WEBVTT 00:00:00.000 --> 00:00:02.100 Imagine for a second that you have an assistant, 00:00:02.459 --> 00:00:05.839 a genuine genius of an assistant. They can quote 00:00:05.839 --> 00:00:09.259 Shakespeare perfectly, solve complex calculus 00:00:09.259 --> 00:00:11.199 and write poetry in three different languages. 00:00:11.380 --> 00:00:13.960 Sounds pretty ideal. Right. But if you ask them 00:00:13.960 --> 00:00:16.179 to transfer money to your savings account and 00:00:16.179 --> 00:00:18.699 you accidentally stutter or change one single 00:00:18.699 --> 00:00:21.940 word in your request, they might just burn the 00:00:21.940 --> 00:00:25.089 money instead. Yeah, that is the. Absolute paradox 00:00:25.089 --> 00:00:27.769 of where we are right now. We treat AI like this 00:00:27.769 --> 00:00:30.070 infallible genius, but would you really trust 00:00:30.070 --> 00:00:32.030 it to fly a plane if it gets completely confused 00:00:32.030 --> 00:00:35.729 by a typo? Most people probably wouldn't. Not 00:00:35.729 --> 00:00:38.789 yet, anyway. Welcome back to The Deep Dive. It 00:00:38.789 --> 00:00:42.829 is Sunday, March 1st, 2026. If you're listening 00:00:42.829 --> 00:00:44.469 to this, you're probably trying to make sense 00:00:44.469 --> 00:00:47.130 of the noise out there. And there's a lot of 00:00:47.130 --> 00:00:49.070 noise lately. It feels like the ground shifts 00:00:49.070 --> 00:00:52.429 every single day. Exactly. So let's just take 00:00:52.429 --> 00:00:55.070 a breath. I want to channel a bit of a calmer, 00:00:55.070 --> 00:00:58.109 more reflective vibe today. We aren't just looking 00:00:58.109 --> 00:01:00.689 at the latest chatbot release or some cool new 00:01:00.689 --> 00:01:02.450 image generator. We're looking at the plumbing. 00:01:02.590 --> 00:01:06.069 The plumbing, yes. The physical and the theoretical 00:01:06.069 --> 00:01:08.730 rails that are being laid down right now to support 00:01:08.730 --> 00:01:11.769 a fully automated world. The infrastructure of 00:01:11.769 --> 00:01:14.290 intelligent automation. It's the difference between 00:01:14.290 --> 00:01:16.989 staring at a shiny new car and actually looking 00:01:16.989 --> 00:01:21.010 at the highway it drives on. That highway is 00:01:21.010 --> 00:01:23.650 getting a massive upgrade. Precisely. So here 00:01:23.650 --> 00:01:26.189 is our roadmap for the next 15 minutes or so. 00:01:26.409 --> 00:01:28.870 First, we're going to unpack Nvidia's massive 00:01:28.870 --> 00:01:32.049 new bet on 6G networks. Which really proved the 00:01:32.049 --> 00:01:34.030 internet isn't for humans anymore. It really 00:01:34.030 --> 00:01:36.689 isn't. Then we'll look at a major breakthrough 00:01:36.689 --> 00:01:39.450 in AI video. One that claims to have finally 00:01:39.450 --> 00:01:41.930 killed the uncanny valley. That one changes the 00:01:41.930 --> 00:01:44.730 entire workflow for creatives. Huge shift there. 00:01:45.200 --> 00:01:46.799 After that, we have to talk about the money. 00:01:47.079 --> 00:01:49.599 We'll look at Block, Jack Dorsey, and the $4 00:01:49.599 --> 00:01:52.680 .5 trillion labor shift happening right under 00:01:52.680 --> 00:01:55.280 our noses. The economic reality check. Right. 00:01:55.340 --> 00:01:57.019 And finally, we're going to end with a different 00:01:57.019 --> 00:01:59.400 kind of reality check. A new study from Princeton 00:01:59.400 --> 00:02:01.780 University about agent reliability that might 00:02:01.780 --> 00:02:04.019 honestly make you think twice about automating 00:02:04.019 --> 00:02:06.980 your life. That Princeton study, it's sobering. 00:02:07.060 --> 00:02:10.020 So let's get into it. Segment one, the nervous 00:02:10.020 --> 00:02:13.330 system. I saw this news about Nvidia forming 00:02:13.330 --> 00:02:16.270 a major alliance with Nokia, SoftBank, and T 00:02:16.270 --> 00:02:19.830 -Mobile. And my first cynical thought was just... 00:02:20.090 --> 00:02:22.689 Great, slightly faster downloads for my phone. 00:02:22.889 --> 00:02:24.990 But I dug into the specs. That's not what this 00:02:24.990 --> 00:02:26.889 is at all, is it? No, not at all. If you think 00:02:26.889 --> 00:02:29.830 6G is just 5G but faster, you're kind of missing 00:02:29.830 --> 00:02:31.889 the entire point. Break it down for us. See, 00:02:31.930 --> 00:02:34.389 5G, the network we've been using, was designed 00:02:34.389 --> 00:02:37.009 for us. It was built for smartphones, for streaming 00:02:37.009 --> 00:02:39.110 AK video, humans talking to humans. Right, it 00:02:39.110 --> 00:02:41.250 was built for TikTok, Zoom calls, that sort of 00:02:41.250 --> 00:02:43.930 thing. Exactly. But NVIDIA and these telecom 00:02:43.930 --> 00:02:46.680 giants... are looking at the next wave. And the 00:02:46.680 --> 00:02:49.379 next wave isn't you or me. It's machines talking 00:02:49.379 --> 00:02:52.060 to machines. We are talking about a network designed 00:02:52.060 --> 00:02:54.659 specifically for robots, autonomous vehicles, 00:02:55.000 --> 00:02:57.360 smart factories, and intelligent cities. So the 00:02:57.360 --> 00:02:59.340 internet is essentially being gentrified for 00:02:59.340 --> 00:03:03.080 robots. In a way, yes. Think about the physical 00:03:03.080 --> 00:03:06.159 limitations of 5G. It has latency. That tiny 00:03:06.159 --> 00:03:08.639 delay between sending a signal and getting a 00:03:08.639 --> 00:03:11.919 response. For a human, 20 milliseconds is totally 00:03:11.919 --> 00:03:14.199 unnoticeable. You literally don't feel it. But 00:03:14.199 --> 00:03:16.900 for a swarm of drones flying in tight formation. 00:03:17.199 --> 00:03:19.580 Or a self -driving car reacting to a kid running 00:03:19.580 --> 00:03:22.219 into the street. Exactly. 20 milliseconds is 00:03:22.219 --> 00:03:24.599 an eternity for a machine. It's the difference 00:03:24.599 --> 00:03:26.599 between a near miss and a complete catastrophe. 00:03:27.080 --> 00:03:29.719 So how is 6G different physically? Is it just 00:03:29.719 --> 00:03:32.240 throwing up more cell towers, denser signals? 00:03:32.620 --> 00:03:34.439 It's actually a move away from physical hardware. 00:03:34.599 --> 00:03:36.659 That's the critical piece. NVIDIA is pushing 00:03:36.659 --> 00:03:39.099 for software -defined AI networks. Let's unpack 00:03:39.099 --> 00:03:42.289 that jargon. Software -defined AI network. That's 00:03:42.289 --> 00:03:45.810 a network run by adaptable code, not hardwired 00:03:45.810 --> 00:03:49.090 physical boxes. Spot on. Traditionally, telecom 00:03:49.090 --> 00:03:52.069 gear is very rigid. You build a metal box, stick 00:03:52.069 --> 00:03:54.330 it on a tower, and it does one thing. It routes 00:03:54.330 --> 00:03:56.909 radio waves in a specific way. Okay. This new 00:03:56.909 --> 00:03:59.030 vision turns the network into a programmable 00:03:59.030 --> 00:04:01.770 intelligence layer. Imagine the airways themselves 00:04:01.770 --> 00:04:04.129 are actually intelligent. Whoa, just imagine 00:04:04.129 --> 00:04:06.729 scaling that to a whole city. The invisible traffic 00:04:06.729 --> 00:04:09.210 of data constantly reorganizing itself around 00:04:09.210 --> 00:04:12.569 us, totally unseen. Billions of requests. It's 00:04:12.569 --> 00:04:16.589 incredible. The AI dynamically routes radio traffic 00:04:16.589 --> 00:04:19.910 in real time. It adapts. It's general purpose 00:04:19.910 --> 00:04:23.170 computing replacing lockdown telecom gear. So 00:04:23.170 --> 00:04:25.250 the network isn't just a dumb pipe anymore. It's 00:04:25.250 --> 00:04:28.290 a brain. It's a nervous system. Consider a factory 00:04:28.290 --> 00:04:31.829 floor right now in 2026. You've got robots moving 00:04:31.829 --> 00:04:35.350 heavy materials, sensors monitoring heat, AI 00:04:35.350 --> 00:04:37.730 cameras checking for defects. And they all need 00:04:37.730 --> 00:04:40.990 flawless wireless connectivity. Right. If that 00:04:40.990 --> 00:04:43.850 network is static and suddenly 10 robots need 00:04:43.850 --> 00:04:45.790 to upload high -res video all at once because 00:04:45.790 --> 00:04:48.209 of an error, the system clogs. And the robots 00:04:48.209 --> 00:04:51.050 stop. Or they crash into each other. But if the 00:04:51.050 --> 00:04:53.310 network is AI -driven, it predicts that surge. 00:04:53.829 --> 00:04:55.850 It says, hey, I see a bottleneck forming in sector 00:04:55.850 --> 00:04:58.970 four and it reroutes bandwidth instantly. Before 00:04:58.970 --> 00:05:01.350 the human operators even know there was a problem. 00:05:01.670 --> 00:05:04.250 Proactive, not reactive. Exactly. And this explains 00:05:04.250 --> 00:05:06.389 why NVIDIA is leading the charge here. They don't 00:05:06.389 --> 00:05:08.449 care about selling you a monthly phone plan. 00:05:08.629 --> 00:05:10.610 They care about selling the chips running that 00:05:10.610 --> 00:05:12.889 brain. Because if the network isn't AI ready, 00:05:13.069 --> 00:05:17.850 physical AI adoption just stalls. 100%. You can't 00:05:17.850 --> 00:05:20.269 sell millions of autonomous robots if they can't 00:05:20.269 --> 00:05:23.199 talk to each other reliably. NVIDIA is essentially 00:05:23.199 --> 00:05:26.079 making sure there is a paved road for the futuristic 00:05:26.079 --> 00:05:29.480 cars they want to sell. Beat. So is this just 00:05:29.480 --> 00:05:32.699 faster Wi -Fi or something fundamentally different? 00:05:32.839 --> 00:05:36.040 It's not speed. It's the network becoming a living, 00:05:36.100 --> 00:05:39.519 thinking software layer. Okay, so we have the 00:05:39.519 --> 00:05:42.240 physical rails. The nervous system is being built. 00:05:42.399 --> 00:05:44.920 Now let's talk about the content flowing through 00:05:44.920 --> 00:05:48.879 those networks. Segment two. the creative breakthrough. 00:05:49.180 --> 00:05:50.779 Oh, this is a fun one. This is the stuff you 00:05:50.779 --> 00:05:52.759 can actually see on your screens right now. We've 00:05:52.759 --> 00:05:55.639 been talking about AI video for years, and visually 00:05:55.639 --> 00:05:58.040 it's been amazing for a while, but there's always 00:05:58.040 --> 00:06:01.259 that uncanny valley. The eyes that look a bit 00:06:01.259 --> 00:06:04.220 dead, with the mouth moving like a weird ventriloquist 00:06:04.220 --> 00:06:06.600 dummy. Right, it always broke the illusion. You'd 00:06:06.600 --> 00:06:09.180 see a stunning cinematic landscape, but the moment 00:06:09.180 --> 00:06:11.439 a character opened their mouth, you instantly 00:06:11.439 --> 00:06:14.220 knew it was fake. It was soul -crushing for storytellers. 00:06:14.319 --> 00:06:16.339 It just pulled you right out of the narrative. 00:06:16.779 --> 00:06:18.879 But according to this new guide on the missing 00:06:18.879 --> 00:06:22.680 piece of AI cinema, that era is effectively over. 00:06:23.079 --> 00:06:26.199 We've solved the lip sync and consistency issues. 00:06:26.480 --> 00:06:28.899 It really looks like we have. The workflow has 00:06:28.899 --> 00:06:31.000 completely shifted. We aren't just generating 00:06:31.000 --> 00:06:33.480 cool five -second clips anymore. We're seeing 00:06:33.480 --> 00:06:37.339 full -blown performances with custom voices and 00:06:37.339 --> 00:06:40.600 perfect lip sync. the guide mentioned a few specific 00:06:40.600 --> 00:06:43.439 techniques that caught my eye they solve very 00:06:43.439 --> 00:06:46.399 specific engineering problems one was called 00:06:46.399 --> 00:06:48.899 the prompt pack yeah that's crucial for temporal 00:06:48.899 --> 00:06:51.300 consistency and just to define that temporal 00:06:51.300 --> 00:06:53.879 consistency means keeping a face looking exactly 00:06:53.879 --> 00:06:56.759 the same frame to frame exactly One of the biggest 00:06:56.759 --> 00:06:59.220 issues with AI video historically is that the 00:06:59.220 --> 00:07:02.600 model forgets what things look like. It's hallucinating 00:07:02.600 --> 00:07:05.240 24 images every single second. So you generate 00:07:05.240 --> 00:07:07.500 a character in one scene, and in the next cut, 00:07:07.620 --> 00:07:09.899 their nose is slightly wider or their jawline 00:07:09.899 --> 00:07:12.759 changes. Right. The prompt pack is essentially 00:07:12.759 --> 00:07:15.139 a rigorous framework. It acts like an anchor. 00:07:15.199 --> 00:07:18.019 It locks the identity data so the AI doesn't 00:07:18.019 --> 00:07:21.259 drift. Like applying perfect digital makeup consistency. 00:07:21.660 --> 00:07:25.100 It forces the model to reference the exact same 00:07:25.100 --> 00:07:27.899 facial structure data in every single frame. 00:07:28.019 --> 00:07:29.939 So the character feels like the same person, 00:07:30.040 --> 00:07:32.379 not just a close cousin of the person. And then 00:07:32.379 --> 00:07:34.399 there's the desert survival case study. That 00:07:34.399 --> 00:07:36.939 sounded pretty intense. It's a great framework 00:07:36.939 --> 00:07:39.759 for handling dialogue. Two -person dialogue in 00:07:39.759 --> 00:07:42.379 AI video used to be a total nightmare. It always 00:07:42.379 --> 00:07:44.980 looked like two stiff robots patiently waiting 00:07:44.980 --> 00:07:48.160 for their turn to speak. Very sequential. Human 00:07:48.160 --> 00:07:50.680 beings don't talk like that. This case study 00:07:50.680 --> 00:07:52.939 breaks down how to actually overlap reactions. 00:07:53.379 --> 00:08:00.399 Yes. We nod, we grimace, we look away, we take 00:08:00.399 --> 00:08:03.560 a breath. This new workflow teaches the AI to 00:08:03.560 --> 00:08:06.300 cut between faces and layer those reactions so 00:08:06.300 --> 00:08:08.759 it feels organic. It creates actual chemistry 00:08:08.759 --> 00:08:11.240 between two generated beings. But my absolute 00:08:11.240 --> 00:08:13.620 favorite part was the soundtrack glue. Ah, the 00:08:13.620 --> 00:08:15.720 soundtrack glue. It's brilliant because it's 00:08:15.720 --> 00:08:17.939 not an engineering fix at all. It's a psychological 00:08:17.939 --> 00:08:20.540 hack. Tell them that works. So even with perfect 00:08:20.540 --> 00:08:22.740 lip sync, you sometimes get these microscopic 00:08:22.740 --> 00:08:25.720 artifacts, tiny little glitches in the audio 00:08:25.720 --> 00:08:28.500 or video where the AI just hiccups for a millisecond. 00:08:28.519 --> 00:08:31.220 Sure. Soundtrack glue is the technique of using 00:08:31.220 --> 00:08:35.220 subtle, melancholic musical underscores or heavy 00:08:35.220 --> 00:08:38.700 ambient noise to mask those physical seams. You're 00:08:38.700 --> 00:08:41.539 hacking human emotion to cover technical flaws. 00:08:41.960 --> 00:08:45.600 It is pure psychoacoustics. If you play a sad, 00:08:45.720 --> 00:08:48.539 swelling cello line, the viewer's brain focuses 00:08:48.539 --> 00:08:50.799 entirely on the emotion of the scene. Your brain 00:08:50.799 --> 00:08:53.440 is busy processing the sadness. Right. So it 00:08:53.440 --> 00:08:55.279 completely ignores the fact that the character's 00:08:55.279 --> 00:08:58.639 upper lip moved weirdly for half a frame. It 00:08:58.639 --> 00:09:01.159 sells the cinematic vibe by distracting the logical 00:09:01.159 --> 00:09:03.039 part of your brain. It's fascinating that we 00:09:03.039 --> 00:09:05.360 are solving highly technical problems with artistic 00:09:05.360 --> 00:09:08.299 sleight of hand. Debeat. Does this mean the era 00:09:08.299 --> 00:09:11.120 of the glitchy AI video is officially over? Yes. 00:09:11.240 --> 00:09:13.820 We've moved from cool clips to actual sustained 00:09:13.820 --> 00:09:16.460 cinematic performance. Which is incredible for 00:09:16.460 --> 00:09:19.080 entertainment, but maybe less great for the people 00:09:19.080 --> 00:09:20.720 who used to make that entertainment or really 00:09:20.720 --> 00:09:23.120 do any other kind of work. This brings us to 00:09:23.120 --> 00:09:25.679 segment three, the economic and social impact. 00:09:25.960 --> 00:09:28.139 The numbers this week are heavy. They really 00:09:28.139 --> 00:09:31.000 are. Looking at the highlights from Today and 00:09:31.000 --> 00:09:33.220 AI, the headline that jumped out at me was Jack 00:09:33.220 --> 00:09:36.820 Dorsey and Block. Yeah, Block, the company behind 00:09:36.820 --> 00:09:39.740 Square and Cash App. Dorsey just replaced over 00:09:39.740 --> 00:09:43.120 4 ,000 workers with AI. That is a massive chunk 00:09:43.120 --> 00:09:45.720 of a real -world workforce. And the market reaction. 00:09:46.399 --> 00:09:50.179 The stock jumped 24%. That is the part that sits 00:09:50.179 --> 00:09:52.240 uncomfortably, doesn't it? It highlights the 00:09:52.240 --> 00:09:55.019 ruthless efficiency of the market. Investors 00:09:55.019 --> 00:09:57.299 see workforce replacement. They translate it 00:09:57.299 --> 00:10:00.279 directly to lower costs and higher margins. And 00:10:00.279 --> 00:10:02.860 they buy the stock. But for the workforce, it's 00:10:02.860 --> 00:10:06.620 a blaring siren. Dorsey. It was very clear here. 00:10:06.720 --> 00:10:09.019 He said this agentic workflow could happen across 00:10:09.019 --> 00:10:11.559 most companies within a single year. Agentic 00:10:11.559 --> 00:10:13.220 workflow. That's a big term. Let's define it. 00:10:13.299 --> 00:10:15.940 It means AI planning and doing the job itself, 00:10:16.120 --> 00:10:18.539 not just helping a human do it faster. Exactly. 00:10:18.679 --> 00:10:20.580 It's the difference between a carpenter using 00:10:20.580 --> 00:10:23.259 a power drill versus a robot actually building 00:10:23.259 --> 00:10:25.179 the cabinet while you watch. The human is just 00:10:25.179 --> 00:10:28.059 the architect now or the observer. And this isn't 00:10:28.059 --> 00:10:30.840 isolated to tech companies. The study noted that 00:10:30.840 --> 00:10:34.919 93 % of U .S. jobs are exposed to AI. That represents 00:10:34.919 --> 00:10:39.259 $4 .5 trillion in labor value. Exposed is such 00:10:39.259 --> 00:10:42.000 a polite economic word. It means the job can 00:10:42.000 --> 00:10:45.740 be done, at least partially. by software. We're 00:10:45.740 --> 00:10:49.000 watching a $4 .5 trillion reorganization of the 00:10:49.000 --> 00:10:52.019 global economy in real time. We've never seen 00:10:52.019 --> 00:10:54.360 a shift happen this fast. The Industrial Revolution 00:10:54.360 --> 00:10:57.179 took decades. This is happening in business quarters. 00:10:57.379 --> 00:10:59.759 It's creating this very weird cultural atmosphere. 00:11:00.120 --> 00:11:02.639 On one hand, people are losing their livelihoods. 00:11:02.639 --> 00:11:04.820 On the other, you have this bizarre simulation 00:11:04.820 --> 00:11:07.519 theory trend going viral. Oh man, the CEO names 00:11:07.519 --> 00:11:10.850 video. I saw that. 2 .3 million views. People 00:11:10.850 --> 00:11:12.909 are genuinely debating if we are living in a 00:11:12.909 --> 00:11:14.750 simulation because the characters running the 00:11:14.750 --> 00:11:17.470 AI revolution have suspiciously thematic names. 00:11:17.629 --> 00:11:19.450 It's funny, but it really speaks to the underlying 00:11:19.450 --> 00:11:22.250 anxiety. When the world changes this fast and 00:11:22.250 --> 00:11:24.250 4 ,000 people lose their jobs while the stock 00:11:24.250 --> 00:11:26.809 market cheers reality starts to feel a bit scripted. 00:11:26.830 --> 00:11:28.549 Like we were just watching a movie we didn't 00:11:28.549 --> 00:11:31.250 write. People look for patterns to make sense 00:11:31.250 --> 00:11:34.169 of the chaos. Speaking of scripted chaos, let's 00:11:34.169 --> 00:11:37.049 touch on the geopolitical side. We saw a headline 00:11:37.049 --> 00:11:40.409 that Trump banned Anthropx Claude. A swift political 00:11:40.409 --> 00:11:43.990 ban. And yet, mere hours later, the U .S. military 00:11:43.990 --> 00:11:46.789 apparently used that exact same technology for 00:11:46.789 --> 00:11:49.389 strike planning in Iran. It's a very stark reminder. 00:11:49.649 --> 00:11:53.529 Policy moves slow. Utility moves fast. Even if 00:11:53.529 --> 00:11:56.450 a government publicly bans a tool, if that tool 00:11:56.450 --> 00:11:59.549 offers a massive strategic advantage in warfare 00:11:59.549 --> 00:12:02.570 or logistics, It's going to get used. AI is just 00:12:02.570 --> 00:12:05.070 too deeply integrated into the fabric of defense 00:12:05.070 --> 00:12:07.629 to be switched off by a press release. It's like 00:12:07.629 --> 00:12:09.490 trying to ban electricity in the military. You 00:12:09.490 --> 00:12:11.470 just can't do it. And it's everywhere, from the 00:12:11.470 --> 00:12:14.509 war room to the dining room. Slang AI just raised 00:12:14.509 --> 00:12:17.730 $68 million for restaurant voice AI. That's the 00:12:17.730 --> 00:12:20.129 consumer side of the labor shift. You call a 00:12:20.129 --> 00:12:22.049 restaurant to book a table. You aren't talking 00:12:22.049 --> 00:12:23.950 to a hostess anymore. You're talking to slang. 00:12:24.480 --> 00:12:27.399 It handles calls 247 and it's driving double 00:12:27.399 --> 00:12:30.019 the reservations. It's relentless. It never sleeps, 00:12:30.139 --> 00:12:32.580 never takes a smoke break, never has a bad day. 00:12:32.740 --> 00:12:35.320 It feels like the efficiency gains are finally 00:12:35.320 --> 00:12:37.340 hitting the Wall Street bottom line, aren't they? 00:12:37.659 --> 00:12:40.120 Absolutely. The exposure is turning into replacement 00:12:40.120 --> 00:12:42.480 and the market is rewarding it. Well, to make 00:12:42.480 --> 00:12:45.159 those gains, we need better tools. Let's do a 00:12:45.159 --> 00:12:46.960 quick run through of what launched this week. 00:12:47.179 --> 00:12:50.299 Segment four, the tool belt, rapid fire. What's 00:12:50.299 --> 00:12:52.590 actually making life easier? A few key things 00:12:52.590 --> 00:12:55.710 focused on speed and connection. First up is 00:12:55.710 --> 00:12:59.149 Claude Import Memory. This is a massive win for 00:12:59.149 --> 00:13:01.289 interoperability. What does that actually mean 00:13:01.289 --> 00:13:03.950 for a normal user? It means the walled gardens 00:13:03.950 --> 00:13:06.009 are finally coming down. You can switch from 00:13:06.009 --> 00:13:08.950 ChatGPT over to Claude, and it just imports your 00:13:08.950 --> 00:13:10.990 memory. It picks up exactly where you left off. 00:13:11.149 --> 00:13:13.049 So I don't have to re -explain my entire life 00:13:13.049 --> 00:13:16.129 story and workflow to a new bot. Exactly. No 00:13:16.129 --> 00:13:18.129 more starting from scratch just because you changed 00:13:18.129 --> 00:13:22.100 brands. Yeah. Next up is Notra. It connects GitHub 00:13:22.100 --> 00:13:24.879 and Slack to automatically write changelogs and 00:13:24.879 --> 00:13:27.960 blog posts. That is an absolute dream for developers 00:13:27.960 --> 00:13:30.700 who hate writing documentation, which is all 00:13:30.700 --> 00:13:33.340 of them. True. But the most technical update 00:13:33.340 --> 00:13:36.980 is OpenAI's new WebSocket mode. It cuts latency 00:13:36.980 --> 00:13:40.240 by 40%. There's that word again, latency. Why 00:13:40.240 --> 00:13:43.000 does a 40 % cut matter here? It's vital for voice 00:13:43.000 --> 00:13:45.960 apps. If you're building a real -time conversational 00:13:45.960 --> 00:13:49.200 agent, 40 % is the difference. And that is a 00:13:49.200 --> 00:13:52.679 massive trap. They tested 14 leading models across 00:13:52.679 --> 00:13:55.460 500 different runs. And they didn't test for 00:13:55.460 --> 00:13:58.299 IQ or creativity. They tested for strict engineering 00:13:58.299 --> 00:14:01.259 reliability. They used aviation principles, right? 00:14:01.379 --> 00:14:03.820 Exactly. Think about it. If you build a passenger 00:14:03.820 --> 00:14:06.299 plane, you don't care if it flies mostly well. 00:14:06.419 --> 00:14:09.059 You care if it flies consistently every single 00:14:09.059 --> 00:14:11.840 time, regardless of the weather. So they defined 00:14:11.840 --> 00:14:15.639 four pillars. Consistency, robustness, predictability, 00:14:15.820 --> 00:14:18.559 and safety. What did they find? They found that 00:14:18.559 --> 00:14:20.919 predictability is massively broken across the 00:14:20.919 --> 00:14:23.139 board. What does that actually mean in practice 00:14:23.139 --> 00:14:26.419 for a user? It means the AI literally does not 00:14:26.419 --> 00:14:29.559 know when it's confused. It sounds just as confident 00:14:29.559 --> 00:14:31.460 when it's completely hallucinating as it does 00:14:31.460 --> 00:14:33.019 when it's telling you the absolute truth. It 00:14:33.019 --> 00:14:35.600 lacks epistemic humility. It can't say, I don't 00:14:35.600 --> 00:14:38.960 know. And that is incredibly dangerous. The study 00:14:38.960 --> 00:14:40.840 found that if you change a single word in your 00:14:40.840 --> 00:14:44.039 prompt, just one synonym, the model might give 00:14:44.039 --> 00:14:45.740 you a completely different answer or just totally 00:14:45.740 --> 00:14:47.659 break down. This comes back to the difference 00:14:47.659 --> 00:14:50.419 between probabilistic and deterministic systems. 00:14:50.700 --> 00:14:53.419 Right. A calculator is deterministic. 2 plus 00:14:53.419 --> 00:14:55.679 2 is always 4. If you enter it a million times, 00:14:55.820 --> 00:15:00.059 it's always 4. But an AI is probabilistic. It's 00:15:00.059 --> 00:15:02.740 guessing the next word based on statistical likelihood. 00:15:02.960 --> 00:15:06.080 It thinks 2 plus 2 is probably 4. But if you 00:15:06.080 --> 00:15:08.960 ask it in a slightly weird way, or if the internal 00:15:08.960 --> 00:15:11.480 temperature setting is off, it might guess 5. 00:15:11.840 --> 00:15:14.779 It's rolling dice. It's not following a rule 00:15:14.779 --> 00:15:17.360 book. See, that's the terrifying part. We are 00:15:17.360 --> 00:15:20.000 integrating this stuff into banks, into hospitals, 00:15:20.179 --> 00:15:22.200 into the military, as we just discussed. And 00:15:22.200 --> 00:15:25.120 you simply cannot run a bank on software that 00:15:25.120 --> 00:15:28.879 completely breaks because a customer used a synonym. 00:15:29.179 --> 00:15:31.519 The industry has been obsessed with larger models, 00:15:31.679 --> 00:15:33.919 making the brain bigger. But Princeton is saying 00:15:33.919 --> 00:15:36.120 it doesn't matter how big the brain is if the 00:15:36.120 --> 00:15:38.840 behavior is fundamentally erratic. A genius who 00:15:38.840 --> 00:15:40.980 is drunk is still dangerous behind the wheel. 00:15:41.240 --> 00:15:43.419 It really makes me think about my own daily usage. 00:15:43.720 --> 00:15:45.799 Yeah, I mean, I'll be totally vulnerable here. 00:15:45.840 --> 00:15:48.080 I still wrestle with prompt drift myself constantly. 00:15:48.340 --> 00:15:51.139 Oh yeah, I'll spend a whole day setting up a 00:15:51.139 --> 00:15:54.899 complex multi -step workflow, testing it, getting 00:15:54.899 --> 00:15:57.299 it totally perfect. I go to sleep feeling like 00:15:57.299 --> 00:15:59.279 an absolute genius. I wake up the next morning, 00:15:59.399 --> 00:16:02.259 run the exact same prompt, and the model has 00:16:02.259 --> 00:16:04.799 inexplicably decided to answer differently. It 00:16:04.799 --> 00:16:06.820 breaks the entire chain. It breaks everything. 00:16:06.919 --> 00:16:09.570 It feels like it's gaslighting me. But that's 00:16:09.570 --> 00:16:11.950 the robustness pillar Princeton is talking about. 00:16:12.070 --> 00:16:14.909 We are building these incredible magical systems, 00:16:15.049 --> 00:16:18.690 but they are so inherently brittle. We are building 00:16:18.690 --> 00:16:21.049 them into critical societal infrastructure, but 00:16:21.049 --> 00:16:23.830 they can shatter like glass if the prompt is 00:16:23.830 --> 00:16:27.269 just slightly wrong. So are we building skyscrapers 00:16:27.269 --> 00:16:29.470 on shaky ground? We're building powerful engines, 00:16:29.570 --> 00:16:31.710 but haven't invented the brakes or the steering 00:16:31.710 --> 00:16:34.149 wheel yet. That is not very comforting, but it 00:16:34.149 --> 00:16:36.610 is the reality we need to understand. Sponsor 00:16:36.610 --> 00:16:39.370 Galvin. All right, let's bring this all together. 00:16:39.570 --> 00:16:42.110 Big idea recap. We've covered a lot of ground 00:16:42.110 --> 00:16:44.509 today. We really have. We started with the physical 00:16:44.509 --> 00:16:47.669 rails. NVIDIA and the telecom giants building 00:16:47.669 --> 00:16:50.470 a 6G network that is essentially a new nervous 00:16:50.470 --> 00:16:53.129 system for machines, not humans. A network that 00:16:53.129 --> 00:16:55.490 dynamically thinks and routes traffic for robots. 00:16:55.950 --> 00:16:58.470 Then we looked at the creative side. The uncanny 00:16:58.470 --> 00:17:01.899 valley is dead. We can now generate video that 00:17:01.899 --> 00:17:05.000 is indistinguishable from reality, using clever 00:17:05.000 --> 00:17:07.579 psychological hacks like soundtrack glue to hide 00:17:07.579 --> 00:17:11.299 the seams. But that immense technological power 00:17:11.299 --> 00:17:14.259 is causing a massive economic reorganization. 00:17:14.440 --> 00:17:17.599 We saw Block replacing thousands of workers with 00:17:17.599 --> 00:17:20.799 agentic workflows and the stock market enthusiastically 00:17:20.799 --> 00:17:24.619 cheering for it. The $4 .5 trillion labor shift 00:17:24.619 --> 00:17:28.200 is officially here. And finally, the cold water. 00:17:28.589 --> 00:17:31.049 A Princeton study. We have all this power, all 00:17:31.049 --> 00:17:32.809 this intelligence, but we lack the fundamental 00:17:32.809 --> 00:17:34.849 reliability to actually trust it with the most 00:17:34.849 --> 00:17:37.170 critical parts of our lives. We have high accuracy, 00:17:37.309 --> 00:17:39.930 but very low predictability. It's a high speed 00:17:39.930 --> 00:17:41.710 train, but we aren't quite sure if the tracks 00:17:41.710 --> 00:17:43.690 ahead are actually finished. So I want to leave 00:17:43.690 --> 00:17:45.970 you with a thought to mull over this week. We've 00:17:45.970 --> 00:17:48.230 talked about 6G networks creating a world where 00:17:48.230 --> 00:17:51.250 machines talk directly to machines. We've talked 00:17:51.250 --> 00:17:54.210 about AI replacing human labor at massive companies. 00:17:54.329 --> 00:17:56.470 We've talked about AI flawlessly generating our 00:17:56.470 --> 00:17:59.359 media. If the AI... creates the network, replaces 00:17:59.359 --> 00:18:01.700 the workers, and generates the media that we 00:18:01.700 --> 00:18:04.740 consume. But as Princeton proved, it lacks the 00:18:04.740 --> 00:18:07.240 self -awareness to know when it is wrong, who 00:18:07.240 --> 00:18:10.160 is actually steering the ship. Are we the captains 00:18:10.160 --> 00:18:12.259 here or are we just passengers on a very fast, 00:18:12.519 --> 00:18:14.839 very erratic ride? That is the defining question 00:18:14.839 --> 00:18:17.079 of the decade. If you enjoyed this deep dive, 00:18:17.259 --> 00:18:19.500 hit follow. We'll see you in the next one. Out 00:18:19.500 --> 00:18:20.000 your own music.