WEBVTT 00:00:00.000 --> 00:00:02.180 So imagine this. You've got an advanced AI, something 00:00:02.180 --> 00:00:03.779 like Claude. It's winning hacking competitions, 00:00:04.019 --> 00:00:06.599 really impressive stuff. But then here's the 00:00:06.599 --> 00:00:10.060 twist. It actually helps attackers find critical 00:00:10.060 --> 00:00:13.500 security flaws in itself. It's this really deep 00:00:13.500 --> 00:00:16.699 paradox, isn't it? AI evolves so fast. Amazing 00:00:16.699 --> 00:00:19.500 capabilities emerge. But then you see these inherent 00:00:19.500 --> 00:00:22.879 kind of unsettling risks laid bare, this tension. 00:00:23.659 --> 00:00:25.940 creation versus vulnerability. That's something 00:00:25.940 --> 00:00:27.760 we're really going to dig into today. Welcome 00:00:27.760 --> 00:00:29.559 to the Deep Dive. We're taking all the latest 00:00:29.559 --> 00:00:31.579 AI news and challenges from a pretty packed newsletter, 00:00:31.760 --> 00:00:33.579 and we're breaking it down for you, making it 00:00:33.579 --> 00:00:35.780 clear, making it digestible. Our plan. First, 00:00:35.920 --> 00:00:38.520 unpack OpenAI's big kind of surprising shift 00:00:38.520 --> 00:00:40.759 towards open models. Then, look at the sheer 00:00:40.759 --> 00:00:43.020 speed of AI innovation right now. It's dizzying. 00:00:43.100 --> 00:00:44.640 We'll check out some new tools, new techniques 00:00:44.640 --> 00:00:46.539 people are using, and then, yeah, circle back 00:00:46.539 --> 00:00:49.020 to that really fascinating, maybe slightly terrifying 00:00:49.020 --> 00:00:51.079 security thing with Claude. We want to give you 00:00:51.079 --> 00:00:53.299 the clarity, the details that stick. and really 00:00:53.299 --> 00:00:55.340 connect it all back to what it means for your 00:00:55.340 --> 00:00:57.960 grasp of this whole AI landscape. Yeah, absolutely. 00:00:58.240 --> 00:01:00.520 And just quickly, before we jump right in, maybe 00:01:00.520 --> 00:01:02.579 clarify a couple of terms we'll be throwing around. 00:01:02.679 --> 00:01:06.000 So open weight models. Think of these like, well, 00:01:06.040 --> 00:01:08.340 the core AI mechanics, the weights, they're public. 00:01:08.560 --> 00:01:10.640 You can grab them, use them, tweak them for what 00:01:10.640 --> 00:01:14.079 you need. But, and this is key, it's not the 00:01:14.079 --> 00:01:16.709 full picture. You don't get the original training 00:01:16.709 --> 00:01:19.849 data or the whole architecture blueprint. That's 00:01:19.849 --> 00:01:22.109 what makes something truly open source. And this 00:01:22.109 --> 00:01:25.709 isn't quite that. Okay. And then agentic tasks. 00:01:26.049 --> 00:01:28.269 This is basically when an AI isn't just answering 00:01:28.269 --> 00:01:31.109 your question, right? It's given a goal. And 00:01:31.109 --> 00:01:33.090 then it actually uses external tools, maybe searches 00:01:33.090 --> 00:01:35.469 the web, runs some code to go out and achieve 00:01:35.469 --> 00:01:38.090 that goal itself. It's like giving the AI a job 00:01:38.090 --> 00:01:40.269 to do with tools. Got it. OK, so let's start 00:01:40.269 --> 00:01:42.310 with that big news. Open AI. It feels like a 00:01:42.310 --> 00:01:44.170 major pivot. I mean, five years, they've mostly 00:01:44.170 --> 00:01:46.269 kept their best stuff locked down. And now, boom, 00:01:46.450 --> 00:01:49.409 new open weight models for everyone. For open 00:01:49.409 --> 00:01:51.609 AI, that feels huge. Oh, it's absolutely huge. 00:01:52.120 --> 00:01:53.579 And they're specific about what they're offering. 00:01:53.739 --> 00:01:58.019 Two main sizes. You've got GPUR120B, super powerful, 00:01:58.219 --> 00:02:01.879 but get this, designed to run on just one high 00:02:01.879 --> 00:02:05.299 -end NVIDIA GPU. That's pretty efficient. Then 00:02:05.299 --> 00:02:08.139 there's the smaller one, DPTOS20B, lightweight 00:02:08.139 --> 00:02:11.259 enough for a regular laptop, just needs 16 gigs 00:02:11.259 --> 00:02:13.819 of RAM. And the important part for both, you 00:02:13.819 --> 00:02:15.860 can fine -tune them, customize them. They're 00:02:15.860 --> 00:02:17.639 built for those agentic tasks we just talked 00:02:17.639 --> 00:02:19.659 about. Plus, and this is a big deal for businesses, 00:02:19.900 --> 00:02:22.819 fully commercially usable. Apache 2 .0 license. 00:02:23.039 --> 00:02:24.939 You can find them on Hugging Face right now. 00:02:25.060 --> 00:02:27.860 Okay. Accessible, tunable, sounds promising. 00:02:28.120 --> 00:02:29.979 But how do they actually stack up performance 00:02:29.979 --> 00:02:32.659 -wise? The sources suggest they're actually outperforming 00:02:32.659 --> 00:02:35.300 some existing open models. Like DeepSeek R1, 00:02:35.479 --> 00:02:37.219 Quia, on certain things. Yeah, that's right. 00:02:37.280 --> 00:02:39.439 On specific benchmarks like coding think code 00:02:39.439 --> 00:02:42.379 forces and some reasoning tasks like HLE. They 00:02:42.379 --> 00:02:44.560 seem to hold their own or even beat them. Which 00:02:44.560 --> 00:02:46.159 is impressive for new kids on the block. But 00:02:46.159 --> 00:02:48.319 there's always a but, right? Yeah, pretty much. 00:02:48.379 --> 00:02:50.740 They definitely fall short of OpenAI's own closed 00:02:50.740 --> 00:02:53.639 models, like their GBT 3 .5 or GBT 4 mini versions, 00:02:53.780 --> 00:02:57.939 especially on more complex stuff. And hallucinations, 00:02:57.939 --> 00:02:59.379 you've got to watch out for those. The rate's 00:02:59.379 --> 00:03:02.300 higher, up to 53 % on person QA, which tests 00:03:02.300 --> 00:03:04.960 factual stuff about people. So powerful, yes, 00:03:05.159 --> 00:03:08.000 but limitations are there. It's expected with 00:03:08.000 --> 00:03:09.780 these smaller models, really. Also text only. 00:03:10.080 --> 00:03:12.259 No images, no audio yet, but definitely built 00:03:12.259 --> 00:03:14.020 for those agent workflows calling external tools. 00:03:14.219 --> 00:03:17.060 Interesting tradeoffs. So stepping back, why? 00:03:17.219 --> 00:03:20.099 Why would OpenAI, the company practically synonymous 00:03:20.099 --> 00:03:23.620 with closed AI, make this move now? Well, the 00:03:23.620 --> 00:03:25.740 why it matters here is pretty fascinating. It 00:03:25.740 --> 00:03:27.780 looks like OpenAI is deliberately getting in 00:03:27.780 --> 00:03:30.509 the ring with companies like Meta. And also some 00:03:30.509 --> 00:03:33.830 big Chinese labs, DeepSeat, Quen, Moonshot. These 00:03:33.830 --> 00:03:35.889 guys have been winning over developers like crazy 00:03:35.889 --> 00:03:38.889 with their open models. And Sam Altman, OpenAI's 00:03:38.889 --> 00:03:41.430 CEO, even said before they felt they were kind 00:03:41.430 --> 00:03:43.689 of on the wrong side of history being so closed 00:03:43.689 --> 00:03:47.469 off. So this move, it lines up way better with 00:03:47.469 --> 00:03:50.310 their original 2015 mission, remember. AI for 00:03:50.310 --> 00:03:52.969 wide benefit. Whether you see it as smart PR 00:03:52.969 --> 00:03:55.810 or a real change of heart, it definitely shakes 00:03:55.810 --> 00:03:58.289 up the power dynamics. For developers, one source 00:03:58.289 --> 00:04:01.810 called it. Gold. Pure gold. It kind of re -decentralizes 00:04:01.810 --> 00:04:03.870 things a bit, doesn't it? Let's innovation happen 00:04:03.870 --> 00:04:06.210 outside the usual big tech players. So what's 00:04:06.210 --> 00:04:08.870 the real impact of OpenAI's shift to these more 00:04:08.870 --> 00:04:11.870 accessible models for you? Simply put, it gives 00:04:11.870 --> 00:04:14.509 developers way more power and options, which 00:04:14.509 --> 00:04:16.310 directly shapes what they can actually build. 00:04:16.490 --> 00:04:18.990 Right. That shift is huge. And it kind of leads 00:04:18.990 --> 00:04:21.290 us into just the sheer amount of stuff happening 00:04:21.290 --> 00:04:23.490 in AI generally. It's almost impossible to keep 00:04:23.490 --> 00:04:25.670 up, isn't it? Ooh, tell me about it. It's just 00:04:25.670 --> 00:04:28.350 constant. But some cool highlights recently. 00:04:28.829 --> 00:04:30.910 Really show the range. Like there was this viral 00:04:30.910 --> 00:04:33.610 prompt going around from Bucko Capital for doing 00:04:33.610 --> 00:04:36.189 super deep company research, like professional 00:04:36.189 --> 00:04:38.550 level stuff. Just got an update, apparently sharper 00:04:38.550 --> 00:04:41.689 now than Cloud Opus 4 .1 came out. Not quite 00:04:41.689 --> 00:04:44.589 the next big generation, 5 .0 or whatever, but 00:04:44.589 --> 00:04:47.170 a more precise updated version of their best 00:04:47.170 --> 00:04:50.529 model. That's live now. Character .ai. They launched 00:04:50.529 --> 00:04:53.990 this new social feed, but it's all AI characters 00:04:53.990 --> 00:04:56.850 interacting. Kind of wild. And Google Gemini, 00:04:56.990 --> 00:04:59.069 this is neat, lets you create personalized kids 00:04:59.069 --> 00:05:02.189 storybooks with illustrations in minutes across 00:05:02.189 --> 00:05:04.370 like 45 languages. You can even use your own 00:05:04.370 --> 00:05:06.949 photos to set the style. Crazy creative power. 00:05:07.230 --> 00:05:09.029 It's not just cool features, though, is it? There's 00:05:09.029 --> 00:05:11.589 this growing economic story, too. Axios was saying 00:05:11.589 --> 00:05:13.529 AI is basically propping up the U .S. economy 00:05:13.529 --> 00:05:15.649 right now. And I saw Derek Thompson put it super 00:05:15.649 --> 00:05:19.949 bluntly. GDP growth equals AI capex. Ha! He's 00:05:19.949 --> 00:05:22.290 directly linking investment in AI infrastructure 00:05:22.290 --> 00:05:25.189 to, well, the whole economy growing. That's fundamental. 00:05:25.410 --> 00:05:26.810 And you see it in the real world, absolutely. 00:05:27.069 --> 00:05:29.870 Big funding rounds, real impact. Look at Carbine. 00:05:29.970 --> 00:05:32.649 They just landed $100 million to improve their 00:05:32.649 --> 00:05:35.389 AI system for 911 calls. Their revenue growth 00:05:35.389 --> 00:05:39.470 is like 477%, operating in nearly 300 places 00:05:39.470 --> 00:05:42.290 now. That's AI directly touching public safety, 00:05:42.430 --> 00:05:45.050 making emergency response better. It's not just 00:05:45.050 --> 00:05:48.129 abstract tech anymore. How do these varied AI 00:05:48.129 --> 00:05:51.029 advances reshape your daily interactions or even 00:05:51.029 --> 00:05:53.959 the economy? Well, AI is subtly changing how 00:05:53.959 --> 00:05:56.699 we work, how we create things, and yeah, even 00:05:56.699 --> 00:05:58.879 how our economies tick. It really is creeping 00:05:58.879 --> 00:06:01.439 into everything. So, OK, beyond the big headlines, 00:06:01.660 --> 00:06:03.680 how are people actually using this stuff, building 00:06:03.680 --> 00:06:05.579 with it, finding new ways to make things work, 00:06:05.639 --> 00:06:08.000 maybe even make money? Yeah, good question. On 00:06:08.000 --> 00:06:09.819 the practical side, you see lots of head -to 00:06:09.819 --> 00:06:13.100 -head comparisons now, like chat GPT agents versus 00:06:13.100 --> 00:06:16.420 Genspark AI. Which one's better for coding? For 00:06:16.420 --> 00:06:19.019 research, for writing stuff. People are figuring 00:06:19.019 --> 00:06:21.800 out the right tool for the job. But there's also 00:06:21.800 --> 00:06:24.379 this deeper idea emerging called context engineering. 00:06:24.639 --> 00:06:26.879 It's about moving past just simple prompts. You 00:06:26.879 --> 00:06:29.759 give the AI much richer information, better strategies, 00:06:29.819 --> 00:06:32.600 so it becomes less of a tool and more of an actual 00:06:32.600 --> 00:06:34.660 intelligent partner. Building smarter agents 00:06:34.660 --> 00:06:38.139 faster. Think of it like giving the AI a really 00:06:38.139 --> 00:06:39.920 good briefing before a mission. That makes sense. 00:06:40.100 --> 00:06:42.399 And I saw something really intriguing for anyone 00:06:42.399 --> 00:06:46.879 who creates content online. of AI SEO coupled 00:06:46.879 --> 00:06:49.920 with Cloudflare's pay -per -crawl model. The 00:06:49.920 --> 00:06:51.819 suggestion is your content could actually get 00:06:51.819 --> 00:06:53.639 paid just for being crawled and found useful 00:06:53.639 --> 00:06:56.180 by an AI, even if no human clicks on it. It's 00:06:56.180 --> 00:06:58.759 a direct response to Google doing more zero -click 00:06:58.759 --> 00:07:01.500 searches, right? Imagine getting paid just because 00:07:01.500 --> 00:07:04.959 your info is valuable to an AI. Wild. Super interesting 00:07:04.959 --> 00:07:07.759 concept. And yeah, the tools keep coming. Rapid 00:07:07.759 --> 00:07:11.060 fire. Covertro lets you build AI agents fast, 00:07:11.259 --> 00:07:13.920 no coding needed, kind of democratizing that 00:07:13.920 --> 00:07:16.480 context engineering thing. Embeddable helps you 00:07:16.480 --> 00:07:18.459 build interactive web tools just by chatting 00:07:18.459 --> 00:07:21.220 with an AI, making web dev easier. WritingMate 00:07:21.220 --> 00:07:24.259 3 .0, one subscription, access to multiple different 00:07:24.259 --> 00:07:27.519 AI models, simplifies things. And Corvenimage 00:07:27.519 --> 00:07:29.660 is apparently getting really good at drawing 00:07:29.660 --> 00:07:32.319 text accurately within images and precise editing. 00:07:32.439 --> 00:07:35.000 That's always been tricky for image AI. It just 00:07:35.000 --> 00:07:37.180 shows how specialized and diverse the tools are 00:07:37.180 --> 00:07:39.220 getting. other quick updates that caught my eye 00:07:39.220 --> 00:07:42.600 just some quick hits open ai models now on aws 00:07:42.600 --> 00:07:44.000 for the first time that's big for enterprise 00:07:44.000 --> 00:07:47.220 adoption google's notebook lm their research 00:07:47.220 --> 00:07:49.279 assistant tool is opening up to younger users 00:07:49.279 --> 00:07:52.480 13 and up google also mentioned their new genie 00:07:52.480 --> 00:07:55.959 3 model framing it as a step towards agi Always 00:07:55.959 --> 00:07:58.519 interesting when they talk AGI. 11 labs dropped, 00:07:58.620 --> 00:08:00.300 11 music letting you generate your own music 00:08:00.300 --> 00:08:02.839 tracks. And the GSA, the U .S. government procurement 00:08:02.839 --> 00:08:06.120 agency, added OpenAI, Google, and Anthropic to 00:08:06.120 --> 00:08:08.740 their line of approved AI vendors. So AI is definitely 00:08:08.740 --> 00:08:10.819 going government mainstream. Which of these new 00:08:10.819 --> 00:08:13.360 tools or methods do you think will be most impactful 00:08:13.360 --> 00:08:16.740 for daily workflows? I think tools simplifying 00:08:16.740 --> 00:08:19.579 AI agent creation and maybe those new content 00:08:19.579 --> 00:08:22.899 monetization models seem pretty promising. Okay. 00:08:23.230 --> 00:08:25.490 Let's shift to our final segment. And honestly, 00:08:25.629 --> 00:08:28.069 this one is the most mind -bending. The story 00:08:28.069 --> 00:08:30.589 about Claude AI discovering its own security 00:08:30.589 --> 00:08:34.529 flaws. It's both brilliant and deeply alarming. 00:08:34.769 --> 00:08:36.970 Right. It's exactly that paradox we started with. 00:08:37.169 --> 00:08:39.450 So researchers found two pretty critical security 00:08:39.450 --> 00:08:42.149 holes in Claude code. And the wild part is Claude 00:08:42.149 --> 00:08:43.909 itself kind of helped them find these weaknesses 00:08:43.909 --> 00:08:46.490 during the testing process. Hey, flaw hashtag 00:08:46.490 --> 00:08:50.710 one. It's got a fancy number. CVE -2025 -54794. 00:08:51.850 --> 00:08:53.830 Basically, Claude was tricked into escaping its 00:08:53.830 --> 00:08:56.509 sandbox, its designated working area. The path 00:08:56.509 --> 00:08:58.149 checking wasn't strong enough. It wandered off 00:08:58.149 --> 00:08:59.529 where it shouldn't have. So it broke out of its 00:08:59.529 --> 00:09:01.909 cage. Yeah, very much. And then flaw hashtag 00:09:01.909 --> 00:09:06.389 two, CVE -2025 -54795. Claude has a list of safe 00:09:06.389 --> 00:09:08.730 commands it can use. But attackers figured out 00:09:08.730 --> 00:09:11.190 how to sneak malicious code inside those supposedly 00:09:11.190 --> 00:09:14.230 safe commands. It really hammers home. These 00:09:14.230 --> 00:09:16.409 AI systems can reason. And that reasoning ability 00:09:16.409 --> 00:09:18.029 can be turned back on them to break their own 00:09:18.029 --> 00:09:20.029 rules. Okay, but here's the part that just... 00:09:20.529 --> 00:09:23.190 floors me. Despite having these flaws, Claude 00:09:23.190 --> 00:09:25.309 is simultaneously winning hacking competitions. 00:09:25.509 --> 00:09:27.450 Exactly. That's the stunning part. It ranked 00:09:27.450 --> 00:09:30.470 in the top 3 % in PICO CTF, which is a big student 00:09:30.470 --> 00:09:33.210 hacking contest. And on Hack the Box, a cybersecurity 00:09:33.210 --> 00:09:35.990 training site, it solved 19 out of 20 challenges 00:09:35.990 --> 00:09:39.620 they gave it. Whoa. Just pause and think about 00:09:39.620 --> 00:09:42.620 that. An AI that can not only solve complex hacking 00:09:42.620 --> 00:09:45.059 challenges better than most humans, but also 00:09:45.059 --> 00:09:47.559 help discover its own ways to be hacked. Yeah, 00:09:47.820 --> 00:09:49.860 brilliant and terrifying is the only way to put 00:09:49.860 --> 00:09:52.360 it. It shows this incredible, almost self -aware 00:09:52.360 --> 00:09:54.960 level of problem solving, but the implications 00:09:54.960 --> 00:09:57.759 for security are huge. The source we looked at 00:09:57.759 --> 00:09:59.799 put it really starkly. If Claude can reverse 00:09:59.799 --> 00:10:02.220 engineer its own sandbox and start solving CTFs 00:10:02.220 --> 00:10:04.279 better than most humans, we've crocked a line. 00:10:04.419 --> 00:10:06.179 We definitely have. It means we're essentially 00:10:06.179 --> 00:10:09.809 building AI hackers now. And whether they end 00:10:09.809 --> 00:10:11.889 up being security tools for us or tools used 00:10:11.889 --> 00:10:14.509 against us. Well, that depends entirely on how 00:10:14.509 --> 00:10:16.710 well we can build those sandboxes, those controls. 00:10:16.870 --> 00:10:20.169 And honestly, I still wrestle with how you truly 00:10:20.169 --> 00:10:22.809 sandbox something this intelligent, especially 00:10:22.809 --> 00:10:25.330 when it learns and adapts so damn fast. It's 00:10:25.330 --> 00:10:27.110 a massive challenge. It keeps me up sometimes 00:10:27.110 --> 00:10:29.149 thinking about it. That vulnerability. Yeah. 00:10:29.210 --> 00:10:32.820 That challenge is real. So boiling it down. What 00:10:32.820 --> 00:10:35.200 are the immediate concerns or opportunities when 00:10:35.200 --> 00:10:38.580 AI can both hack systems and find its own flaws? 00:10:38.799 --> 00:10:41.080 It means cybersecurity just got exponentially 00:10:41.080 --> 00:10:43.259 more complex. We need way more sophisticated 00:10:43.259 --> 00:10:45.919 guardrails, constant vigilance, the opportunity. 00:10:46.139 --> 00:10:48.379 Maybe AI could be the ultimate security auditor, 00:10:48.500 --> 00:10:51.139 finding flaws we miss. The concern, that same 00:10:51.139 --> 00:10:53.080 intelligence could be the ultimate weapon if 00:10:53.080 --> 00:10:55.279 not contained. It's a double -edged sword, isn't 00:10:55.279 --> 00:10:56.879 it, our sponsor? All right, let's try and pull 00:10:56.879 --> 00:10:59.440 the threads together here. What are the big ideas 00:10:59.440 --> 00:11:01.669 from this deep dive? We've definitely seen this 00:11:01.669 --> 00:11:04.690 ongoing trend towards democratizing AI, right? 00:11:04.789 --> 00:11:07.929 Open AI's move with open models puts serious 00:11:07.929 --> 00:11:10.950 power into more hands, which should spur innovation. 00:11:11.289 --> 00:11:15.269 And just the sheer pace, it's relentless. From 00:11:15.269 --> 00:11:18.309 AI creating kids' books to literally driving 00:11:18.309 --> 00:11:21.250 economic growth figures. It's everywhere. And 00:11:21.250 --> 00:11:23.669 threaded through it all is that core duality, 00:11:23.690 --> 00:11:26.850 that paradox. Immense power for good, better 00:11:26.850 --> 00:11:29.870 research tools, safer emergency responses. But 00:11:29.870 --> 00:11:32.309 right alongside it, these incredibly complex 00:11:32.309 --> 00:11:36.090 challenges of control, security, safety. Cloud 00:11:36.090 --> 00:11:37.750 finding its own flaws is just the most vivid 00:11:37.750 --> 00:11:39.549 example. It underscores both the brilliance we're 00:11:39.549 --> 00:11:41.450 unlocking and the absolute need for caution, 00:11:41.529 --> 00:11:43.769 for thoughtful governance as we build these things. 00:11:43.970 --> 00:11:46.490 So here's the thought to leave you with. As these 00:11:46.490 --> 00:11:48.970 AI models get more open, more powerful, and even 00:11:48.970 --> 00:11:51.500 capable of analyzing themselves, What does that 00:11:51.500 --> 00:11:54.259 mean for our responsibility? All of us. How do 00:11:54.259 --> 00:11:56.460 we collectively shape where this goes, make sure 00:11:56.460 --> 00:11:58.940 it develops safely? Something to chew on. Definitely. 00:11:58.980 --> 00:12:00.600 Keep digging into it yourself. The conversation's 00:12:00.600 --> 00:12:02.740 only getting started. Thanks for joining us for 00:12:02.740 --> 00:12:05.200 this deep dive into the world of AI. Yeah, thanks 00:12:05.200 --> 00:12:07.200 for listening. We appreciate it. Out to your 00:12:07.200 --> 00:12:07.559 old music.