WEBVTT 00:00:00.000 --> 00:00:02.700 Think about the last time you spoke to an AI. 00:00:02.960 --> 00:00:05.519 Yeah. It is usually a pretty jarring feeling. 00:00:05.700 --> 00:00:07.799 You try to interrupt the machine. An awkward 00:00:07.799 --> 00:00:10.779 silence stretches out. The system clunks through 00:00:10.779 --> 00:00:13.820 a messy reset. It just feels completely unnatural. 00:00:14.080 --> 00:00:16.160 It really does. It breaks the illusion immediately. 00:00:16.239 --> 00:00:19.059 Now, change that picture entirely. What if your 00:00:19.059 --> 00:00:21.579 AI could, you know, see you frown? What if it 00:00:21.579 --> 00:00:23.339 stopped talking mid -sentence and asked, what 00:00:23.339 --> 00:00:25.399 is wrong? That completely changes the paradigm. 00:00:25.579 --> 00:00:28.359 It does. Welcome to this custom deep dive. We 00:00:28.359 --> 00:00:30.980 are very glad you are here with us. Today, we 00:00:30.980 --> 00:00:33.159 are exploring a massive technological shift. 00:00:33.609 --> 00:00:36.270 We are tracking the move from clunky AI tools 00:00:36.270 --> 00:00:40.810 to fluid proactive partners. It is a big one. 00:00:40.990 --> 00:00:43.689 First, we will unpack Google's brand new Android 00:00:43.689 --> 00:00:46.609 updates. Then we are going to examine the staggering 00:00:46.609 --> 00:00:48.729 capital engines operating behind the scenes. 00:00:48.829 --> 00:00:52.109 We are talking about OpenAI, SoftBank, and Anthropic. 00:00:52.229 --> 00:00:54.509 The scale of the money is just wild. It truly 00:00:54.509 --> 00:00:57.109 is. Finally, we will look at a major breakthrough 00:00:57.109 --> 00:01:00.149 from Miramarati's Thinking Machines Lab. It is 00:01:00.149 --> 00:01:02.490 a fundamental shift that promises to end the 00:01:02.490 --> 00:01:05.620 awkward AI silence forever. It is a profound 00:01:05.620 --> 00:01:09.219 transition. We are moving past the novelty phase 00:01:09.219 --> 00:01:11.900 of AI. Yeah. The technology is weaving itself 00:01:11.900 --> 00:01:14.799 into our daily routines. It is basically becoming 00:01:14.799 --> 00:01:17.620 an invisible layer of reality. That is exactly 00:01:17.620 --> 00:01:19.620 what struck me about Google's latest announcements. 00:01:19.900 --> 00:01:22.099 They just wrapped up the Android show. It was 00:01:22.099 --> 00:01:24.620 their special IO edition. Right. And honestly, 00:01:24.799 --> 00:01:27.219 the whole ecosystem is shifting. We are looking 00:01:27.219 --> 00:01:29.920 at a radical redesign of how we interact with 00:01:29.920 --> 00:01:32.120 software. We are seeing a definitive move away 00:01:32.120 --> 00:01:35.420 from reactive tools. Consumer tech is becoming 00:01:35.420 --> 00:01:38.459 a proactive assistant. It anticipates what you 00:01:38.459 --> 00:01:40.799 need. It doesn't just wait around for a type 00:01:40.799 --> 00:01:43.500 command anymore. Right. And the clearest example 00:01:43.500 --> 00:01:45.760 of this is the Google book. They announced this 00:01:45.760 --> 00:01:48.560 entirely new type of computer. It is built from 00:01:48.560 --> 00:01:50.319 the ground up for Gemini. The operating system 00:01:50.319 --> 00:01:53.239 architecture is basically inverted. How so? Well, 00:01:53.280 --> 00:01:55.579 traditionally, the OS manages files and apps. 00:01:55.719 --> 00:01:59.019 Here, the OS centers entirely on the AI model. 00:01:59.180 --> 00:02:02.299 The AI is the kernel. Wow. They showcased a feature 00:02:02.299 --> 00:02:04.519 called the Magic Pointer too. It is essentially 00:02:04.519 --> 00:02:07.260 an AI -powered cursor. Yeah, that was fascinating. 00:02:07.439 --> 00:02:09.240 What is really interesting is how it bridges 00:02:09.240 --> 00:02:12.520 devices. You can use all your mobile apps on 00:02:12.520 --> 00:02:15.719 the big screen seamlessly. That removes a massive 00:02:15.719 --> 00:02:19.550 layer of friction. And the cursor itself... actually 00:02:19.550 --> 00:02:22.009 understands context. Right. It uses computer 00:02:22.009 --> 00:02:25.030 vision. Exactly. It knows what you are hovering 00:02:25.030 --> 00:02:28.030 over. If you hover over an address, it predicts 00:02:28.030 --> 00:02:30.569 you want a map. It anticipates your next action 00:02:30.569 --> 00:02:32.729 before you click. Okay, but the vibe -coded widgets 00:02:32.729 --> 00:02:35.069 feature is what actually stopped me in my tracks. 00:02:35.289 --> 00:02:37.610 Oh, man. You just describe a widget you want, 00:02:37.689 --> 00:02:39.490 maybe something incredibly niche that doesn't 00:02:39.490 --> 00:02:42.090 exist yet, and Android just builds it for you 00:02:42.090 --> 00:02:45.090 on the spot. Oh, it is wild. It completely flips 00:02:45.090 --> 00:02:47.469 how we interact with our phone's interface. You 00:02:47.469 --> 00:02:49.139 don't search an app store anymore? Nope. You 00:02:49.139 --> 00:02:51.800 just articulate a desire. The operating system 00:02:51.800 --> 00:02:54.900 handles the creation instantly. It is like magic. 00:02:55.180 --> 00:02:57.280 Under the hood, a lightweight language model 00:02:57.280 --> 00:03:00.539 writes the code. It compiles the UI components 00:03:00.539 --> 00:03:03.280 in real time. But wait, let's unpack that. Isn't 00:03:03.280 --> 00:03:05.520 that a massive threat to developers? Oh, absolutely. 00:03:05.780 --> 00:03:08.060 If the OS dynamically generates custom widgets, 00:03:08.379 --> 00:03:11.060 nobody is buying third -party apps for those 00:03:11.060 --> 00:03:13.900 microtasks anymore. Right? And that is the hidden 00:03:13.900 --> 00:03:16.919 disruption here. It democratizes software creation 00:03:16.919 --> 00:03:20.259 for the user. but it absolutely cannibalizes 00:03:20.259 --> 00:03:23.979 the lower tier of the developer ecosystem. The 00:03:23.979 --> 00:03:26.400 middleman is eliminated. That makes sense. We 00:03:26.400 --> 00:03:29.000 are also seeing major upgrades to Gemini intelligence. 00:03:29.789 --> 00:03:32.030 It is actually taking real -world actions now. 00:03:32.150 --> 00:03:34.990 Yes. And this is the shift from a passive oracle 00:03:34.990 --> 00:03:38.270 to an active agent. Right. It moves beyond generating 00:03:38.270 --> 00:03:42.229 text into executing multi -step workflows. The 00:03:42.229 --> 00:03:44.169 example they showed was brilliant. You take a 00:03:44.169 --> 00:03:46.169 photo of a concert flyer on a telephone pole. 00:03:46.409 --> 00:03:49.669 You simply tell Gemini to book a hotel. It goes 00:03:49.669 --> 00:03:51.990 out, navigates the travel sites autonomously, 00:03:52.030 --> 00:03:54.210 and finds a room near the venue. Think about 00:03:54.210 --> 00:03:56.629 the underlying mechanics there. It parses the 00:03:56.629 --> 00:03:59.280 image using a vision model. Yeah. It extracts 00:03:59.280 --> 00:04:01.879 the location, the band name and the date. It 00:04:01.879 --> 00:04:04.159 cross references your calendar. It accesses your 00:04:04.159 --> 00:04:06.400 payment preferences. It is doing all of that 00:04:06.400 --> 00:04:08.860 in the background. Exactly. Then it executes 00:04:08.860 --> 00:04:11.280 a complex sequence of web navigation tasks. It 00:04:11.280 --> 00:04:14.199 acts exactly like a human assistant. It is staggering. 00:04:14.280 --> 00:04:16.319 And alongside these massive shifts, there are 00:04:16.319 --> 00:04:18.819 smaller quality of life improvements. Right, 00:04:18.879 --> 00:04:21.379 like in Gboard. Yes. Gboard now has a feature 00:04:21.379 --> 00:04:24.560 called Rambler. It scrubs your voice -to -text 00:04:24.560 --> 00:04:27.879 inputs before they ever hit the screen. It automatically 00:04:27.879 --> 00:04:31.439 removes your awes and mid -sentence corrections. 00:04:31.480 --> 00:04:34.540 It uses on -device models to filter acoustic 00:04:34.540 --> 00:04:37.779 garbage. It is so helpful. It makes human communication 00:04:37.779 --> 00:04:40.720 look much cleaner. It basically edits our natural 00:04:40.720 --> 00:04:43.379 verbal stumbling in real time. Then there's the 00:04:43.379 --> 00:04:46.839 pause point feature. This one feels deeply necessary. 00:04:47.120 --> 00:04:49.839 I love this one. It forces you to wait 10 seconds 00:04:49.839 --> 00:04:52.319 before opening distracting apps, things like 00:04:52.319 --> 00:04:54.680 TikTok or Instagram. It gives you a moment to 00:04:54.680 --> 00:04:57.040 decide if you want to read a book instead. It 00:04:57.040 --> 00:04:59.100 is a great intervention. I have to admit something 00:04:59.100 --> 00:05:01.360 here. I still wrestle with doom scrolling myself. 00:05:01.680 --> 00:05:04.100 We all do. Our brains are hardwired for that 00:05:04.100 --> 00:05:06.959 quick dopamine hit. Yeah. PausePoint acts as 00:05:06.959 --> 00:05:09.160 a digital speed bump. Yes. It inserts a crucial 00:05:09.160 --> 00:05:11.279 moment of reflection into a compulsive habit. 00:05:11.459 --> 00:05:14.240 It fights algorithm -induced addiction with a 00:05:14.240 --> 00:05:16.040 counter -algorithm. They also announced some 00:05:16.040 --> 00:05:18.360 practical quick share updates. You can easily 00:05:18.360 --> 00:05:21.620 share files with iPhones via QR codes now. Moving 00:05:21.620 --> 00:05:24.899 your whole digital life from iOS to Android is 00:05:24.899 --> 00:05:27.439 getting much smoother. They are actively breaking 00:05:27.439 --> 00:05:30.399 down the walled gardens. The strategy is to make 00:05:30.399 --> 00:05:33.899 switching ecosystems entirely painless. Exactly. 00:05:34.519 --> 00:05:37.100 Lowering the barrier to entry. Security got a 00:05:37.100 --> 00:05:41.139 fascinating AI upgrade too. There is a new theft 00:05:41.139 --> 00:05:44.589 detection lock. This is very clever. The AI actually 00:05:44.589 --> 00:05:46.790 senses when a phone has been snatched out of 00:05:46.790 --> 00:05:49.129 your hand. It automatically locks the device. 00:05:49.449 --> 00:05:52.029 It also aggressively reduces the number of PI 00:05:52.029 --> 00:05:54.649 attempts a thief can make. That is a brilliant 00:05:54.649 --> 00:05:57.269 use of hardware sensors. It uses the phone's 00:05:57.269 --> 00:05:59.589 accelerometers and gyroscopes. Right. The machine 00:05:59.589 --> 00:06:01.569 learning model was trained on the physical signature 00:06:01.569 --> 00:06:05.410 of a theft. It knows the exact G -force and angle 00:06:05.410 --> 00:06:07.209 of a phone being grabbed and someone running 00:06:07.209 --> 00:06:10.410 away. Wow. It reacts faster than human reflexes 00:06:10.410 --> 00:06:13.209 could. And on a lighter note, they fully redesigned 00:06:13.209 --> 00:06:16.430 all 4 ,000 3D emojis. Right, because it is the 00:06:16.430 --> 00:06:18.290 little things that keep users tethered to an 00:06:18.290 --> 00:06:20.509 ecosystem. Looking at all this together, the 00:06:20.509 --> 00:06:23.290 ecosystem shift is vivid. It's like turning your 00:06:23.290 --> 00:06:25.910 phone from a passive filing cabinet into an active 00:06:25.910 --> 00:06:29.209 copilot. That is the perfect analogy. The device 00:06:29.209 --> 00:06:31.509 is no longer just waiting for your manual input. 00:06:31.689 --> 00:06:34.980 It is observing, predicting, and acting. But 00:06:34.980 --> 00:06:37.720 this raises a profound question. Does removing 00:06:37.720 --> 00:06:40.040 all friction with things like Gemini intelligence 00:06:40.040 --> 00:06:43.100 make us lose our own cognitive maps? That is 00:06:43.100 --> 00:06:45.620 a very valid concern. If the AI does everything, 00:06:45.939 --> 00:06:48.720 do we forget how to do things ourselves? That 00:06:48.720 --> 00:06:51.339 is the classic dilemma of cognitive offloading. 00:06:51.500 --> 00:06:54.959 Every new tool fundamentally changes our baseline 00:06:54.959 --> 00:06:57.579 expectations. Like when we got GPS. Exactly. 00:06:57.699 --> 00:07:01.060 When we offloaded navigation to GPS, our spatial 00:07:01.060 --> 00:07:03.939 memory demonstrably weakened. But it freed up 00:07:03.939 --> 00:07:06.060 mental bandwidth for other things. Right. We 00:07:06.060 --> 00:07:08.560 are outsourcing routine digital logistics to 00:07:08.560 --> 00:07:11.160 AI now. We might actually forget how to manually 00:07:11.160 --> 00:07:13.079 navigate a travel site or build a spreadsheet. 00:07:13.360 --> 00:07:16.000 But the theory is we gain time and energy for 00:07:16.000 --> 00:07:18.639 higher level thinking. It is a permanent tradeoff. 00:07:18.680 --> 00:07:20.459 So we trade a little self -reliance for ultimate 00:07:20.459 --> 00:07:23.579 daily efficiency. Got it. We're redefining what 00:07:23.579 --> 00:07:25.720 we consider a worthwhile use of our human time. 00:07:25.839 --> 00:07:28.579 If we are outsourcing our daily logistics to 00:07:28.579 --> 00:07:31.120 these seamless assistants, it takes a terrifying 00:07:31.120 --> 00:07:33.160 amount of server power to keep that illusion 00:07:33.160 --> 00:07:36.279 alive without lagging. Oh, the compute power 00:07:36.279 --> 00:07:38.899 required is mind boggling. Which brings us to 00:07:38.899 --> 00:07:41.120 the actual physical engines running all this. 00:07:41.319 --> 00:07:44.079 The numbers we are seeing are massive. The scale 00:07:44.079 --> 00:07:46.360 of capital deployment right now is unprecedented. 00:07:46.939 --> 00:07:49.759 We are watching a historical shift in global 00:07:49.759 --> 00:07:53.240 resource allocation. Let us look at OpenAI. They 00:07:53.240 --> 00:07:55.600 just had a massive liquidity event last fall. 00:07:55.939 --> 00:08:00.319 It was valued at $6 .6 billion. It is hard to 00:08:00.319 --> 00:08:02.740 even fathom that number. Roughly 600 employees 00:08:02.740 --> 00:08:05.870 cashed out. Around 75 of those employees walked 00:08:05.870 --> 00:08:09.519 away with over $30 million each. That fundamentally 00:08:09.519 --> 00:08:12.500 changes the internal dynamics of a company. You 00:08:12.500 --> 00:08:14.439 suddenly have a large group of incredibly wealthy 00:08:14.439 --> 00:08:17.079 engineers. Yeah. It shifts the culture from a 00:08:17.079 --> 00:08:19.399 scrappy research lab to a legacy institution 00:08:19.399 --> 00:08:21.920 managing massive wealth. It certainly changes 00:08:21.920 --> 00:08:24.319 the stakes. But the infrastructure spending happening 00:08:24.319 --> 00:08:26.779 outside the company is even bigger. Way bigger. 00:08:27.139 --> 00:08:30.000 SoftBank is making aggressive moves. Their CEO 00:08:30.000 --> 00:08:33.759 is discussing a $100 billion AI investment in 00:08:33.759 --> 00:08:36.320 France. That is an astonishing figure. It is 00:08:36.320 --> 00:08:40.509 literally larger than the G. It includes building 00:08:40.509 --> 00:08:43.870 massive new AI data centers. They want to radically 00:08:43.870 --> 00:08:46.690 expand the country's computing power infrastructure. 00:08:47.149 --> 00:08:49.750 This is deeply tied to national AI sovereignty. 00:08:50.289 --> 00:08:53.149 Countries are realizing that compute power is 00:08:53.149 --> 00:08:56.429 the new oil. If you don't control your own compute, 00:08:56.690 --> 00:08:59.149 you are at the mercy of foreign tech giants. 00:08:59.350 --> 00:09:02.409 Ain't true. SoftBank is positioning itself as 00:09:02.409 --> 00:09:05.570 the primary financier of this new global infrastructure. 00:09:05.870 --> 00:09:07.899 They are pouring the concrete. The corporate 00:09:07.899 --> 00:09:10.500 drama behind the scenes is just as intense as 00:09:10.500 --> 00:09:13.580 the spending. Sam Altman recently testified under 00:09:13.580 --> 00:09:16.299 oath. That testimony was wild. He shared some 00:09:16.299 --> 00:09:19.059 fascinating and honestly bizarre details about 00:09:19.059 --> 00:09:21.139 his relationship with Elon Musk. The history 00:09:21.139 --> 00:09:23.240 between those two founders is incredibly complicated. 00:09:23.519 --> 00:09:26.779 It is the defining origin story of modern AI. 00:09:27.080 --> 00:09:29.559 Altman testified that Musk once suggested open 00:09:29.559 --> 00:09:32.000 AI could pass to his children if he died. I still 00:09:32.000 --> 00:09:33.879 cannot believe he said that. Altman called it 00:09:33.879 --> 00:09:35.740 one of the hair -raising moments of their relationship. 00:09:36.320 --> 00:09:38.360 It highlights the philosophical clashes they 00:09:38.360 --> 00:09:41.299 had early on. They were just arguing over a software 00:09:41.299 --> 00:09:43.840 product. They were arguing over who should legally 00:09:43.840 --> 00:09:46.759 control artificial general intelligence. It sounds 00:09:46.759 --> 00:09:49.019 like science fiction, but to them, the stakes 00:09:49.019 --> 00:09:51.590 were very real. Meanwhile, Anthropic is throwing 00:09:51.590 --> 00:09:54.590 its own weight around. They are hiring a literal 00:09:54.590 --> 00:09:59.590 Claude evangelist. The job pays up to $315 ,000 00:09:59.590 --> 00:10:03.309 a year. They want someone specifically to help 00:10:03.309 --> 00:10:06.970 founders and VCs adopt their AI products. It 00:10:06.970 --> 00:10:09.169 shows how critical developer adoption has become. 00:10:09.409 --> 00:10:12.049 The best model doesn't win automatically. You 00:10:12.049 --> 00:10:14.549 have to fight a ground war for mindshare among 00:10:14.549 --> 00:10:16.830 the people building the apps. Anthropic is also 00:10:16.830 --> 00:10:19.940 flexing its capital in acquisitions. reportedly 00:10:19.940 --> 00:10:22.360 in late -stage talks to buy a startup called 00:10:22.360 --> 00:10:25.299 Stainless. This is a huge deal. It builds crucial 00:10:25.299 --> 00:10:28.259 developer tools. OpenAI and Google actually already 00:10:28.259 --> 00:10:31.019 use Stainless. And Throdnick might buy it for 00:10:31.019 --> 00:10:34.419 over $300 million. That is a brilliant, aggressive 00:10:34.419 --> 00:10:37.159 chess move. If Anthropic owns the underlying 00:10:37.159 --> 00:10:40.399 tools that their biggest rivals rely on, they 00:10:40.399 --> 00:10:43.000 gain a massive structural advantage. It is about 00:10:43.000 --> 00:10:45.679 owning the plumbing of the entire ecosystem. 00:10:46.080 --> 00:10:48.659 Speaking of advantages, OpenAI is fighting back 00:10:48.659 --> 00:10:50.360 with new technology. They just released something 00:10:50.360 --> 00:10:52.620 called Daybreak. This is their new security play. 00:10:53.049 --> 00:10:55.789 It is their direct answer to anthropic -sclawed 00:10:55.789 --> 00:10:58.990 mythos. Security is rapidly becoming the next 00:10:58.990 --> 00:11:01.289 major battleground for these models. Daybreak 00:11:01.289 --> 00:11:04.809 combines two distinct systems. It uses GPT -5 00:11:04.809 --> 00:11:08.950 .5 cyber and codex security, an AI guard dog 00:11:08.950 --> 00:11:11.970 catching software bugs before hackers do. The 00:11:11.970 --> 00:11:15.659 AI writes the code. But now it also audits the 00:11:15.659 --> 00:11:18.360 code autonomously. Which is crazy to think about. 00:11:18.500 --> 00:11:21.240 It runs adversarial attacks against itself. It 00:11:21.240 --> 00:11:23.700 hardens critical systems against zero -day exploits 00:11:23.700 --> 00:11:26.340 at machine speed. There are also some incredibly 00:11:26.340 --> 00:11:29.279 fun details emerging from all this coding. The 00:11:29.279 --> 00:11:31.919 team behind Cloud Code shared a strange tip. 00:11:32.100 --> 00:11:34.399 Oh, the markdown thing. Yes. They said developers 00:11:34.399 --> 00:11:36.379 should avoid using markdown files for instructions. 00:11:37.279 --> 00:11:39.360 Funny enough, they called the tip unreasonably 00:11:39.360 --> 00:11:41.639 effective. Sometimes the behavior of these massive 00:11:41.639 --> 00:11:44.039 models defines logical explanation. It really 00:11:44.039 --> 00:11:46.860 does. It is likely a quirky artifact hidden deep 00:11:46.860 --> 00:11:49.820 in the training data. The model just pays better 00:11:49.820 --> 00:11:53.120 attention to raw text. And everyday people are 00:11:53.120 --> 00:11:55.559 using these massive models for entirely different, 00:11:55.679 --> 00:11:59.340 hilarious things. Like what? A fan used AI video 00:11:59.340 --> 00:12:01.679 tools to insert himself into the movie Titanic. 00:12:02.179 --> 00:12:05.639 He somehow fixed every single fan complaint. 00:12:06.009 --> 00:12:08.950 in a single video. He saved Jack. Right, and 00:12:08.950 --> 00:12:10.889 he basically became the hero the movie needed. 00:12:11.070 --> 00:12:14.190 Exactly. It shows the incredible, chaotic, creative 00:12:14.190 --> 00:12:16.809 potential of these tools once they hit the mainstream. 00:12:17.399 --> 00:12:19.620 But looking at the big picture, a massive question 00:12:19.620 --> 00:12:22.580 remains. With SoftBank dropping $100 billion 00:12:22.580 --> 00:12:25.720 and OpenAI minting millionaires overnight, are 00:12:25.720 --> 00:12:28.139 we building foundational plumbing or is this 00:12:28.139 --> 00:12:30.259 just an unsustainable arms race? It certainly 00:12:30.259 --> 00:12:32.080 looks like an arms race from the outside. The 00:12:32.080 --> 00:12:34.320 valuations are undeniably astronomical. Right. 00:12:34.460 --> 00:12:35.759 But we have to look at what they're actually 00:12:35.759 --> 00:12:38.019 building with that cash. The data centers are 00:12:38.019 --> 00:12:40.120 real. The power grids are real. They are physical 00:12:40.120 --> 00:12:43.419 assets. Exactly. They are pouring actual concrete 00:12:43.419 --> 00:12:46.080 and laying thousands of miles of fiber optics. 00:12:46.650 --> 00:12:49.710 Even if an economic bubble bursts, that physical 00:12:49.710 --> 00:12:52.769 infrastructure remains. It will power the next 00:12:52.769 --> 00:12:56.250 generation of software, regardless of which specific 00:12:56.250 --> 00:13:00.309 company wins the AI war. Massive capital is literally 00:13:00.309 --> 00:13:02.870 laying down the physical plumbing for the future. 00:13:03.029 --> 00:13:06.210 Makes sense. Yes. It is very similar to the telecom 00:13:06.210 --> 00:13:08.889 boom of the late 90s. How so? The companies went 00:13:08.889 --> 00:13:11.549 bust, but the fiber they laid gave us the modern 00:13:11.549 --> 00:13:14.299 Internet. Wait, I have to push back there. Telecom 00:13:14.299 --> 00:13:17.519 laid neutral fiber that anyone could use. These 00:13:17.519 --> 00:13:20.720 AI data centers are highly proprietary walled 00:13:20.720 --> 00:13:23.740 gardens. SoftBank isn't building a public park. 00:13:23.940 --> 00:13:25.679 Isn't that fundamentally different? That is a 00:13:25.679 --> 00:13:27.659 very sharp distinction. You are right. Yeah. 00:13:27.759 --> 00:13:30.740 The physical hardware exists, but the access 00:13:30.740 --> 00:13:33.120 is strippedly gated by the corporate giants. 00:13:33.419 --> 00:13:36.799 It is infrastructure, but it is privatized infrastructure. 00:13:37.299 --> 00:13:40.000 Sponsor. The billions of dollars we just discussed 00:13:40.000 --> 00:13:43.139 are all chasing one holy grail. They all want 00:13:43.139 --> 00:13:45.480 to make computers feel completely, seamlessly 00:13:45.480 --> 00:13:48.080 human. That is the ultimate goal. This leads 00:13:48.080 --> 00:13:51.080 perfectly into Mira Marotti's new venture. It's 00:13:51.080 --> 00:13:52.720 perhaps the most exciting technical development 00:13:52.720 --> 00:13:55.240 we are covering today. It changes the core interaction 00:13:55.240 --> 00:13:58.100 paradigm completely. Maradi is the former CTO 00:13:58.100 --> 00:14:01.240 of OpenAI. She recently founded Thinking Machines 00:14:01.240 --> 00:14:04.159 Lab. We will call it TML. Right. They just dropped 00:14:04.159 --> 00:14:06.600 a research preview that is stunning. It focuses 00:14:06.600 --> 00:14:08.980 on the one thing we actually hate about talking 00:14:08.980 --> 00:14:12.000 to AI. The awkward silence. Yeah. The fundamental 00:14:12.000 --> 00:14:14.600 lack of human conversational rhythm. Exactly. 00:14:14.940 --> 00:14:17.539 They are trying to end the chat bubble interface 00:14:17.539 --> 00:14:21.500 forever. TML built a system that collaborates 00:14:21.500 --> 00:14:24.539 in a live, continuous streaming loop. More waiting. 00:14:24.700 --> 00:14:27.840 It processes your voice, video, and text simultaneously. 00:14:28.039 --> 00:14:30.480 This is a radical departure from turn -based 00:14:30.480 --> 00:14:33.480 interactions. Standard AI operates like a walkie 00:14:33.480 --> 00:14:35.899 -talkie. Yeah. You speak, you wait. It speaks. 00:14:36.440 --> 00:14:38.720 With TML, you no longer wait for the model to 00:14:38.720 --> 00:14:41.580 finish generating a response. The model perceives 00:14:41.580 --> 00:14:44.580 and processes data in tiny rapid bursts. It operates 00:14:44.580 --> 00:14:46.940 in chunks of just 200 milliseconds at a time. 00:14:47.000 --> 00:14:49.500 That is incredibly fast. It is. This allows it 00:14:49.500 --> 00:14:52.000 to make natural human noises. It says, mm -hmm, 00:14:52.080 --> 00:14:54.419 yeah, and got it while you are actively speaking. 00:14:54.700 --> 00:14:57.639 Human conversation is incredibly complex sociologically. 00:14:57.879 --> 00:15:00.360 We rely heavily on those subtle back -channel 00:15:00.360 --> 00:15:02.840 signals. They are so important. They let us know 00:15:02.840 --> 00:15:05.519 the other person is tracking our thoughts. Standard 00:15:05.519 --> 00:15:07.840 AI lacks this entirely, which is why it feels 00:15:07.840 --> 00:15:10.399 dead. It even reacts to your facial expressions 00:15:10.399 --> 00:15:12.779 before you finish your sentence. Wow. If you 00:15:12.779 --> 00:15:15.360 look confused mid -thought, the AI stops and 00:15:15.360 --> 00:15:17.539 clarifies its point. That requires incredibly 00:15:17.539 --> 00:15:20.580 low system latency. 200 milliseconds is the magic 00:15:20.580 --> 00:15:23.340 number. Why that specific number? It matches 00:15:23.340 --> 00:15:26.519 human conversational gap times perfectly. If 00:15:26.519 --> 00:15:28.960 it takes longer than that, our brains register 00:15:28.960 --> 00:15:32.200 it as an awkward pause. To stay that impossibly 00:15:32.200 --> 00:15:35.419 fast... TML uses a unique two layer architecture. 00:15:36.000 --> 00:15:38.320 This is a brilliant engineering solution to a 00:15:38.320 --> 00:15:40.879 very hard physics problem. Because you cannot 00:15:40.879 --> 00:15:43.840 run a massive intelligence model in 200 milliseconds. 00:15:44.039 --> 00:15:46.259 Exactly. You just can't. So they split it up. 00:15:46.299 --> 00:15:49.440 They use a live model to handle immediate superficial 00:15:49.440 --> 00:15:52.019 interaction. Then a background model does the 00:15:52.019 --> 00:15:54.320 heavy intellectual lifting. One brain for quick 00:15:54.320 --> 00:15:57.100 reactions, another brain for deep thinking. Exactly. 00:15:57.100 --> 00:16:00.279 It is like the human amygdala versus the prefrontal 00:16:00.279 --> 00:16:03.200 cortex. That makes sense. The live model manages 00:16:03.200 --> 00:16:06.799 the social dynamics and the hmms. The background 00:16:06.799 --> 00:16:09.759 model pulls the factual data and structures the 00:16:09.759 --> 00:16:12.879 complex argument. They work in tandem seamlessly. 00:16:13.279 --> 00:16:15.980 Because it processes video in real time, the 00:16:15.980 --> 00:16:19.519 real -world capabilities are wild. The model 00:16:19.519 --> 00:16:22.220 can literally watch you exercise through your 00:16:22.220 --> 00:16:24.519 phone camera. Oh, yeah. It will count your workout 00:16:24.519 --> 00:16:26.840 reps out loud. It acts as an interactive physical 00:16:26.840 --> 00:16:30.159 coach. It sees your posture, analyzes your form, 00:16:30.240 --> 00:16:32.620 and corrects it instantly. It can also translate 00:16:32.620 --> 00:16:36.240 live speech directly from a TV screen. It can 00:16:36.240 --> 00:16:38.460 speak up proactively if it sees something change 00:16:38.460 --> 00:16:39.980 in your environment. Right, like telling you 00:16:39.980 --> 00:16:42.840 your coffee has got to stop. It has genuine situational 00:16:42.840 --> 00:16:45.879 awareness. That is a massive leap forward from 00:16:45.879 --> 00:16:48.360 a text box. Think about standard AI voice modes 00:16:48.360 --> 00:16:50.940 today. If you have ever tried to interrupt one, 00:16:51.059 --> 00:16:53.379 you know how clunky it feels. Oh, it is terrible. 00:16:53.580 --> 00:16:56.240 It refuses to stop or it glitches. It breaks 00:16:56.240 --> 00:16:58.750 the illusion of intelligence immediately. TML 00:16:58.750 --> 00:17:01.610 is placing a massive contrarian bet here. They 00:17:01.610 --> 00:17:04.029 believe the future of AI isn't just about raw, 00:17:04.210 --> 00:17:07.069 escalating intelligence. No, it is deeply about 00:17:07.069 --> 00:17:10.910 latency, empathy, and conversational flow. Whoa. 00:17:11.609 --> 00:17:15.410 Beat it. Imagine scaling those 200 millisecond 00:17:15.410 --> 00:17:18.309 reactions to a billion queries. It requires an 00:17:18.309 --> 00:17:20.250 entirely different approach to server architecture. 00:17:20.609 --> 00:17:23.130 That is why the two -layer system is so crucial. 00:17:23.529 --> 00:17:27.190 Yeah. They are basically rewriting how data moves. 00:17:27.430 --> 00:17:30.230 But I have to ask a strategic question. If TML 00:17:30.230 --> 00:17:32.670 is betting everything on latency and flow rather 00:17:32.670 --> 00:17:35.109 than just raw smarts, can they actually survive 00:17:35.109 --> 00:17:37.529 as a standalone company? That is the million 00:17:37.529 --> 00:17:39.569 dollar question. Or is this just a preview of 00:17:39.569 --> 00:17:42.150 a feature that Google or OpenAI will eventually 00:17:42.150 --> 00:17:44.789 absorb? That is the ultimate Silicon Valley dilemma 00:17:44.789 --> 00:17:47.730 right now. Interaction models do create a powerful 00:17:47.730 --> 00:17:50.759 user mode. True. If users fall in love with that 00:17:50.759 --> 00:17:53.180 fluidity, they will not want to go back to clunky 00:17:53.180 --> 00:17:56.599 walkie -talkie AI. However, tech giants have 00:17:56.599 --> 00:17:59.220 immense distribution power. Right. Google already 00:17:59.220 --> 00:18:01.839 has the phones. Exactly. Google can push a software 00:18:01.839 --> 00:18:03.980 update to billions of Android phones overnight. 00:18:04.579 --> 00:18:07.140 TML has to build their user base from absolute 00:18:07.140 --> 00:18:09.900 scratch. That is a tough climb. They are pioneering 00:18:09.900 --> 00:18:12.359 the interface, but the giants are watching closely, 00:18:12.480 --> 00:18:14.740 and they have infinite capital to copy it. They 00:18:14.740 --> 00:18:17.079 either redefine the whole industry or get swallowed 00:18:17.079 --> 00:18:20.339 by a giant. fascinating it will be a thrilling 00:18:20.339 --> 00:18:22.319 narrative to watch unfold over the next year 00:18:22.319 --> 00:18:24.319 let us step back and look at the big picture 00:18:24.319 --> 00:18:26.920 how do all these disparate pieces fit together 00:18:26.920 --> 00:18:30.000 today we can trace A very clear through line 00:18:30.000 --> 00:18:32.720 across all these stories. AI is fundamentally 00:18:32.720 --> 00:18:36.519 moving from reactive text boxes to omnipresent 00:18:36.519 --> 00:18:40.059 real time partners. We see it in Google Androids, 00:18:40.079 --> 00:18:42.759 proactive widgets building themselves. We see 00:18:42.759 --> 00:18:45.160 it in the massive physical infrastructure being 00:18:45.160 --> 00:18:47.720 built globally to support these complex agents. 00:18:47.920 --> 00:18:50.880 Right. And we see it most clearly in TML's 200 00:18:50.880 --> 00:18:53.980 millisecond conversational bursts. The new metric 00:18:53.980 --> 00:18:56.539 for AI isn't just how smart it is on a test. 00:18:56.779 --> 00:18:59.569 Exactly. The metric is. how seamlessly it flows 00:18:59.569 --> 00:19:02.589 with human latency. It is about emotional intelligence 00:19:02.589 --> 00:19:05.170 and frictionless interaction. It is a profound 00:19:05.170 --> 00:19:07.569 shift in our relationship with technology. I 00:19:07.569 --> 00:19:09.109 want to leave you with a final thought to mull 00:19:09.109 --> 00:19:11.549 over. Two sec silence. We talked heavily about 00:19:11.549 --> 00:19:13.650 ending the awkward silence today. We talked about 00:19:13.650 --> 00:19:16.990 fluid, proactive interactions. If an AI can read 00:19:16.990 --> 00:19:19.539 your micro expressions perfectly. and interrupt 00:19:19.539 --> 00:19:21.759 you with the precise level of empathy in just 00:19:21.759 --> 00:19:24.720 200 milliseconds, at what point does talking 00:19:24.720 --> 00:19:26.799 to a machine become more comforting than talking 00:19:26.799 --> 00:19:29.579 to a human? Wow. Thank you for joining us on 00:19:29.579 --> 00:19:32.680 this deep dive. We will see you next time. U2RO 00:19:32.680 --> 00:19:32.859 Music.