WEBVTT 00:00:00.000 --> 00:00:03.620 Is the AI boom of today a bubble? A big one, 00:00:03.700 --> 00:00:06.599 maybe. Potentially even bigger, perhaps riskier, 00:00:06.700 --> 00:00:10.119 than the dot -com frenzy back in the 90s. Two 00:00:10.119 --> 00:00:12.679 -step silence. That's quite a thought, isn't 00:00:12.679 --> 00:00:14.580 it? Especially when you think how much AI is 00:00:14.580 --> 00:00:17.420 already part of things. Welcome to the Deep Talk. 00:00:17.519 --> 00:00:18.920 Today, we're going to pull back the curtain a 00:00:18.920 --> 00:00:21.519 bit on some really big recent insights into this 00:00:21.519 --> 00:00:24.300 AI landscape. It's moving so fast. Yeah, it really 00:00:24.300 --> 00:00:27.579 is exponentially fast, it feels like. Almost 00:00:27.579 --> 00:00:29.359 impossible to keep track of everything. Right. 00:00:29.519 --> 00:00:32.140 Absolutely. So our mission today, really, is 00:00:32.140 --> 00:00:34.039 to try and cut through some of that noise for 00:00:34.039 --> 00:00:36.140 you. We're going to take a journey, look at a 00:00:36.140 --> 00:00:39.159 few key things. First up... The big money questions 00:00:39.159 --> 00:00:41.840 is all this AI hype sustainable, really. Then 00:00:41.840 --> 00:00:43.920 we'll switch gears, look at some fascinating 00:00:43.920 --> 00:00:46.880 actual applications, things happening, you know, 00:00:46.880 --> 00:00:50.539 on the ground, industry moves, stuff away from 00:00:50.539 --> 00:00:52.859 just the stock market headlines. And then finally, 00:00:52.979 --> 00:00:54.759 we're going to wrap up with something kind of 00:00:54.759 --> 00:00:57.619 mind bending, an idea that might really change 00:00:57.619 --> 00:01:00.140 how we think about AI itself. What if all these 00:01:00.140 --> 00:01:03.119 different AI models, what if they're all... Basically 00:01:03.119 --> 00:01:05.640 learning the same thing, converging on some kind 00:01:05.640 --> 00:01:07.859 of, I don't know, universal understanding. Exactly. 00:01:07.879 --> 00:01:09.939 It's a wild concept. Makes you stop and think. 00:01:10.140 --> 00:01:12.480 OK, so let's start there with that big claim 00:01:12.480 --> 00:01:14.579 from the top. We've been looking at what Torsten 00:01:14.579 --> 00:01:16.939 Slick, he's Apollo's chief economist, has been 00:01:16.939 --> 00:01:18.900 saying. And he doesn't really pull any punches. 00:01:19.000 --> 00:01:23.280 He says today's AI boom is more dangerously inflated 00:01:23.280 --> 00:01:25.760 than the 90s tech bubble. Wow. And this time 00:01:25.760 --> 00:01:29.239 he argues there's just much more at stake. Yeah. 00:01:29.340 --> 00:01:31.260 And what's really striking is the data he points 00:01:31.260 --> 00:01:34.500 to. He looked at the top 10 companies in the 00:01:34.500 --> 00:01:38.060 S &P 500. You know the ones. NVIDIA, Meta, Microsoft, 00:01:38.480 --> 00:01:40.879 Google, Amazon. His internal numbers are saying 00:01:40.879 --> 00:01:42.900 they're more overvalued right now than the biggest 00:01:42.900 --> 00:01:45.079 tech names were back then, like at the absolute 00:01:45.079 --> 00:01:47.939 peak of the late 90s. That is a really strong 00:01:47.939 --> 00:01:51.319 statement. What's behind that specifically? Well, 00:01:51.379 --> 00:01:54.120 it's their price to earnings ratios. P .E. ratios. 00:01:54.459 --> 00:01:57.540 He calls them dot combusting levels. And for 00:01:57.540 --> 00:02:00.340 anyone maybe not tracking markets day to day, 00:02:00.400 --> 00:02:02.819 the P .E. ratio is basically how much investors 00:02:02.819 --> 00:02:05.439 are willing to pay for, you know, a dollar of 00:02:05.439 --> 00:02:07.519 the company's profit. Right. High P .E. means 00:02:07.519 --> 00:02:09.840 high expectations for the future. Exactly. But 00:02:09.840 --> 00:02:13.599 dot combusting levels that kind of suggests expectations 00:02:13.599 --> 00:02:16.219 might be, let's say, a bit out of whack with 00:02:16.219 --> 00:02:19.280 reality or potential reality. And what's different 00:02:19.280 --> 00:02:21.539 this time? If he's right, if it is a bubble. 00:02:22.150 --> 00:02:24.469 What makes it different from the 90s one? The 00:02:24.469 --> 00:02:26.370 big thing, and this is what makes it kind of 00:02:26.370 --> 00:02:29.129 concerning, is the concentration risk. It's extreme, 00:02:29.330 --> 00:02:32.469 almost 40%. Think about that. 40 % of the entire 00:02:32.469 --> 00:02:35.330 S &P 500 is tied up in just those 10 companies. 00:02:35.530 --> 00:02:37.930 Wow. So if you're just buying a broad market 00:02:37.930 --> 00:02:40.590 index fund thinking you're diversified, you're 00:02:40.590 --> 00:02:43.169 not really. You're making a massive bet on AI 00:02:43.169 --> 00:02:45.550 hype continuing. So if those few stumble. The 00:02:45.550 --> 00:02:47.830 whole market feels it, much more severely potentially 00:02:47.830 --> 00:02:50.469 than back in the 90s. It really makes you ask, 00:02:50.509 --> 00:02:52.550 you know, how much of this value is real, tangible 00:02:52.550 --> 00:02:55.169 stuff, and how much is, well, just hype, belief. 00:02:55.719 --> 00:02:57.639 It's just staggering amounts of money moving 00:02:57.639 --> 00:03:00.180 around. And we see these huge financial plays 00:03:00.180 --> 00:03:03.000 that seem to, well, feed that bubble perception. 00:03:03.620 --> 00:03:06.400 NVIDIA, for example, reportedly planning to spend 00:03:06.400 --> 00:03:10.860 $500 billion on AI factory. $500 billion. It's 00:03:10.860 --> 00:03:13.379 hard to even picture. Right. And then Meta. Yeah. 00:03:14.120 --> 00:03:16.879 Supposedly offering like $100 million signing 00:03:16.879 --> 00:03:20.479 bonuses for top AI people. $100 million. For 00:03:20.479 --> 00:03:22.620 one person. It's a talent war out there, for 00:03:22.620 --> 00:03:25.219 sure. And it keeps going. Amazon maybe putting 00:03:25.219 --> 00:03:27.900 another $8 billion into Anthropic. And OpenAI, 00:03:28.020 --> 00:03:31.060 remember they bought ScaleAI for, what, $14 billion? 00:03:31.460 --> 00:03:33.580 Yeah. And then pretty soon after laid off 200 00:03:33.580 --> 00:03:36.139 people. It makes you wonder about the strategy 00:03:36.139 --> 00:03:38.620 sometimes. Is it real growth or just grabbing 00:03:38.620 --> 00:03:40.840 land in a gold rush? You know, if you look at 00:03:40.840 --> 00:03:43.020 the market sentiment overall, it gets even clearer 00:03:43.020 --> 00:03:45.800 why people are using words like frothy. There's 00:03:45.800 --> 00:03:49.479 this index. BUZZ Next Gen AI Sentiment Index 00:03:49.479 --> 00:03:52.060 tracks AI hype with retail traders, basically. 00:03:52.360 --> 00:03:55.780 It's apparently a 45 % and just 16 once, trading 00:03:55.780 --> 00:03:58.419 29 % above its 200 -day moving average. Which 00:03:58.419 --> 00:04:00.639 is pretty high. Yeah, the 200 -day average is 00:04:00.639 --> 00:04:02.979 a key thing traders watch for long -term trends. 00:04:03.479 --> 00:04:07.360 Being that far above it, it signals extreme optimism. 00:04:07.819 --> 00:04:10.520 Almost like you said, a fever pitch. Right. BT's 00:04:10.520 --> 00:04:12.419 analyst just came out and said it's looking very 00:04:12.419 --> 00:04:14.780 frothy. That's the direct quote. Very frothy. 00:04:14.900 --> 00:04:16.740 It really does feel like everyone's jumping on 00:04:16.740 --> 00:04:18.660 the same train and it's just going faster and 00:04:18.660 --> 00:04:21.240 faster. So let's connect this back. What does 00:04:21.240 --> 00:04:24.660 this mean for you listening right now? If this 00:04:24.660 --> 00:04:28.480 bubble or frothiness or whatever, if it pops, 00:04:28.620 --> 00:04:30.860 well, the good news probably is it won't kill 00:04:30.860 --> 00:04:33.790 AI itself, the tech. Yeah, just like the dot 00:04:33.790 --> 00:04:35.850 -com crash didn't kill the internet. Right, exactly. 00:04:36.410 --> 00:04:38.689 The underlying tech is definitely moving forward. 00:04:38.730 --> 00:04:41.670 That innovation is real. But that doesn't mean 00:04:41.670 --> 00:04:43.290 NVIDIA should be priced like it's going to run 00:04:43.290 --> 00:04:45.810 the whole global economy. Fair point. So the 00:04:45.810 --> 00:04:48.230 fallout from a big correction could be pretty 00:04:48.230 --> 00:04:50.529 rough. It could hit general market funds, maybe 00:04:50.529 --> 00:04:53.269 even hurt funding for new AI startups trying 00:04:53.269 --> 00:04:56.569 to get off the ground. So thinking about right 00:04:56.569 --> 00:04:59.490 now this market, what's the biggest immediate 00:04:59.490 --> 00:05:01.810 risk for an investor, would you say? I'd say 00:05:01.810 --> 00:05:04.170 overexposure, being too heavily invested in just 00:05:04.170 --> 00:05:06.750 a few of these really volatile, highly valued 00:05:06.750 --> 00:05:10.850 companies. If expectations slip, losses could 00:05:10.850 --> 00:05:15.149 be significant. Okay. Let's pivot then. Away 00:05:15.149 --> 00:05:17.649 from the finance side, the potential frothiness, 00:05:17.829 --> 00:05:19.769 let's look at what's actually happening on the 00:05:19.769 --> 00:05:22.350 ground. Because beyond all the market talk, there 00:05:22.350 --> 00:05:26.110 are some really interesting real world uses and 00:05:26.110 --> 00:05:29.009 industry moves going on. Yeah, absolutely. And 00:05:29.009 --> 00:05:31.189 what's kind of wild is how fast some of these 00:05:31.189 --> 00:05:33.850 tools are scaling up. Look at Perplexity, the 00:05:33.850 --> 00:05:36.509 AI answer engine. They partnered with ITEL over 00:05:36.509 --> 00:05:39.470 in India, offered free Perplexity Pro for a year 00:05:39.470 --> 00:05:44.029 to all 360 million of ITEL's customer. 360 million. 00:05:44.170 --> 00:05:46.949 Yep. And boom, Perplexity shot straight to number 00:05:46.949 --> 00:05:49.339 one on the app charts in India. Overnight, practically. 00:05:49.339 --> 00:05:52.240 That's a whole new level of user acquisition. 00:05:52.300 --> 00:05:54.379 It really shows a new model for getting AI out 00:05:54.379 --> 00:05:56.579 there globally super fast. Incredible reach. 00:05:56.680 --> 00:05:58.000 And it shows people really want these tools. 00:05:58.180 --> 00:06:00.480 But speaking of services, there's this interesting 00:06:00.480 --> 00:06:03.720 point Mert Davici made that maybe AI won't replace 00:06:03.720 --> 00:06:05.779 a lot of services after all. Oh, yeah. Why is 00:06:05.779 --> 00:06:08.819 that? His argument is basically customers want 00:06:08.819 --> 00:06:12.019 to pay software prices for software, not salary 00:06:12.019 --> 00:06:15.139 prices. Ah, OK. Makes sense. So if an AI service 00:06:15.139 --> 00:06:18.699 costs as much as hiring a person. Maybe people 00:06:18.699 --> 00:06:20.879 just stick with the person. Yeah. Yeah. Who wants 00:06:20.879 --> 00:06:23.819 to pay a robot a full salary for like writing 00:06:23.819 --> 00:06:25.779 emails? There's definitely a psychological thing 00:06:25.779 --> 00:06:28.459 there. Definitely. And then on the voice side, 00:06:28.620 --> 00:06:32.000 Hume AI, their EVI 2 model was already pretty 00:06:32.000 --> 00:06:34.699 cool, making AI voices sound unique, kind of 00:06:34.699 --> 00:06:38.779 emotional. Now they've got EVI 3 out, supposedly 00:06:38.779 --> 00:06:40.980 captures even more personality, more nuance. 00:06:41.240 --> 00:06:42.939 So what can you do with that? Well, think about 00:06:42.939 --> 00:06:45.500 building really natural language coaches or podcast 00:06:45.500 --> 00:06:48.120 hosts that sound totally real or even like compassionate 00:06:48.120 --> 00:06:50.759 AI companions. It's about making AI interaction 00:06:50.759 --> 00:06:54.519 feel much more human. And here's something I 00:06:54.519 --> 00:06:56.060 think a lot of our listeners will find really 00:06:56.060 --> 00:06:57.839 interesting. An open AI researcher who actually 00:06:57.839 --> 00:07:00.540 just joined Meta. gave this 30 minute masterclass 00:07:00.540 --> 00:07:03.579 online about how to create wealth in the AI economy. 00:07:03.740 --> 00:07:05.620 Oh, really? Yeah. And it apparently got like 00:07:05.620 --> 00:07:08.180 at least a million views super quickly. Wow. 00:07:08.379 --> 00:07:10.480 People are clearly hungry for that practical 00:07:10.480 --> 00:07:12.920 side. How do I actually use this stuff? Exactly. 00:07:13.120 --> 00:07:15.139 It shows AI isn't just abstract tech anymore. 00:07:15.220 --> 00:07:18.819 It's affecting careers, business models right 00:07:18.819 --> 00:07:21.250 now. And for practical tools that kind of do 00:07:21.250 --> 00:07:23.730 things for you, there's Perplexity Comet. People 00:07:23.730 --> 00:07:26.769 are calling it scary good. No. Imagine an AI 00:07:26.769 --> 00:07:30.230 assistant that hooks into your app's email, calendar, 00:07:30.449 --> 00:07:32.949 project tools. Okay. And just does stuff for 00:07:32.949 --> 00:07:36.509 you autonomously. They highlighted like 10 powerful 00:07:36.509 --> 00:07:39.329 use cases, summarizing meetings, drafting complex 00:07:39.329 --> 00:07:41.720 emails. That sounds like it. A proper assistant. 00:07:42.180 --> 00:07:44.480 Finally. Feels like a real step towards AI that 00:07:44.480 --> 00:07:46.779 doesn't just respond, but takes initiative. That 00:07:46.779 --> 00:07:49.279 could be a huge productivity boost. And speaking 00:07:49.279 --> 00:07:51.339 of things taking off, another funding story that's 00:07:51.339 --> 00:07:54.680 just explosive. Lovable, the Swedish AI app builder. 00:07:54.839 --> 00:07:57.620 Lovable, yeah. Heard about them. They hit $75 00:07:57.620 --> 00:08:01.220 million in annual recurring revenue. In seven 00:08:01.220 --> 00:08:04.079 months. Seven months, seriously. Yeah, 2 .3 million 00:08:04.079 --> 00:08:07.259 users, 180 ,000 paying customers. And they just 00:08:07.259 --> 00:08:10.800 raised $200 million Series A. at a $1 .8 billion 00:08:10.800 --> 00:08:13.040 valuation only eight months after they launched. 00:08:13.259 --> 00:08:15.860 Okay, that's insane growth. Just shows if you 00:08:15.860 --> 00:08:17.439 build something AI -powered that people actually 00:08:17.439 --> 00:08:20.660 need. It can scale incredibly fast. So looking 00:08:20.660 --> 00:08:23.079 at all these different applications, Perplexity, 00:08:23.100 --> 00:08:27.220 Hume, Lovable, what's the common thread? What 00:08:27.220 --> 00:08:29.639 stands out about how these tools are actually 00:08:29.639 --> 00:08:31.680 finding their place? I think it's that they're 00:08:31.680 --> 00:08:34.480 solving specific high -value problems. They create 00:08:34.480 --> 00:08:37.360 real value for users, and because of that, they 00:08:37.360 --> 00:08:40.110 can scale with just... Unbelievable speed. Right. 00:08:40.149 --> 00:08:41.909 Let's shift again. Just a few quick hits now. 00:08:42.009 --> 00:08:44.669 Kind of snapshots of how AI is evolving day by 00:08:44.669 --> 00:08:46.789 day almost. Lots of little interesting bits. 00:08:46.970 --> 00:08:50.029 Okay. Rapid fire. We saw this piece on 10 examples 00:08:50.029 --> 00:08:53.029 of how brutal robot testing made them more robust. 00:08:53.909 --> 00:08:56.610 It's kind of cool looking at the extreme engineering 00:08:56.610 --> 00:08:58.970 needed to make physical robots tough enough for 00:08:58.970 --> 00:09:01.549 the real world. It's not all just software. Right. 00:09:01.610 --> 00:09:05.250 The hardware side. And Elon Musk's XAI, they're 00:09:05.250 --> 00:09:07.970 launching something called Baby Grok, a chatbot 00:09:07.970 --> 00:09:10.769 for kids learning. Huh. AI for kids. That's interesting. 00:09:10.909 --> 00:09:13.210 Shows how it's moving into education and sensitive 00:09:13.210 --> 00:09:15.750 areas. Raises questions about safety, too. Definitely. 00:09:15.789 --> 00:09:19.029 And for users. Anyone trying to get better results 00:09:19.029 --> 00:09:21.669 from AI. And Fropic put out some advice on writing 00:09:21.669 --> 00:09:23.759 effective prompts. Well, that's useful. Yeah. 00:09:23.860 --> 00:09:25.799 I mean, I still wrestle with prompt drift myself 00:09:25.799 --> 00:09:28.179 sometimes. You start asking for one thing and 00:09:28.179 --> 00:09:29.960 the AI kind of wanders off. Totally know what 00:09:29.960 --> 00:09:31.620 you mean. So yeah, any tips on prompting are 00:09:31.620 --> 00:09:34.840 gold. And another practical thing, DuckDuckGo, 00:09:34.960 --> 00:09:38.080 the search engine, they added a feature to let 00:09:38.080 --> 00:09:40.889 you block AI -generated images. Do you want to 00:09:40.889 --> 00:09:43.110 filter those out? Oh, interesting. Giving users 00:09:43.110 --> 00:09:46.610 control over authenticity. And then on the policy 00:09:46.610 --> 00:09:49.350 front, kind of big news. Meta decided not to 00:09:49.350 --> 00:09:52.789 sign the EU's AI code of practice. After reviewing 00:09:52.789 --> 00:09:55.529 it carefully, they said no. OK, why is that significant? 00:09:55.830 --> 00:09:58.429 Well, it signals a potential split, right? Between 00:09:58.429 --> 00:10:01.549 how major tech companies and regulators like 00:10:01.549 --> 00:10:04.529 the EU see AI governance shaping up could mean 00:10:04.529 --> 00:10:06.710 different rules in different places. Raises questions 00:10:06.710 --> 00:10:09.149 about how global companies navigate that. OK, 00:10:09.210 --> 00:10:11.549 so just from. Those few quick examples. Right. 00:10:11.929 --> 00:10:15.370 Robots, kids chat bots, prompting, image blocking, 00:10:15.570 --> 00:10:18.830 EU codes. What's the main takeaway? I'd say AI 00:10:18.830 --> 00:10:20.950 is getting really specialized really fast. And 00:10:20.950 --> 00:10:24.230 that's creating totally new needs for users and 00:10:24.230 --> 00:10:26.190 definitely raising some tricky new regulatory 00:10:26.190 --> 00:10:28.710 challenges. Amid a role sponsor replace holder. 00:10:29.070 --> 00:10:31.950 Okay. Let's unpack this next part. Because here's 00:10:31.950 --> 00:10:34.330 where it gets, I think, really interesting. The 00:10:34.330 --> 00:10:37.350 big idea we sort of hinted at earlier. Imagine 00:10:37.350 --> 00:10:43.840 this. What if all AI models, all the different 00:10:43.840 --> 00:10:46.120 ones, different companies, different data, what 00:10:46.120 --> 00:10:48.799 if they're all fundamentally learning the same 00:10:48.799 --> 00:10:51.659 core thing? Like they're all uncovering pieces 00:10:51.659 --> 00:10:54.620 of the same universal puzzle about how the world 00:10:54.620 --> 00:10:57.059 works. Exactly. That's the fascinating premise 00:10:57.059 --> 00:10:59.620 here. The idea is as these models get bigger, 00:10:59.700 --> 00:11:03.039 get more complex, get smarter, essentially, they 00:11:03.039 --> 00:11:05.299 start learning the same basic relationships about 00:11:05.299 --> 00:11:08.000 the world. And they represent those relationships 00:11:08.000 --> 00:11:10.120 inside themselves in ways that are surprisingly 00:11:10.120 --> 00:11:12.059 similar. Even if they were trained differently. 00:11:12.240 --> 00:11:14.279 Yeah. It kind of suggests there might be one 00:11:14.279 --> 00:11:16.679 ideal sort of universal way to understand and 00:11:16.679 --> 00:11:19.019 represent knowledge. And maybe all these big, 00:11:19.080 --> 00:11:21.440 powerful models are just converging on it. Like 00:11:21.440 --> 00:11:23.299 they're all accidentally tuning into the same, 00:11:23.340 --> 00:11:25.460 I don't know, cosmic frequency of understanding. 00:11:25.840 --> 00:11:28.419 And this isn't just like a philosophical idea 00:11:28.419 --> 00:11:31.440 someone had. No, no. MIT actually put out research 00:11:31.440 --> 00:11:35.179 formalizing this in 2024. They showed that bigger 00:11:35.179 --> 00:11:37.990 models. both for vision, like understanding images, 00:11:38.129 --> 00:11:41.789 and for language, they learn features internally 00:11:41.789 --> 00:11:44.929 that are strikingly similar. So basically, the 00:11:44.929 --> 00:11:47.350 smarter they get, the more their internal maps 00:11:47.350 --> 00:11:49.190 of the world start to look alike. That's the 00:11:49.190 --> 00:11:51.049 gist of it, yeah. It's like they're all finding 00:11:51.049 --> 00:11:53.269 the same fundamental patterns in reality, no 00:11:53.269 --> 00:11:55.009 matter what specific data they started with. 00:11:55.090 --> 00:11:58.850 Okay, wow. Now, the really wild part of this 00:11:58.850 --> 00:12:01.830 research you mentioned, trying to turn the AI's 00:12:01.830 --> 00:12:04.850 internal thought that abstract number stuff, 00:12:04.970 --> 00:12:07.210 the vector, back into the exact words that went 00:12:07.210 --> 00:12:09.350 in. Yeah, that's the really hard part. Like trying 00:12:09.350 --> 00:12:12.370 to rebuild a complex blueprint just by, I don't 00:12:12.370 --> 00:12:14.210 know, feeling the vibes in the finished building. 00:12:14.309 --> 00:12:16.330 It's super difficult. So how on earth did they 00:12:16.330 --> 00:12:19.169 manage that? Well, they used this method, iterative 00:12:19.169 --> 00:12:21.710 refinement, and millions, literally millions 00:12:21.710 --> 00:12:24.090 of queries, basically asking the AI again and 00:12:24.090 --> 00:12:26.330 again, is this closer? How about this? Refining 00:12:26.330 --> 00:12:28.289 the output until it matched the original input. 00:12:28.570 --> 00:12:31.429 And did it work? One team hit 94 % accuracy. 00:12:32.059 --> 00:12:36.500 On long sentences. 94%. Yeah. Think about that. 00:12:36.799 --> 00:12:39.740 They could take the AI's abstract internal state, 00:12:39.899 --> 00:12:42.519 its thought, and translate it back into perfectly 00:12:42.519 --> 00:12:45.080 readable human language. That's like getting 00:12:45.080 --> 00:12:47.740 a peek inside its head. It really is. A remarkable 00:12:47.740 --> 00:12:51.000 glimpse into the AI's sort of cognitive process. 00:12:51.320 --> 00:12:54.899 Okay. So if this is true, if all big models are 00:12:54.899 --> 00:12:57.159 essentially starting to speak the same internal 00:12:57.159 --> 00:13:00.159 language, what does that actually mean? What 00:13:00.159 --> 00:13:02.580 are the implications practically? Philosophically. 00:13:02.620 --> 00:13:05.740 Oh, the implications are huge. Potentially revolutionary, 00:13:06.059 --> 00:13:08.659 honestly. If we can really understand this universal 00:13:08.659 --> 00:13:12.139 internal language, we could build universal inverters, 00:13:12.200 --> 00:13:15.679 tools to actually understand why an AI gives 00:13:15.679 --> 00:13:18.519 a certain output, not just what the output is. 00:13:18.639 --> 00:13:20.740 Which would be massive for trust, for debugging. 00:13:21.019 --> 00:13:23.779 Exactly. For finding hidden biases, for making 00:13:23.779 --> 00:13:26.379 AI more transparent. Huge. And it could mean 00:13:26.379 --> 00:13:29.519 we can translate between closed models. Making 00:13:29.519 --> 00:13:31.360 different AIs from different companies and be 00:13:31.360 --> 00:13:33.259 built totally differently, making them talk to 00:13:33.259 --> 00:13:36.379 each other seamlessly. So like a Microsoft AI 00:13:36.379 --> 00:13:38.679 could understand an open AI model, even if they 00:13:38.679 --> 00:13:40.679 weren't designed to connect. Potentially, yeah. 00:13:41.419 --> 00:13:43.600 Imagine the possibilities for combining different 00:13:43.600 --> 00:13:46.940 AI strengths. Interoperability could just explode. 00:13:47.340 --> 00:13:49.700 OK, that alone is huge. But you mentioned going 00:13:49.700 --> 00:13:52.919 even further. Right. So if there's this shared 00:13:52.919 --> 00:13:56.090 structure of understanding. Maybe it could help 00:13:56.090 --> 00:13:58.490 us decode things we don't understand, like ancient 00:13:58.490 --> 00:14:01.190 human texts. Think about Linear A, the Minoan 00:14:01.190 --> 00:14:03.529 script. We still can't fully read it after decades 00:14:03.529 --> 00:14:06.590 of trying. What if an AI, seeing these universal 00:14:06.590 --> 00:14:10.330 patterns, could crack it? Whoa. An AI unlocking 00:14:10.330 --> 00:14:13.250 secrets of human history. That's humbling. Isn't 00:14:13.250 --> 00:14:15.549 it? And then, here's the really mind -bending 00:14:15.549 --> 00:14:17.009 one, the one that makes you question everything. 00:14:17.269 --> 00:14:19.950 What if, what if we could talk to whales? Talk 00:14:19.950 --> 00:14:23.059 to whales. Yeah. There's already Project CETI 00:14:23.059 --> 00:14:25.720 using AI to try and decode whale communication. 00:14:26.039 --> 00:14:28.559 If there is some kind of shared conceptual space, 00:14:28.700 --> 00:14:31.220 this universal language, could AI bridge the 00:14:31.220 --> 00:14:33.159 gap? Could we find shared semantic structures 00:14:33.159 --> 00:14:35.480 between human thought and whale thought? Whoa. 00:14:35.799 --> 00:14:37.899 Right. Imagine scaling that. A billion queries. 00:14:38.399 --> 00:14:40.580 Maybe it decodes ancient languages. Maybe it 00:14:40.580 --> 00:14:42.179 lets us communicate with another intelligent 00:14:42.179 --> 00:14:45.740 species on Earth. Wow. It makes you rethink what 00:14:45.740 --> 00:14:48.159 intelligence even is. Artificial, biological. 00:14:48.539 --> 00:14:52.220 Yeah. Yeah, breathtaking. Does this convergence 00:14:52.220 --> 00:14:55.759 idea, does it mean AI might one day actually 00:14:55.759 --> 00:14:58.940 understand us, not just process our words, but 00:14:58.940 --> 00:15:01.620 get the concepts, the feelings behind them? It 00:15:01.620 --> 00:15:04.320 certainly suggests a path towards that, a shared 00:15:04.320 --> 00:15:07.399 conceptual space, a deep structure. Yeah. It 00:15:07.399 --> 00:15:10.100 feels like it could bridge human and AI thought 00:15:10.100 --> 00:15:12.840 in a way that goes beyond just processing, maybe 00:15:12.840 --> 00:15:16.100 towards real comprehension. So, OK, we've covered 00:15:16.100 --> 00:15:18.860 a lot. We navigated the choppy financial waters 00:15:18.860 --> 00:15:21.629 of the AI market. The frothiness. Right. We've 00:15:21.629 --> 00:15:23.669 seen all these diverse real world applications 00:15:23.669 --> 00:15:25.769 popping up everywhere, scaling incredibly fast. 00:15:26.009 --> 00:15:27.870 And then we peered into this possible future 00:15:27.870 --> 00:15:30.549 where AI models might actually share a universal 00:15:30.549 --> 00:15:32.990 language. It's quite the deep dive from economics 00:15:32.990 --> 00:15:35.750 all the way to, well, philosophy almost. Yeah. 00:15:35.789 --> 00:15:37.370 And I think the key tension is clear, right? 00:15:37.389 --> 00:15:39.049 Despite all the market craziness, the volatility, 00:15:39.269 --> 00:15:41.269 the hype, the underlying technology is advancing, 00:15:41.549 --> 00:15:43.950 seriously advancing. And it keeps revealing these 00:15:43.950 --> 00:15:46.649 profound, almost philosophical questions about 00:15:46.649 --> 00:15:48.990 intelligence, about reality, maybe even about 00:15:48.990 --> 00:15:51.720 ourselves. So as you go about your day today, 00:15:51.860 --> 00:15:53.980 maybe just take a moment to think about this. 00:15:54.299 --> 00:15:58.000 If AI models are converging on some single ideal 00:15:58.000 --> 00:16:01.639 way to understand the world, what does that tell 00:16:01.639 --> 00:16:04.100 us about reality itself? Is there some inherent 00:16:04.100 --> 00:16:07.100 structure just waiting there to be discovered 00:16:07.100 --> 00:16:09.600 by any intelligence smart enough, whether it's 00:16:09.600 --> 00:16:11.480 made of silicon or carbon? And the follow -on 00:16:11.480 --> 00:16:13.700 question, could this shared internal language 00:16:13.700 --> 00:16:16.159 of AI actually lead us to discover some universal 00:16:16.159 --> 00:16:19.169 truths? Truths about intelligence itself, both 00:16:19.169 --> 00:16:21.690 the artificial kind we're building and the biological 00:16:21.690 --> 00:16:23.990 kind we possess. It's definitely a question worth 00:16:23.990 --> 00:16:26.370 pondering as we all move into this future. We 00:16:26.370 --> 00:16:28.250 really hope you feel a little more informed after 00:16:28.250 --> 00:16:30.750 this and maybe, maybe a lot more curious. Thanks 00:16:30.750 --> 00:16:33.129 so much for joining us on the Deep Dive. Audioro 00:16:33.129 --> 00:16:33.330 music.