1 00:00:00,000 --> 00:00:01,420 - Okay, let's dive right in. 2 00:00:01,420 --> 00:00:02,780 Welcome to Crash News. 3 00:00:02,780 --> 00:00:05,840 Today is May 22nd, 2025. 4 00:00:05,840 --> 00:00:07,980 You just imagine this for a second. 5 00:00:07,980 --> 00:00:11,300 A pocket-sized device, totally screen-free. 6 00:00:11,300 --> 00:00:12,140 - Right. 7 00:00:12,140 --> 00:00:14,580 - Designed by the person, Joni Ive, 8 00:00:14,580 --> 00:00:17,980 who basically helped shape the look and feel 9 00:00:17,980 --> 00:00:19,940 of the iPhone, the MacBook, 10 00:00:19,940 --> 00:00:22,980 aiming to completely change how we use AI. 11 00:00:22,980 --> 00:00:24,660 - That's a pretty wild concept. 12 00:00:24,660 --> 00:00:28,660 - Or how about AI that digs through scientific papers, 13 00:00:28,660 --> 00:00:30,660 lab data all on its own, 14 00:00:30,660 --> 00:00:32,300 and points researchers toward 15 00:00:32,300 --> 00:00:34,600 promising new medical treatments, 16 00:00:34,600 --> 00:00:36,980 like speeding up drug discovery massively. 17 00:00:36,980 --> 00:00:38,180 - Yeah, that's actually happening. 18 00:00:38,180 --> 00:00:40,060 - And then there's the tech job market. 19 00:00:40,060 --> 00:00:41,620 It's a bit stark right now. 20 00:00:41,620 --> 00:00:44,120 Hiring for new grads, apparently down over 50% 21 00:00:44,120 --> 00:00:44,960 since before the pandemic. 22 00:00:44,960 --> 00:00:45,800 - A huge drop. 23 00:00:45,800 --> 00:00:48,820 - But experienced AI engineers, they're like gold dust. 24 00:00:48,820 --> 00:00:50,580 One company is apparently poaching talent 25 00:00:50,580 --> 00:00:52,220 from the big players at an incredible rate. 26 00:00:52,220 --> 00:00:53,740 - The competition is fierce. 27 00:00:53,740 --> 00:00:55,320 - So welcome to today's deep dive. 28 00:00:55,320 --> 00:00:57,540 We've got a stack of sources, industry reports, 29 00:00:57,540 --> 00:01:00,380 company news, research updates, newsletters, the works. 30 00:01:00,380 --> 00:01:01,540 - All focused on AI. 31 00:01:01,540 --> 00:01:03,180 - Exactly, and our mission really 32 00:01:03,180 --> 00:01:04,660 is to cut through all that noise 33 00:01:04,660 --> 00:01:05,980 and give you the essential briefing 34 00:01:05,980 --> 00:01:09,140 on what's actually moving the needle in AI, like right now. 35 00:01:09,140 --> 00:01:11,860 - Think of it as your shortcut to staying current, 36 00:01:11,860 --> 00:01:13,660 because things are moving so fast 37 00:01:13,660 --> 00:01:15,100 across so many different fronts. 38 00:01:15,100 --> 00:01:16,220 - They really are. 39 00:01:16,220 --> 00:01:18,480 It's hardware, it's software models, 40 00:01:18,480 --> 00:01:21,100 it's how AI is showing up in the real world, 41 00:01:21,100 --> 00:01:23,440 even who's getting hired and where. 42 00:01:23,440 --> 00:01:24,280 - It's a lot to track. 43 00:01:24,280 --> 00:01:25,520 - Absolutely. 44 00:01:25,520 --> 00:01:27,340 So speaking of those big moves, 45 00:01:27,340 --> 00:01:30,820 let's start with maybe the biggest headline grabber. 46 00:01:30,820 --> 00:01:33,860 OpenAI making a massive bet on hardware, 47 00:01:33,860 --> 00:01:35,380 and they're doing it with Joni Ive. 48 00:01:35,380 --> 00:01:36,820 - Yeah, this is huge news. 49 00:01:36,820 --> 00:01:39,020 OpenAI has fully acquired IO, 50 00:01:39,020 --> 00:01:41,580 that was the company born from their joint venture. 51 00:01:41,580 --> 00:01:42,780 - Acquired them, okay. 52 00:01:42,780 --> 00:01:45,580 - For a reported $6.5 billion in equity. 53 00:01:45,580 --> 00:01:48,540 One source said 6.4 cents, another $6.5 billion, 54 00:01:48,540 --> 00:01:49,380 but either way. 55 00:01:49,380 --> 00:01:51,740 - That's a serious number for hardware. 56 00:01:51,740 --> 00:01:54,180 - It absolutely signals they are dead serious 57 00:01:54,180 --> 00:01:56,140 about moving beyond just software. 58 00:01:56,140 --> 00:01:57,340 This isn't a side project. 59 00:01:57,340 --> 00:01:58,540 - And bringing in Joni Ive, 60 00:01:58,540 --> 00:02:01,660 I mean the design legend from Apple, close to Steve Jobs, 61 00:02:01,660 --> 00:02:03,140 that sends a message, doesn't it? 62 00:02:03,140 --> 00:02:05,860 They're aiming for something big, something mainstream. 63 00:02:05,860 --> 00:02:06,700 - Definitely. 64 00:02:06,700 --> 00:02:09,060 And Ive, through his firm love from, 65 00:02:09,060 --> 00:02:12,460 he's going to oversee design for OpenAI's hardware 66 00:02:12,460 --> 00:02:13,460 and software now. 67 00:02:13,460 --> 00:02:14,700 - So he's deeply involved. 68 00:02:14,700 --> 00:02:15,900 - Fully integrated. 69 00:02:15,900 --> 00:02:19,460 And IO, the company itself, was only founded in 2024 70 00:02:19,460 --> 00:02:21,100 with other Apple veterans. 71 00:02:21,100 --> 00:02:24,060 About 55 of their engineers are joining OpenAI. 72 00:02:24,060 --> 00:02:24,900 - Wow. 73 00:02:24,900 --> 00:02:28,260 - Okay, so it's a whole team, a whole vision being absorbed. 74 00:02:28,260 --> 00:02:29,860 - Exactly, it's a full integration 75 00:02:29,860 --> 00:02:33,500 of top-tier hardware design into a leading AI company. 76 00:02:33,500 --> 00:02:35,640 - So what are they actually trying to build, 77 00:02:35,640 --> 00:02:37,340 this screen-free device? 78 00:02:37,340 --> 00:02:38,340 What's the vision? 79 00:02:38,340 --> 00:02:41,460 - Well, they're talking about a pocket-sized AI companion. 80 00:02:41,460 --> 00:02:43,440 Screen-free is the key part. 81 00:02:43,440 --> 00:02:44,280 - Right. 82 00:02:44,280 --> 00:02:46,480 - And the ambition is kind of mind-blowing. 83 00:02:46,480 --> 00:02:48,220 They're talking about potentially shipping 84 00:02:48,220 --> 00:02:49,660 a hundred million units. 85 00:02:49,660 --> 00:02:50,500 - A hundred million. 86 00:02:50,500 --> 00:02:53,340 - And they're positioning it as maybe a third core device, 87 00:02:53,340 --> 00:02:55,140 like alongside the iPhone and the MacBook, 88 00:02:55,140 --> 00:02:57,100 or maybe even replacing some of their functions. 89 00:02:57,100 --> 00:02:59,940 - A third core device, screen-free. 90 00:02:59,940 --> 00:03:01,140 That's a fundamental shift 91 00:03:01,140 --> 00:03:02,640 in how we think about personal tech. 92 00:03:02,640 --> 00:03:03,480 - It is. 93 00:03:03,480 --> 00:03:05,840 And the idea, they say, is to tackle the quote, 94 00:03:05,840 --> 00:03:08,220 "Unintended consequences of smartphones." 95 00:03:08,220 --> 00:03:11,020 - Like distraction maybe, endless scrolling. 96 00:03:11,020 --> 00:03:12,740 - That seems to be the implication. 97 00:03:12,740 --> 00:03:15,540 They want a more natural, more intuitive way 98 00:03:15,540 --> 00:03:16,940 to interact with AI, 99 00:03:16,940 --> 00:03:18,660 calling it a totally new kind of thing. 100 00:03:18,660 --> 00:03:20,900 Sam Altman and Joni Ive apparently even called it 101 00:03:20,900 --> 00:03:22,940 potentially the coolest piece of technology. 102 00:03:22,940 --> 00:03:23,940 High praise from them. 103 00:03:23,940 --> 00:03:26,660 Okay, and they're aiming for a 2026 launch. 104 00:03:26,660 --> 00:03:30,100 That feels incredibly fast for something so different. 105 00:03:30,100 --> 00:03:31,940 - It's super aggressive, no question. 106 00:03:31,940 --> 00:03:35,540 But the key takeaway here for you listening is this. 107 00:03:35,540 --> 00:03:38,180 When a major AI player like OpenAI 108 00:03:38,180 --> 00:03:40,540 teams up with someone like Joni Ive 109 00:03:40,540 --> 00:03:44,720 and bets billions on dedicated screen-free hardware, 110 00:03:44,720 --> 00:03:46,820 it signals a serious potential shift 111 00:03:46,820 --> 00:03:48,780 in how we interact with AI day to day. 112 00:03:48,780 --> 00:03:50,740 It's moving AI beyond the screen, 113 00:03:50,740 --> 00:03:53,580 maybe into something more ambient, more integrated. 114 00:03:53,580 --> 00:03:54,760 - Which could change our habits, 115 00:03:54,760 --> 00:03:56,720 our whole relationship with tech. 116 00:03:56,720 --> 00:03:58,300 Okay, huge move. 117 00:03:58,300 --> 00:04:00,640 But let's shift gears because it's not just about hardware. 118 00:04:00,640 --> 00:04:03,360 Our sources are also showing just a flood of new AI models, 119 00:04:03,360 --> 00:04:05,660 new tools, new capabilities hitting the market. 120 00:04:05,660 --> 00:04:06,700 It's almost hard to keep up. 121 00:04:06,700 --> 00:04:07,540 - It really is. 122 00:04:07,540 --> 00:04:08,380 It's incredibly dynamic. 123 00:04:08,380 --> 00:04:10,480 And what's interesting is how specialized 124 00:04:10,480 --> 00:04:11,940 these models are becoming. 125 00:04:11,940 --> 00:04:12,940 - Specialized how? 126 00:04:12,940 --> 00:04:14,540 What are some examples from the sources? 127 00:04:14,540 --> 00:04:17,260 - Well, one big area is coding agents. 128 00:04:17,260 --> 00:04:19,500 Mistral AI, working with All Hands AI, 129 00:04:19,500 --> 00:04:21,540 just released something called Devstral. 130 00:04:21,540 --> 00:04:22,380 - Devstral, okay. 131 00:04:22,380 --> 00:04:23,780 - It's an open source model 132 00:04:23,780 --> 00:04:27,020 trained specifically for software engineering tasks. 133 00:04:27,020 --> 00:04:29,980 It's relatively small, 24 billion parameters, 134 00:04:29,980 --> 00:04:31,900 so maybe usable on a laptop. 135 00:04:31,900 --> 00:04:34,420 And the key thing is it can understand context 136 00:04:34,420 --> 00:04:36,780 across different files in a project. 137 00:04:36,780 --> 00:04:39,100 - Okay, you used the term agentic model there. 138 00:04:39,100 --> 00:04:40,980 We hear agent a lot. 139 00:04:40,980 --> 00:04:42,580 What does that actually mean for coding? 140 00:04:42,580 --> 00:04:44,860 It's more than just suggesting the next line of code, right? 141 00:04:44,860 --> 00:04:47,100 - Exactly, that's the crucial difference. 142 00:04:47,100 --> 00:04:49,060 A regular code completion tool. 143 00:04:49,060 --> 00:04:52,040 you suggestions for the immediate line or function. 144 00:04:52,040 --> 00:04:53,400 Right, like autocomplete. 145 00:04:53,400 --> 00:04:57,640 Kind of. But an agentic model, like DevStraw is aiming to be, 146 00:04:57,640 --> 00:05:03,240 is designed to act more autonomously. It tries to understand the bigger picture of your code, 147 00:05:03,240 --> 00:05:08,280 navigate files, grasp the context, and potentially handle more complex tasks on its own. 148 00:05:08,280 --> 00:05:09,640 Like what kind of tasks? 149 00:05:09,640 --> 00:05:14,840 Things like finding bugs that span multiple parts of the code may be suggesting larger refactors, 150 00:05:14,840 --> 00:05:19,240 or even generating whole chunks of code across several files based on a description. 151 00:05:19,240 --> 00:05:22,840 DevStraw is showing good benchmark results, apparently, 152 00:05:22,840 --> 00:05:27,400 and being open source is a big deal for developers, even though it's still a research preview. 153 00:05:27,400 --> 00:05:29,880 So like having an AI assistant engineer. 154 00:05:29,880 --> 00:05:34,360 Interesting. What about the creative side? The demos there are often pretty stunning. 155 00:05:34,360 --> 00:05:39,480 Oh, absolutely. ByteDance has Bagel, a multimodal model for image generation, 156 00:05:39,480 --> 00:05:42,440 putting it in the same league as things like GPT 4.0. 157 00:05:42,440 --> 00:05:45,240 Okay. But Google's new video generator, 158 00:05:45,240 --> 00:05:51,160 VO3, sounds really impressive. It can apparently create videos complete with sound effects and 159 00:05:51,160 --> 00:05:55,000 even dialogue just from a text prompt. Yeah, I saw that example in the sources, 160 00:05:55,000 --> 00:05:57,160 the pharma commercial creator. That really stood out. 161 00:05:57,160 --> 00:06:01,000 Didn't it? This person apparently used to spend like half a million dollars on a commercial shoe. 162 00:06:01,000 --> 00:06:02,360 $500,000. Yeah. 163 00:06:02,360 --> 00:06:08,600 And they were able to generate one using VO3 in less than a day for about $500 in credits. 164 00:06:08,600 --> 00:06:14,840 That's disruptive potential right there, reducing cost and time drastically in creative work. 165 00:06:14,840 --> 00:06:19,320 It's potentially revolutionary for those fields. And there are other tools popping up too. 166 00:06:19,320 --> 00:06:23,000 Google Flow for filmmaking, Luma Labs for camera motion effects, 167 00:06:23,000 --> 00:06:28,760 Geometry Crafter, Triplane Turbo. Lots happening in creative AI generation. 168 00:06:28,760 --> 00:06:32,040 And for the developers who want to build on top of these models, 169 00:06:32,040 --> 00:06:36,040 what's happening with API Access? Access is definitely improving. 170 00:06:36,040 --> 00:06:41,880 Anthropic, for instance, announced their upcoming Claude 4 Sonnet and Opus models will be available. 171 00:06:41,880 --> 00:06:44,040 Okay. But Lambda's offering caught my eye. 172 00:06:44,040 --> 00:06:48,600 They've launched an inference API with what they call unrestricted access to models like DeepSeek 173 00:06:48,600 --> 00:06:53,160 V3, Llama 4 Maverick, even the big DeepSeek R1 671B. 174 00:06:53,160 --> 00:06:54,360 Unrestricted. Yeah. 175 00:06:54,360 --> 00:06:59,400 Their pitch is no rate limits, no tiers, no throttling, plus competitive pricing. 176 00:06:59,400 --> 00:07:04,440 For developers needing reliable, scalable AI access without hitting surprise limits, 177 00:07:04,440 --> 00:07:06,440 that could be really appealing. Got it. 178 00:07:06,440 --> 00:07:09,640 So the models are getting better, more specialized, easier to access. 179 00:07:09,640 --> 00:07:13,320 But how do people actually get good results? The source. 180 00:07:13,500 --> 00:07:15,060 to seem to emphasize prompting a lot. 181 00:07:15,060 --> 00:07:16,860 - Yeah, prompt optimization basically, 182 00:07:16,860 --> 00:07:19,720 how you ask the AI for things is becoming super important. 183 00:07:19,720 --> 00:07:20,780 It's a real skill. 184 00:07:20,780 --> 00:07:24,600 The sources mentioned a tutorial covering five prompt types. 185 00:07:24,600 --> 00:07:28,880 Things like summarize, reframe for a specific audience, 186 00:07:28,880 --> 00:07:32,460 brainstorm ideas, compare things, and role play scenarios. 187 00:07:32,460 --> 00:07:33,900 - Why is the way you ask so critical? 188 00:07:33,900 --> 00:07:35,900 Can't you just type what you want? 189 00:07:35,900 --> 00:07:37,980 - You can, but the quality you get back 190 00:07:37,980 --> 00:07:39,820 really depends on how clear 191 00:07:39,820 --> 00:07:41,880 and well-structured your prompt is. 192 00:07:41,880 --> 00:07:44,700 It's like giving instructions to a person, 193 00:07:44,700 --> 00:07:47,540 vague instructions lead to vague results. 194 00:07:47,540 --> 00:07:49,880 Good prompting lets you specify the format, 195 00:07:49,880 --> 00:07:51,780 the tone, any constraints, 196 00:07:51,780 --> 00:07:54,980 even the personality you want the AI to adopt. 197 00:07:54,980 --> 00:07:56,680 The sources basically say product teams 198 00:07:56,680 --> 00:07:58,060 really need to focus on this 199 00:07:58,060 --> 00:08:01,460 because good prompting makes the AI vastly more useful. 200 00:08:01,460 --> 00:08:03,740 It's like learning a new way to communicate effectively. 201 00:08:03,740 --> 00:08:05,620 - A new kind of literacy, makes sense. 202 00:08:05,620 --> 00:08:07,360 Were there other little tools mentioned 203 00:08:07,360 --> 00:08:08,740 that help people work with AI? 204 00:08:08,740 --> 00:08:10,980 - Yeah, a few interesting ones briefly mentioned. 205 00:08:10,980 --> 00:08:14,140 Tools like snap deck AI to make Figma slides from text, 206 00:08:14,140 --> 00:08:15,620 something called unGPT 207 00:08:15,620 --> 00:08:17,820 to make AI writing sound less robotic. 208 00:08:17,820 --> 00:08:19,140 - Oh, that could be useful. 209 00:08:19,140 --> 00:08:22,060 - Right, also idocs.ai-redact 210 00:08:22,060 --> 00:08:25,100 for automatically blacking out sensitive info in documents, 211 00:08:25,100 --> 00:08:28,780 respond.io for managing different communication channels, 212 00:08:28,780 --> 00:08:30,140 studypal for learning, 213 00:08:30,140 --> 00:08:33,420 and liner for finding citations for AI claims. 214 00:08:33,420 --> 00:08:36,660 It shows a whole ecosystem building up around AI workflows. 215 00:08:36,660 --> 00:08:39,000 - Okay, so tons of models, tons of tools, 216 00:08:39,000 --> 00:08:42,000 but how do we even know which models are actually good? 217 00:08:42,000 --> 00:08:43,000 How do we compare them? 218 00:08:43,000 --> 00:08:45,880 - Ah, yes, benchmarking and evaluation. 219 00:08:45,880 --> 00:08:49,520 That's a huge and actually quite tricky area right now. 220 00:08:49,520 --> 00:08:50,360 - Tricky how? 221 00:08:50,360 --> 00:08:52,360 - Well, the sources mentioned LM Arena 222 00:08:52,360 --> 00:08:53,900 raising a hundred million dollars, 223 00:08:53,900 --> 00:08:55,280 which shows there's big money 224 00:08:55,280 --> 00:08:57,580 going into figuring out evaluation. 225 00:08:57,580 --> 00:08:58,820 But there are ongoing debates 226 00:08:58,820 --> 00:09:00,540 about whether the leaderboards are fair 227 00:09:00,540 --> 00:09:02,780 or if they truly reflect how well these models work 228 00:09:02,780 --> 00:09:03,620 in the real world. 229 00:09:03,620 --> 00:09:04,780 - Right, like teaching to the test. 230 00:09:04,780 --> 00:09:07,200 - Exactly, and models are evolving so fast, 231 00:09:07,200 --> 00:09:10,080 gaining new abilities like understanding images and video, 232 00:09:10,080 --> 00:09:12,300 that the tests themselves struggle to keep up. 233 00:09:12,300 --> 00:09:13,540 It's a moving target. 234 00:09:13,540 --> 00:09:16,380 - So even how we measure progress 235 00:09:16,380 --> 00:09:18,540 is still kind of being figured out. 236 00:09:18,540 --> 00:09:20,340 Okay, let's shift focus a bit. 237 00:09:20,340 --> 00:09:23,260 From the tech to the people, the talent building AI 238 00:09:23,260 --> 00:09:25,520 and the trust needed to use it. 239 00:09:25,520 --> 00:09:27,520 The sources had some really interesting, 240 00:09:27,520 --> 00:09:30,900 maybe sobering stats on the job market. 241 00:09:30,900 --> 00:09:33,740 - Yeah, the SignalFire report painted a stark picture, 242 00:09:33,740 --> 00:09:35,940 new grad hiring in tech generally. 243 00:09:35,940 --> 00:09:39,180 down significantly over 50%. 244 00:09:39,180 --> 00:09:39,520 from pre-pandemic levels. 245 00:09:39,520 --> 00:09:41,800 - Wow, that's tough for people starting out. 246 00:09:41,800 --> 00:09:43,480 - It is, but at the same time, 247 00:09:43,480 --> 00:09:45,540 the demand for experienced engineers, 248 00:09:45,540 --> 00:09:49,280 especially those with real AI skills, is incredibly high. 249 00:09:49,280 --> 00:09:50,600 They're in huge demand. 250 00:09:50,600 --> 00:09:52,560 - So it's a tale of two markets almost. 251 00:09:52,560 --> 00:09:54,520 Is there a specific talent war heating up 252 00:09:54,520 --> 00:09:56,520 between the big AI labs? 253 00:09:56,520 --> 00:09:57,560 - Seems like it. 254 00:09:57,560 --> 00:10:00,280 The report highlighted that companies really focused on AI, 255 00:10:00,280 --> 00:10:02,160 like Anthropic, are doing remarkably well 256 00:10:02,160 --> 00:10:04,280 at keeping their employees, high loyalty. 257 00:10:04,280 --> 00:10:05,120 - Okay. 258 00:10:05,120 --> 00:10:07,080 - And here's the really eye-popping stat. 259 00:10:07,080 --> 00:10:09,680 Apparently, talent is eight times more likely 260 00:10:09,680 --> 00:10:11,840 to move to Anthropic from places like open AI 261 00:10:11,840 --> 00:10:13,560 or DeepMind than the other way around. 262 00:10:13,560 --> 00:10:14,520 - Eight times? 263 00:10:14,520 --> 00:10:16,000 - Yeah, it's-- - That's a massive differential. 264 00:10:16,000 --> 00:10:17,320 What does that suggest? 265 00:10:17,320 --> 00:10:18,600 Something Anthropic is doing right 266 00:10:18,600 --> 00:10:20,080 or maybe something the others aren't? 267 00:10:20,080 --> 00:10:22,320 - It certainly suggests Anthropic has built 268 00:10:22,320 --> 00:10:25,520 a very attractive environment for top AI talent right now. 269 00:10:25,520 --> 00:10:28,120 For you, our listener, perhaps someone in tech 270 00:10:28,120 --> 00:10:30,400 or thinking about it, what does this mean? 271 00:10:30,400 --> 00:10:31,440 - Yeah, what's the takeaway? 272 00:10:31,440 --> 00:10:33,440 - It really highlights the premium being placed 273 00:10:33,440 --> 00:10:36,580 on deep, practical AI experience. 274 00:10:36,580 --> 00:10:38,720 While getting that first general tech job 275 00:10:38,720 --> 00:10:40,720 might be harder for new grads now, 276 00:10:40,720 --> 00:10:44,040 specializing in AI development, research or application 277 00:10:44,040 --> 00:10:46,040 could put you in a very strong position. 278 00:10:46,040 --> 00:10:49,900 And for the industry, this talent flow really shapes 279 00:10:49,900 --> 00:10:52,640 who's likely to lead the next wave of innovation. 280 00:10:52,640 --> 00:10:53,480 - Yeah, makes sense. 281 00:10:53,480 --> 00:10:56,400 Talent follows the most interesting work, often. 282 00:10:56,400 --> 00:10:58,400 Okay, beyond attracting talent, 283 00:10:58,400 --> 00:11:00,880 there's the challenge of building trust in AI. 284 00:11:00,880 --> 00:11:02,320 That includes security, right? 285 00:11:02,320 --> 00:11:03,520 Prompt injection came up. 286 00:11:03,520 --> 00:11:04,340 - It did. 287 00:11:04,340 --> 00:11:06,440 Prompt injection is a major headache. 288 00:11:06,440 --> 00:11:08,800 It's basically when someone tricks the AI 289 00:11:08,800 --> 00:11:10,700 by hiding malicious instructions 290 00:11:10,700 --> 00:11:12,560 inside a normal-looking prompt, 291 00:11:12,560 --> 00:11:14,680 trying to get it to bypass its safety rules 292 00:11:14,680 --> 00:11:15,940 or reveal things it shouldn't. 293 00:11:15,940 --> 00:11:17,000 - Right, that have vulnerability. 294 00:11:17,000 --> 00:11:17,840 - Exactly. 295 00:11:17,840 --> 00:11:19,640 Google DeepMind has been working hard 296 00:11:19,640 --> 00:11:21,920 on making their models more resistant to this, 297 00:11:21,920 --> 00:11:25,760 using techniques like adversarial training for Gemini 2.5, 298 00:11:25,760 --> 00:11:27,980 basically training the AI on tricky prompts 299 00:11:27,980 --> 00:11:28,900 to make it tougher. 300 00:11:28,900 --> 00:11:29,820 - And is it working? 301 00:11:29,820 --> 00:11:31,440 Are the defenses getting better? 302 00:11:31,440 --> 00:11:34,800 - According to their research, yes, significantly better. 303 00:11:34,800 --> 00:11:37,280 They reported that combining different defense techniques 304 00:11:37,280 --> 00:11:40,320 dropped the success rate of these attacks on Gemini 2.5 305 00:11:40,320 --> 00:11:41,920 from almost 95%. 306 00:11:41,920 --> 00:11:42,760 - That high, wow. 307 00:11:42,760 --> 00:11:44,680 - Down to just over 6%. 308 00:11:44,680 --> 00:11:47,240 That's a huge improvement and a really important step 309 00:11:47,240 --> 00:11:50,040 for making these powerful models safer to use widely. 310 00:11:50,040 --> 00:11:51,600 - Definitely need that trust. 311 00:11:51,600 --> 00:11:53,840 And related to trust is knowing what's real, right? 312 00:11:53,840 --> 00:11:55,840 Distinguishing AI-generated stuff. 313 00:11:55,840 --> 00:11:58,000 - Yes, the authenticity issue. 314 00:11:58,000 --> 00:12:01,000 As AI gets better at creating text, images, 315 00:12:01,000 --> 00:12:03,600 even video that looks totally different, 316 00:12:03,600 --> 00:12:06,520 really real, we need ways to tell what's human made and what's not. The sources mentioned 317 00:12:06,520 --> 00:12:12,360 tools like Google's Syntheday Detector, which aims to help identify AI generated content. 318 00:12:12,360 --> 00:12:15,920 That's crucial for tackling misinformation and maintaining trust online. 319 00:12:15,920 --> 00:12:23,140 Absolutely. Okay, let's zoom out a bit. We've talked hardware, models, talent, trust. How 320 00:12:23,140 --> 00:12:27,640 is all this AI actually showing up in the real world now? What tangible impacts are 321 00:12:27,640 --> 00:12:28,640 we seeing? 322 00:12:28,640 --> 00:12:33,460 As soon as some really concrete applications emerge, take Medical Discovery, a company 323 00:12:33,460 --> 00:12:38,260 called Future House, used what they call an autonomous AI system. 324 00:12:38,260 --> 00:12:39,260 Autonomous meaning? 325 00:12:39,260 --> 00:12:44,140 Meaning the AI did parts of the research process itself, analyzing scientific literature, experimental 326 00:12:44,140 --> 00:12:49,720 data. And through that process, it identified an existing drug, Rupasudol, as potentially 327 00:12:49,720 --> 00:12:54,540 promising for treating age-related blindness. They're releasing their results publicly soon, 328 00:12:54,540 --> 00:12:59,860 May 27th. It just shows how AI can really accelerate finding new scientific leads. 329 00:12:59,860 --> 00:13:02,080 And AI finding potential drug candidates. 330 00:13:02,080 --> 00:13:03,080 Yeah. 331 00:13:03,080 --> 00:13:07,300 That's incredible potential. What about in the physical world? Robotics? 332 00:13:07,300 --> 00:13:11,240 Tesla's optimist robot is making progress there. Recent demos showed it doing basic 333 00:13:11,240 --> 00:13:17,000 household chores, sweeping, vacuuming, tearing paper towels, stirring things, opening cabinets. 334 00:13:17,000 --> 00:13:18,000 Stuff around the house, yeah. 335 00:13:18,000 --> 00:13:21,580 And what's interesting is how it's learning. They're training its neural network using 336 00:13:21,580 --> 00:13:24,820 videos recorded from a first-person human perspective. 337 00:13:24,820 --> 00:13:27,100 So it learns by watching humans do tasks? 338 00:13:27,100 --> 00:13:31,600 Essentially, yes. And the plan is to expand that to learning from third-person videos 339 00:13:31,600 --> 00:13:36,620 and even just general videos from the internet. It's a fascinating approach to teaching robots 340 00:13:36,620 --> 00:13:37,920 physical skills. 341 00:13:37,920 --> 00:13:43,100 So AI is not just thinking, it's learning to do physically. And it's getting embedded 342 00:13:43,100 --> 00:13:44,820 in other daily items too. 343 00:13:44,820 --> 00:13:50,100 Definitely. Google's working with Samsung on AI-powered smart glasses. They're aiming 344 00:13:50,100 --> 00:13:53,660 for 2026 running something called Android XR. 345 00:13:53,660 --> 00:13:54,940 Smart glasses. Again. 346 00:13:54,940 --> 00:14:00,860 Yeah. But maybe with more AI focus this time, they might have options like an in-lens display, 347 00:14:00,860 --> 00:14:06,340 real-time translation, interpreting what you're seeing, embedding AI right into your view 348 00:14:06,340 --> 00:14:11,700 of the world. And Volvo is putting Google's Gemini assistant into their cars. So AI is 349 00:14:11,700 --> 00:14:16,700 moving into our vehicles, our eyewear, becoming part of the environment. 350 00:14:16,700 --> 00:14:19,940 So it's becoming less of a destination we go to on a screen and more part of the fabric 351 00:14:19,940 --> 00:14:20,940 around us. 352 00:14:20,940 --> 00:14:24,860 What about the impact on businesses, the economy? Eric Schmidt's views were mentioned. 353 00:14:24,860 --> 00:14:27,020 Right. The former Google CEO. He's been... 354 00:14:27,020 --> 00:14:30,800 arguing quite forcefully that AI is actually under-hyped, not over-hyped. 355 00:14:30,800 --> 00:14:31,800 Under-hyped. 356 00:14:31,800 --> 00:14:32,800 Really. 357 00:14:32,800 --> 00:14:33,960 With all the buzz. 358 00:14:33,960 --> 00:14:37,280 His point is that its potential goes way beyond what we're seeing now. 359 00:14:37,280 --> 00:14:42,360 He talks about AI agents being able to manage entire business processes end to end just 360 00:14:42,360 --> 00:14:44,400 using natural language instructions. 361 00:14:44,400 --> 00:14:45,400 Manage whole processes. 362 00:14:45,400 --> 00:14:46,400 Yeah. 363 00:14:46,400 --> 00:14:51,200 And he suggests that could lead to efficiency gains of maybe 90 percent in some areas. 364 00:14:51,200 --> 00:14:52,200 The 90 percent. 365 00:14:52,200 --> 00:14:54,240 That would fundamentally change how businesses operate. 366 00:14:54,240 --> 00:14:55,240 Absolutely. 367 00:14:55,240 --> 00:14:56,960 It would reshape industries. 368 00:14:56,960 --> 00:15:00,080 And Schmidt connects this to bigger strategic issues, too. 369 00:15:00,080 --> 00:15:05,360 He warns about a serious AI arms race between the U.S. and China for technological dominance. 370 00:15:05,360 --> 00:15:06,360 An arms race. 371 00:15:06,360 --> 00:15:07,360 Yeah. 372 00:15:07,360 --> 00:15:11,840 And he also points out the massive increase in computing power that future AI will need 373 00:15:11,840 --> 00:15:14,700 maybe a hundred to a thousand times what we have now. 374 00:15:14,700 --> 00:15:18,840 These aren't just technical forecasts, they're about global competition and economic power. 375 00:15:18,840 --> 00:15:23,260 So the stakes are incredibly high geopolitically and economically. 376 00:15:23,260 --> 00:15:27,240 The sources also touched on this idea of an AI trough of disillusionment. 377 00:15:27,240 --> 00:15:28,240 What's that about? 378 00:15:28,240 --> 00:15:29,240 Uh. 379 00:15:29,240 --> 00:15:31,720 That comes from the Gartner hype cycle idea. 380 00:15:31,720 --> 00:15:36,100 After the initial peak of excitement, there's often a period where progress seems to slow 381 00:15:36,100 --> 00:15:40,420 down, challenges become clearer, and some of the initial hype fades. 382 00:15:40,420 --> 00:15:41,420 Right. 383 00:15:41,420 --> 00:15:42,420 Reality sets in. 384 00:15:42,420 --> 00:15:43,420 Exactly. 385 00:15:43,420 --> 00:15:47,740 And the discussion is around what it takes to get through that phase and make AI work 386 00:15:47,740 --> 00:15:49,440 sustainably. 387 00:15:49,440 --> 00:15:54,440 Things like strong leadership, continued investment in research labs, and importantly, involving 388 00:15:54,440 --> 00:15:58,300 the crowd users providing data, feedback, real world testing. 389 00:15:58,300 --> 00:16:00,540 So keeping the momentum going through the hard part. 390 00:16:00,540 --> 00:16:01,540 Precisely. 391 00:16:01,540 --> 00:16:05,980 There was also that interesting, maybe slightly provocative mention of Sean Pury's idea that 392 00:16:05,980 --> 00:16:10,060 the next Zuckerberg might use AI companions to tackle loneliness. 393 00:16:10,060 --> 00:16:11,660 AI companions for loneliness. 394 00:16:11,660 --> 00:16:12,660 Wow. 395 00:16:12,660 --> 00:16:16,460 It suggests AI could start impacting really fundamental human social needs, for better 396 00:16:16,460 --> 00:16:17,460 or worse. 397 00:16:17,460 --> 00:16:19,800 So let's connect these big picture ideas back to you, the listener. 398 00:16:19,800 --> 00:16:20,800 Yeah. 399 00:16:20,800 --> 00:16:25,320 How does talk of an AI arms race or a trough of disillusionment or these super efficient 400 00:16:25,320 --> 00:16:28,120 AI agents, how does that affect you personally? 401 00:16:28,120 --> 00:16:31,540 Well, these large scale trends shape everything. 402 00:16:31,540 --> 00:16:37,680 National focus on an arms race means huge government and private investment in AI R&D, 403 00:16:37,680 --> 00:16:40,880 which influences technology direction and job creation. 404 00:16:40,880 --> 00:16:41,880 Okay. 405 00:16:41,880 --> 00:16:47,100 Navigating a trough means the focus shifts to practical, valuable applications. 406 00:16:47,100 --> 00:16:51,500 where many job opportunities might actually lie in making AI useful day to day. 407 00:16:51,500 --> 00:16:55,140 and AI agents automating business processes, 408 00:16:55,140 --> 00:16:57,240 that could fundamentally change the nature of work. 409 00:16:57,240 --> 00:16:58,080 - How so? 410 00:16:58,080 --> 00:17:00,680 - It might shift jobs away from performing routine tasks 411 00:17:00,680 --> 00:17:02,200 towards managing, overseeing, 412 00:17:02,200 --> 00:17:04,180 and directing these AI systems. 413 00:17:04,180 --> 00:17:05,620 Skills like prompt engineering, 414 00:17:05,620 --> 00:17:09,360 strategic thinking about AI deployment, ethical oversight, 415 00:17:09,360 --> 00:17:11,060 these could become much more crucial. 416 00:17:11,060 --> 00:17:13,860 It forces all of us really to think about how we adapt 417 00:17:13,860 --> 00:17:15,860 and what skills we need in a world where AI 418 00:17:15,860 --> 00:17:17,820 is becoming more integrated and autonomous. 419 00:17:17,820 --> 00:17:19,360 - Right, it's not just about using the tools, 420 00:17:19,360 --> 00:17:20,600 but shaping how they're used. 421 00:17:20,600 --> 00:17:21,440 - Yeah. 422 00:17:21,440 --> 00:17:22,260 - We've got a ton today. 423 00:17:22,260 --> 00:17:24,660 Open AI and Joanie Ive's big hardware play. 424 00:17:24,660 --> 00:17:25,740 - A huge bet. 425 00:17:25,740 --> 00:17:28,700 - The absolute explosion of new AI models 426 00:17:28,700 --> 00:17:30,800 and tools for developers, for creatives, for everyone. 427 00:17:30,800 --> 00:17:32,180 - The pace is relentless. 428 00:17:32,180 --> 00:17:35,240 - The dynamics in the AI talent market, tough for some, 429 00:17:35,240 --> 00:17:36,440 huge demand for others. 430 00:17:36,440 --> 00:17:37,940 - A real shift happening there. 431 00:17:37,940 --> 00:17:41,720 - Plus the critical work on trust, security, 432 00:17:41,720 --> 00:17:43,060 detecting AI content. 433 00:17:43,060 --> 00:17:44,220 - Essential foundations. 434 00:17:44,220 --> 00:17:46,900 - And then all these real world applications popping up 435 00:17:46,900 --> 00:17:50,160 from medicine to robotics to smart classes, 436 00:17:50,160 --> 00:17:52,560 alongside these massive strategic questions 437 00:17:52,560 --> 00:17:54,860 about the economy and global competition. 438 00:17:54,860 --> 00:17:57,540 - Yeah, putting it all together from these sources, 439 00:17:57,540 --> 00:18:01,040 the picture is one of really rapid fundamental change. 440 00:18:01,040 --> 00:18:02,840 It's not just tech updates. 441 00:18:02,840 --> 00:18:06,180 It's shifts in industry structure, jobs, 442 00:18:06,180 --> 00:18:08,680 and potentially how we live and interact. 443 00:18:08,680 --> 00:18:09,520 - Absolutely. 444 00:18:09,520 --> 00:18:11,880 And look, if you found this deep dive valuable, 445 00:18:11,880 --> 00:18:14,580 if you appreciate us trying to connect these dots for you. 446 00:18:14,580 --> 00:18:15,720 - Which we hope you do. 447 00:18:15,720 --> 00:18:17,580 - The best way you can help us keep doing this 448 00:18:17,580 --> 00:18:18,660 is actually free. 449 00:18:18,660 --> 00:18:20,060 Just take in a second to rate the show, 450 00:18:20,060 --> 00:18:22,160 like it, subscribe on whatever platform you use, 451 00:18:22,160 --> 00:18:24,620 Spotify, Apple Podcasts, wherever. 452 00:18:24,620 --> 00:18:26,560 It genuinely helps other people find us. 453 00:18:26,560 --> 00:18:27,860 - It really does make a difference. 454 00:18:27,860 --> 00:18:29,400 - You could also join our Discord community. 455 00:18:29,400 --> 00:18:32,300 It's growing, lots of smart people discussing these topics. 456 00:18:32,300 --> 00:18:34,840 Or if you're able, becoming a Patreon member 457 00:18:34,840 --> 00:18:37,000 offers more ways to support us directly. 458 00:18:37,000 --> 00:18:38,940 - We appreciate any support you can offer. 459 00:18:38,940 --> 00:18:40,820 It lets us keep digging into these sources 460 00:18:40,820 --> 00:18:42,820 and bringing you this analysis. 461 00:18:42,820 --> 00:18:45,380 - So to leave you with something to chew on, 462 00:18:45,380 --> 00:18:46,660 given everything we've talked about, 463 00:18:46,660 --> 00:18:50,120 the speed, the talent race, the push towards new forms, 464 00:18:50,120 --> 00:18:55,120 like screen-free AI, AI in your glasses, AI driving cars. 465 00:18:55,120 --> 00:18:58,600 As AI shifts from being this thing on our screens 466 00:18:58,600 --> 00:19:02,140 to potentially being this always-on, context-aware presence 467 00:19:02,140 --> 00:19:03,900 woven into our lives, 468 00:19:03,900 --> 00:19:05,860 what's the biggest unintended consequence 469 00:19:05,860 --> 00:19:07,700 we should be thinking hardest about right now? 470 00:19:07,700 --> 00:19:10,020 What happens when AI is truly ubiquitous? 471 00:19:10,020 --> 00:19:12,020 - That's a really important question to ponder 472 00:19:12,020 --> 00:19:13,040 as all this unfolds. 473 00:19:13,040 --> 00:19:14,380 - Definitely something to think about. 474 00:19:14,380 --> 00:19:16,420 Thanks for joining us for this deep dive.