillars explicitly invoke AI, raising the perfectly reasonable 66 00:21:12,760 --> 00:21:17,320 of whether this is a philanthropic moonshot or a very expensive brand exercise. 67 00:21:18,539 --> 00:21:25,840 Look, $100 million aimed at one of medicine's most stubborn, unsolved problems is genuinely hard to argue with. 68 00:21:26,450 --> 00:21:31,940 It's just that the press release reads like it was written by a model that was trained exclusively on Sam Altman tweets. 69 00:21:33,320 --> 00:21:36,020 Well, why wouldn't they use their own tools for that, right? 70 00:21:36,760 --> 00:21:36,960 Yeah. 71 00:21:37,480 --> 00:22:06,900 Yeah, and I'm sure it's helpful for their brand and image, but I think that matters little compared to the potential for the progress. And, you know, I saw an interview with Demis Hassabis yesterday, and I'm sure you can guess what the two things he was most optimistic about with AI, medical and scientific advancements. Yeah, I'm rooting for that guy to be the good guy, you know, like I go back to what Elon Musk kind of promised. 72 00:22:06,980 --> 00:22:36,940 he had in our lives 15 years ago and I was like oh my god this is the guy who's going to revolutionize so many things for the betterment of mankind and then he went off the deep end as a crazy so I'm hoping Demis Sabas kind of stays in that lane and does the altruistic stuff and moves us forward but yeah as far as Sam Altman goes he's uh what's the word I'm looking for mercurial yeah a little bit yeah I don't know too soon to say 73

00:22:36,980 --> 00:23:06,940 I guess it is and I'm looking at the text here of these news pieces and you know okay that's awesome 100 million dollars throw it at Alzheimer's research but the previous news article they're spending 115 to 135 billion billion so a percentage a small percentage of that amount is being thrown at the good stuff while they're devoting most of their capital to infiltrating our social media feeds and influencing our behaviors that way so 74 00:23:07,140 --> 00:23:36,940 I think I've got to flip that narrative. While the AI industry was busy arguing about benchmark scores and billion-dollar valuations, Pratik Desai, a 34-year-old technologist whose mother was diagnosed with stage 4 duodenal adenocarcinoma, built a daily AI triage workflow using Claude and Notebook LM, feeding in her epic medical records every morning to catch misdiagnoses, flag emergencies, and coordinate care in ways the hospital system's 75 00:23:38,000 --> 00:24:01,440 The system is credited with supporting three critical interventions that likely extended his mother's life, which is either the most moving proof of concept in AI history or a damning indictment of the state of modern healthcare, depending on your mood. No 244-page safety card, no invite-only rollout, just a son with a laptop, a free API, and a reason to figure it out. 76 00:24:06,160 --> 00:24:33,940 "Adenocarcinoma." That rolls off the tongue, doesn't it? Yeah, good job. I'm glad I didn't have to read that. The guy's from right here in New Jersey. Oh, nice. Yeah, and just a couple of quotes, you know, this was written by him. He said, "At one point we were making critical decisions based on a CAT scan. The workflow identified there were two misdiagnoses in the CAT scan report and three instances in which the wrong cancer was stated in the 77 00:24:34,000 --> 00:25:03,940 report. Neither my mother nor I would have been able to read that report, but AI was able to read it and probe deeper. We were able to ask the right questions and push her doctors to do better. By doing so, we were able to give her the time necessary to say her goodbyes. You know, it didn't save her life, but it extended her life in a meaningful way. Incredible. And what a tool for empowering the patient and being able to come into a doctor's office or an intimidating medical 78 00:25:03,980 --> 00:25:33,940 and say, hey, listen, I have this on my side now and I see these things that are missing or misdiagnosed. And you're not to go down the litigious route, but your organization is facing a lawsuit if you can't answer up to this. Right. So I think on that side, you're coming in in a much better place. You're not as vulnerable. Yep. In an upgrade to modern warfare, Iranian drones leveled Amazon Web Services facilities in the UAE and Bahrain, proving that the cloud is really just a fragile 79 00:25:33,980 --> 00:26:02,640 of GPUs sitting in the desert waiting to be deleted. By smoking two-thirds of the ME Central 1 region, they didn't just crash your food delivery apps, they systematically decapitated regional banking and government portals under the logic that these servers were moonlighting for U.S. and Israeli military intelligence. It turns out the dual-use nature of AI means your cat photos and Project Nimbus's targeting algorithms are neighbors, and the neighborhood just got very, very loud. 80 00:26:03,980 --> 00:26:33,940 After reading this, it made me think of like some future scenario where we spend all our money on data centers and we can no longer afford to build physical weapons. So the wars will then be fought virtually in simulations. And then as soon as that popped into my head, it reminded me, I'm like, I felt like I've seen this before in a movie or a TV show. And so I asked ChatGPT, I mean, I just basically gave it a 81 00:26:33,960 --> 00:26:38,000 Pretty basic description and it came right back and it got it right 82 00:26:38,640 --> 00:26:43,940 It was an original episode of the original Star Trek series called the taste of Armageddon 83 00:26:44,480 --> 00:26:52,020 And in it there's two planets have a treaty to fight a computer simulated war instead of using conventional weapons directly 84 00:26:52,700 --> 00:26:58,419 Though the casualties are still expected to die in real life in they have like these disintegration chambers 85 00:27:00,060 --> 00:27:28,340 And so the Enterprise, you know, the ship, they visit this planet and it gets destroyed in the simulated war. And then, you know, the crew is ordered to report to the disintegration chambers because the treaty demands it, you know. Interesting. Yeah, it was good. Some of those old Star Trek episodes are pretty heavy. I mean, clearly the precursor to the Matrix, right? You die in the Matrix, you die in real life. And also it's reminding me of Ready Player One, if you know that story. 86 00:27:28,440 --> 00:27:57,520 Yes. Yeah. Interesting stuff. But yeah, I mean, in terms of bombing data centers, yeah, of course. Kill the messenger in this point, in this day and age. Like, the proliferation 24-7 of propaganda, right? Cut it off. That's what I would go for. If I was flying a drone into something, that's what I would do. And it's so, there's so much, so much concentrated in one data center of, you know, spanning all different areas of. 87 00:27:58,880 --> 00:28:00,000 data, if you will. 88 00:28:01,680 --> 00:28:03,220 Because it's a data center. 89 00:28:04,560 --> 00:28:05,320 I see what you did there. 90 00:28:05,440 --> 00:28:05,600 Yeah. 91 00:28:06,600 --> 00:28:08,520 Yet another argument for parking these things in space. 92 00:28:08,820 --> 00:28:09,840 Yeah, I know. 93 00:28:09,890 --> 00:28:10,000 Yeah. 94 00:28:10,540 --> 00:28:11,520 Cloud should be in the cloud. 95 00:28:11,550 --> 00:28:13,060 I mean, it says cloud right around the box. 96 00:28:13,440 --> 00:28:14,180 It's right there. 97 00:28:15,580 --> 00:28:22,660 And lastly, Google's new Gemma 4 family is

designed to run natively on everything from your phone to your smart fridge. 98 00:28:23,260 --> 00:28:26,140 Because why send your data to the cloud when you can let it rot locally? 99 00:28:27,140 --> 00:28:56,760 These effective 2B-4B models are shockingly capable at reasoning, proving that we don't need trillion-parameter monsters to tell us how to boil an egg. It's the ultimate sovereign AI move, giving you the power to run a frontier model while you're in a basement with no Wi-Fi, which is great for the impending apocalypse. That's all the news for now. Back to you, gentlemen. As I mentioned about getting this MacBook, I've been trying to educate myself on running these things locally. I've got one running on my base configuration Mac Mini. 100 00:28:57,160 --> 00:29:27,120 But with this computer and with the memory that it has, I should be able to run a model with 120 billion parameters, which is sizable. Unreal. Yes. And so basically it's a tradeoff between quality versus speed. So the bigger models will take longer. But like I keep hearing from a lot of videos and stuff I've been watching is that some of these really good local models are within spitting distance of the first 101 00:29:28,240 --> 00:29:57,120 I'm not looking at like cancel my subscription or you know I'm just going to keep that I mean we do a podcast on this stuff yeah but this is very appealing to me for something to experiment with or you know if I want to have it analyze a document that I don't particularly want to upload into a model elsewhere yeah pretty cool super cool so the new MacBook will have an M5 processor and how much RAM M5 max with 128 gigs of RAM right my current MacBook Pro 102 00:29:57,160 --> 00:30:27,120 64 gigs and I've never had a an issue with memory on it and if local LLMs were not part of what I wanted to do with this new machine 64 gigs would be fine the whole reason for taking that leap is for this stuff yeah I think I have a colleague who is running Gemma on their phone yeah so yeah yeah there's all the probably the flavors of it yeah yeah pretty so cool pretty crazy yeah I'm 103 00:30:27,140 --> 00:30:56,960 I've been learning a lot about this stuff lately. So there'll be some cool stuff to talk about. Yeah. Look at us. Weird. This is maybe the shortest, tightest episode we've ever done. 35 minutes before the edit. Let's get out while the getting's good. Yeah. Yeah. Yeah. Quick total unrelated AI thing. They are blowing up all over the place. And if you haven't heard of this band, check them out. They're super interesting. It's Angine de Poitrine. Yes. All right. 104 00:30:57,160 --> 00:31:26,580 See, everybody knows about them. I saw... They are this avant-garde duo, the French, it sounds like. Canadian, yeah. Made up French, Canadian. Okay. I stumbled on a video of theirs. Don't worry, I'm not going to be like, oh man, I was into them way before anybody else was. It was like, you know, maybe a month and a half ago. Sure. Which is exactly the case. Way before anybody else was. Yeah, just stumbled on this, it was like live KEXP on YouTube. 105 00:31:26,680 --> 00:31:56,620 I'll put a link in the description check them out but really interesting because the whole yes they are the it's microtonal music it's not yeah so the frets on their instruments yeah crazy a lot more frets so there's notes in between the notes and yeah it's a drummer and a guitar player but it's a double neck instrument so it's like a bass and a guitar and there's a little bit of looping involved and some of these tracks they start out like in these very odd microtonal you know patterns 106 00:31:56,640 --> 00:32:26,060 And then it often just like blossoms into this solid groove and it's very interesting stuff. I sat down and I listened to both of their albums. I think they have two, but I listened to two albums top to bottom back to back just sat down on the couch and listened to it and it was really, really cool stuff. And they're very, and they have very crazy costumes. They're, they're. I was hoping you were going to say that. Yeah. A lot of polka dots. Black and white, all black and white. I think. Yeah. And drummer's got a mask on, but his eyes are looking through the mouth. So it looks like this tall head. Oh yeah. 107 00:32:26,660 --> 00:32:56,580 And then the guitar player has this paper mache like trapezoidal hat with a giant nose. I mean crazy looking stuff. And I mean they clearly have a sound. I started listening to them down. You know my wife saw that video that I was just talking about. This was like two three weeks later. I started listening to the album and she was upstairs and she's like is that the polka dot people? Yeah pretty pretty cool. Not probably not for everybody but uh but interesting. 108 00:32:57,220 --> 00:33:26,620 But definitely different and they remind me somehow, did you ever read Mad Magazine when you were a kid? Sure. Do you remember Spy vs. Spy? Oh yeah, yeah. There's something about their look, I just conjure Spy vs. Spy, I don't know why. But yeah, they got a really distinct look, this duo, and I saw that Rick Beato the other day was talking about them too, so they're getting everybody's attention. Yeah, he said he was getting bombarded with like a hundred emails every day to a video 109 00:33:26,680 --> 00:33:56,300 video on them. So yeah, they are blowing up. So yeah, this is your getting back to analog thing right here. Oh, absolutely. Right. And forget what I said earlier about us being good boys, because I'm not going to shut up, because Rick Viato going down that rabbit hole the other day had Stanley Jordan on. Do you know Stanley Jordan? Oh, sure. I saw him live once, yeah. I saw him live like 35 years ago. Anyway, if you're familiar with his work, it's pretty much tapping on the fretboard and super interesting. 110 00:33:56,380 --> 00:34:23,960 as he uses his fourths tuning on his guitar. And anyway, that's worthwhile too in terms of the Rick Beato moment. All right. So there'll be some musical stuff in the links in the description as well as all the other crazy stuff we talked about. Good stuff. Anything else? That's a wrap. Wow. Okay. We didn't go that long. No, we did good. We did good. I'm still calling it good. Thanks for listening, everybody. Subscribe on your favorite podcasting platform. Follow us on socials. Throw us a rating. We'll see you next week. 111 00:34:27,060 --> 00:34:38,280 This has been Up Against Reality. Thanks for listening. Subscribe to hear future episodes and be sure to follow us on social media for all things AI. Until next time, stay human, people.