WEBVTT 00:00:01.079 --> 00:00:03.419 We used to open all these different apps to do 00:00:03.419 --> 00:00:05.540 our work, but now it feels like the apps are 00:00:05.540 --> 00:00:09.199 just doing the work for us. Intro music. Welcome 00:00:09.199 --> 00:00:11.419 to the Deep Dive. We have a completely new technological 00:00:11.419 --> 00:00:14.019 reality to explore with you today. Yeah, I am 00:00:14.019 --> 00:00:16.199 so incredibly ready for this one. It is going 00:00:16.199 --> 00:00:20.100 to be big. We are unpacking OpenAI's GPT 5 .5 00:00:20.100 --> 00:00:23.300 super app ambitions today. We are also looking 00:00:23.300 --> 00:00:26.379 at the dramatic reality of an AI -driven workforce. 00:00:26.879 --> 00:00:30.359 Google is literally generating 75 % of its code 00:00:30.359 --> 00:00:33.259 autonomously right now. Which is just wild to 00:00:33.259 --> 00:00:36.340 think about. Right. We are also tracking AI robots 00:00:36.340 --> 00:00:39.939 dominating physical sports. And finally, a fascinating 00:00:39.939 --> 00:00:42.700 system using Wikipedia to predict our technological 00:00:42.700 --> 00:00:45.460 future. Let's start with the catalyst for this 00:00:45.460 --> 00:00:49.200 massive shift. OpenAI just released GPT 5 .5. 00:00:49.320 --> 00:00:51.539 It basically aims to be the only software you 00:00:51.539 --> 00:00:54.200 ever need. Yeah, it merges everything into one 00:00:54.200 --> 00:00:57.479 single agentic powerhouse. Right. Your browser, 00:00:57.640 --> 00:00:59.700 your code editor, and your daily assistant, they 00:00:59.700 --> 00:01:02.179 are all just combined. It officially ends the 00:01:02.179 --> 00:01:04.420 constant app hopping we do every day. I mean, 00:01:04.420 --> 00:01:07.260 we saw leaks about the 3D capabilities earlier. 00:01:07.560 --> 00:01:10.840 this year, but the official launch focuses heavily 00:01:10.840 --> 00:01:13.959 on autonomy. And core efficiency. Yeah. Sam Altman 00:01:13.959 --> 00:01:16.219 has a very specific vision for this. He really 00:01:16.219 --> 00:01:18.799 does. He wants a unified everything app for work 00:01:18.799 --> 00:01:21.900 and creation. But Elon Musk has a totally different 00:01:21.900 --> 00:01:23.980 approach with X. Much more compartmentalized. 00:01:24.099 --> 00:01:27.640 Exactly. Musk is focusing purely on social interactions 00:01:27.640 --> 00:01:30.879 and finance. We are watching two very different 00:01:30.879 --> 00:01:33.719 philosophies collide. It's basically a super 00:01:33.719 --> 00:01:38.640 app arms race right now. and GBT 5 .5 is currently 00:01:38.640 --> 00:01:41.260 leading the performance benchmarks. It beats 00:01:41.260 --> 00:01:45.620 Gemini 3 .1 Pro and Claude Opus 4 .7 handily. 00:01:46.040 --> 00:01:48.659 The real standout here is their new science leap 00:01:48.659 --> 00:01:51.159 capability. Oh, yeah? They are seeing massive 00:01:51.159 --> 00:01:53.640 gains in AI -led drug discovery. These agents 00:01:53.640 --> 00:01:56.680 can run much longer, highly complex reasoning 00:01:56.680 --> 00:01:59.239 loops, and they do it without breaking down. 00:01:59.439 --> 00:02:01.920 Which is huge. I still wrestle with prompt drift 00:02:01.920 --> 00:02:05.079 myself. Oh, same. You spend hours refining a 00:02:05.079 --> 00:02:08.870 specific goal. Then? 20 prompts later, the AI 00:02:08.870 --> 00:02:11.129 hallucinates a completely different outcome. 00:02:11.289 --> 00:02:13.909 It just forgets the core directive. It is incredibly 00:02:13.909 --> 00:02:16.530 frustrating. Right. And that is exactly the problem 00:02:16.530 --> 00:02:19.469 this new architecture solves. Older models would 00:02:19.469 --> 00:02:22.050 just lose the plot over time. Their context windows 00:02:22.050 --> 00:02:24.889 would essentially overflow. Exactly. But GPT 00:02:24.889 --> 00:02:28.330 -5 .5 maintains strict focus without bankrupting 00:02:28.330 --> 00:02:31.389 your API budget. The underlying logic is purely 00:02:31.389 --> 00:02:34.680 about contextual memory retention. It uses fewer 00:02:34.680 --> 00:02:38.060 tokens for its basic reasoning tasks. So it mathematically 00:02:38.060 --> 00:02:40.280 has more memory left to remember your original 00:02:40.280 --> 00:02:42.620 goal. That is exactly why the drug discovery 00:02:42.620 --> 00:02:45.800 applications are so successful. The AI analyzes 00:02:45.800 --> 00:02:48.300 complex molecular structures for hours on end. 00:02:48.439 --> 00:02:50.659 And it never forgets the baseline parameters. 00:02:50.879 --> 00:02:53.319 Yeah. Beat. Beat. But I look at this ecosystem 00:02:53.319 --> 00:02:55.340 and I kind of have to push back. Okay, let's 00:02:55.340 --> 00:02:58.099 hear it. Why is this specific super app model 00:02:58.099 --> 00:03:01.099 truly necessary? Why not just keep our specialized 00:03:01.099 --> 00:03:04.319 tools separate? Well. Separate tools create immense 00:03:04.319 --> 00:03:07.259 cognitive friction for human users. Switching 00:03:07.259 --> 00:03:09.800 contexts constantly destroys our deep focus. 00:03:10.039 --> 00:03:13.020 Right. An agentic super app removes those barriers 00:03:13.020 --> 00:03:15.740 entirely. It handles the complex routing between 00:03:15.740 --> 00:03:18.379 specific tasks entirely in the background. So 00:03:18.379 --> 00:03:20.400 it's about having one invisible assistant for 00:03:20.400 --> 00:03:23.159 absolutely everything. Precisely. You state your 00:03:23.159 --> 00:03:26.169 goal. And the system handles the routing. GPT 00:03:26.169 --> 00:03:28.789 5 .5 represents the grand vision of our digital 00:03:28.789 --> 00:03:32.990 future. But big tech is living this AI first 00:03:32.990 --> 00:03:35.210 reality right now. Yeah. The corporate reality 00:03:35.210 --> 00:03:37.270 right now is actually pretty dramatic to witness. 00:03:37.449 --> 00:03:40.250 It really is. Massive structural shifts in tech 00:03:40.250 --> 00:03:42.990 labor are happening everywhere. Google just revealed 00:03:42.990 --> 00:03:45.969 an absolutely staggering internal metric. 75 00:03:45.969 --> 00:03:49.270 % of its internal code is now AI generated. That 00:03:49.270 --> 00:03:51.870 is a staggering volume of automated production. 00:03:52.129 --> 00:03:54.050 But let's look at why. they are actually doing 00:03:54.050 --> 00:03:56.250 this. Right. It fundamentally improves their 00:03:56.250 --> 00:03:58.550 security and their backend operations. Because 00:03:58.550 --> 00:04:02.389 humans make typos. Machines don't get tired or 00:04:02.389 --> 00:04:05.490 overlook simple syntax errors. Exactly. They 00:04:05.490 --> 00:04:08.270 follow strict security protocols with absolute 00:04:08.270 --> 00:04:11.150 rigid perfection. It changes the entire engineering 00:04:11.150 --> 00:04:14.610 pipeline. Meanwhile, Microsoft and Meta are aggressively 00:04:14.610 --> 00:04:17.649 cutting thousands of jobs. Yeah. Microsoft is 00:04:17.649 --> 00:04:21.470 offering buyouts to about 7 % of its staff. Meta 00:04:21.470 --> 00:04:24.250 is executing a flat 10 % reduction across the 00:04:24.250 --> 00:04:27.529 board. Yet they are pouring billions into new 00:04:27.529 --> 00:04:30.550 AI productivity tools. A huge shift. They are 00:04:30.550 --> 00:04:33.990 quite literally trading human headcount for artificial 00:04:33.990 --> 00:04:36.490 intelligence capital. We see this capital shift 00:04:36.490 --> 00:04:39.290 reflected in venture valuations, too. Look at 00:04:39.290 --> 00:04:41.759 the autonomous software startup Cognition. They 00:04:41.759 --> 00:04:44.259 just raised money at a $25 billion valuation. 00:04:44.720 --> 00:04:47.660 That is a massive jump from their previous $10 00:04:47.660 --> 00:04:50.500 .2 billion. And they specialize in autonomous 00:04:50.500 --> 00:04:53.120 deployment. Let me clarify what that means for 00:04:53.120 --> 00:04:56.339 you. It's software that writes, tests, and launches 00:04:56.339 --> 00:04:58.860 itself without human help. That completely rewrites 00:04:58.860 --> 00:05:00.980 the core economics of running a tech company. 00:05:01.220 --> 00:05:03.639 You simply don't need massive teams of junior 00:05:03.639 --> 00:05:07.490 developers anymore. And... Humans are understandably 00:05:07.490 --> 00:05:10.089 scrambling to adapt to this. Oh, yeah. Stanford 00:05:10.089 --> 00:05:13.029 University is hosting this viral AI Coachella 00:05:13.029 --> 00:05:16.589 class right now. Over 500 students are packing 00:05:16.589 --> 00:05:18.810 the lecture halls every week. It's packed. They 00:05:18.810 --> 00:05:21.269 are desperate to hear directly from Sam Altman 00:05:21.269 --> 00:05:24.269 and Satya Nadella. They are panicking, but they 00:05:24.269 --> 00:05:27.209 also see the massive opportunity. Even retail 00:05:27.209 --> 00:05:30.189 investors want a piece of this action. AngelList 00:05:30.189 --> 00:05:33.180 just opened a brand new AI fund. For the masses. 00:05:33.439 --> 00:05:36.279 It already has over 4 .8 million views online. 00:05:36.639 --> 00:05:39.120 People want early stage access to foundational 00:05:39.120 --> 00:05:41.779 companies. Like OpenAI and Anthropic. They want 00:05:41.779 --> 00:05:43.899 to invest before these automated systems completely 00:05:43.899 --> 00:05:47.339 take over. Beat. I look at Cognition's massive 00:05:47.339 --> 00:05:51.160 $25 billion valuation. I can't help but feel 00:05:51.160 --> 00:05:54.759 like syntax coders are essentially extinct. Yeah. 00:05:54.860 --> 00:05:57.759 Convince me I'm wrong here. Is human coding effectively 00:05:57.759 --> 00:06:00.579 dead? It isn't dead, but it is fundamentally 00:06:00.579 --> 00:06:03.879 evolving right before our eyes. Humans are stepping 00:06:03.879 --> 00:06:06.800 away from raw manual syntax creation. They are 00:06:06.800 --> 00:06:08.920 focusing entirely on system architecture and 00:06:08.920 --> 00:06:12.300 high -level logic now. The AI handles the tedious 00:06:12.300 --> 00:06:14.860 line -by -line implementation for them. Right. 00:06:14.980 --> 00:06:18.079 Developers are now editors managing AI, not just 00:06:18.079 --> 00:06:21.420 typing code. Exactly. It elevates the human to 00:06:21.420 --> 00:06:26.420 a high -level managerial role. Sponsor. We have 00:06:26.420 --> 00:06:29.620 explored how AI conquered our digital workspaces, 00:06:29.839 --> 00:06:32.980 but we need to look at physical latency and hardware 00:06:32.980 --> 00:06:35.779 next. Yes, definitely. AI is crossing the boundary 00:06:35.779 --> 00:06:38.040 into physical hardware right now. The physical 00:06:38.040 --> 00:06:40.019 breakthrough blew my mind entirely this week. 00:06:40.100 --> 00:06:43.060 Really? Sony built an advanced AI robot named 00:06:43.060 --> 00:06:46.420 Innis for table tennis. It just beat elite human 00:06:46.420 --> 00:06:49.480 players with near perfect consistency. Robotics 00:06:49.480 --> 00:06:51.019 teams have failed at this specific challenge 00:06:51.019 --> 00:06:54.500 for decades. Ping pong requires incredibly complex 00:06:54.500 --> 00:06:56.980 physics calculations in absolute real time. Yeah. 00:06:57.060 --> 00:06:59.660 It is a perfect textbook example of Moravec's 00:06:59.660 --> 00:07:02.540 paradox. Yeah. High level logic is computationally 00:07:02.540 --> 00:07:05.420 easy for AI, but basic physical movement and 00:07:05.420 --> 00:07:07.660 spatial reasoning are incredibly hard. Right. 00:07:07.740 --> 00:07:09.800 You can actually watch the match video online 00:07:09.800 --> 00:07:12.319 right now. The robot anticipates the ball's spin 00:07:12.319 --> 00:07:15.019 and trajectory flawlessly. It calculates the 00:07:15.019 --> 00:07:18.120 ball's aerodynamics in your milliseconds. If 00:07:18.120 --> 00:07:21.459 AI can solve physical latency in a fast -paced 00:07:21.459 --> 00:07:26.100 game. Yeah. It is a massive leap for real world 00:07:26.100 --> 00:07:29.040 physical AI applications. Think about the broader 00:07:29.040 --> 00:07:31.500 implications here. If a robot reacts to a ping 00:07:31.500 --> 00:07:33.959 pong ball instantly. Imagine a manufacturing 00:07:33.959 --> 00:07:37.060 facility. Exactly. Or navigating unpredictable 00:07:37.060 --> 00:07:39.759 environments during dangerous search and rescue 00:07:39.759 --> 00:07:42.639 missions. But speed in the real world creates 00:07:42.639 --> 00:07:45.259 incredible vulnerabilities. Organizations are 00:07:45.259 --> 00:07:47.199 rushing frantically to protect their proprietary 00:07:47.199 --> 00:07:50.279 operational data. Anthropic has an unreleased, 00:07:50.319 --> 00:07:52.860 highly advanced model called Mythos. right now. 00:07:52.980 --> 00:07:55.519 And people are tracking its usage using a tool 00:07:55.519 --> 00:07:58.839 called Mythoswatch. Over 51 governments and major 00:07:58.839 --> 00:08:00.740 banks are already listed there. Wow. They're 00:08:00.740 --> 00:08:03.459 testing it quietly behind tightly closed doors. 00:08:03.720 --> 00:08:06.180 Because public AI models are essentially data 00:08:06.180 --> 00:08:08.709 sieves. Corporations are terrified of feeding 00:08:08.709 --> 00:08:11.410 private internal data into public models. Absolutely. 00:08:11.790 --> 00:08:14.610 OpenAI just launched a local privacy filter to 00:08:14.610 --> 00:08:18.050 explicitly address this. It actively stops API 00:08:18.050 --> 00:08:21.269 keys and private data from leaking out. It runs 00:08:21.269 --> 00:08:24.310 completely locally with over 96 % proven accuracy. 00:08:24.730 --> 00:08:26.910 Which is incredible. You can grab it for free 00:08:26.910 --> 00:08:30.029 on GitHub right now. It scrubs your prompts before 00:08:30.029 --> 00:08:32.690 they ever leave your physical computer. We're 00:08:32.690 --> 00:08:35.129 also seeing tools securely integrate these agents 00:08:35.129 --> 00:08:38.230 internally. Right. A platform called Colab brings 00:08:38.230 --> 00:08:40.950 these agents directly inside Slack. It links 00:08:40.950 --> 00:08:43.509 your tools without switching apps or compromising 00:08:43.509 --> 00:08:46.830 data. It keeps the context alive across every 00:08:46.830 --> 00:08:50.169 single project securely. Beat. Let's explore 00:08:50.169 --> 00:08:53.120 the broader enterprise implications here. Will 00:08:53.120 --> 00:08:56.360 localized AI filters simply become the new standard 00:08:56.360 --> 00:08:58.700 for enterprise? Absolutely. You cannot legally 00:08:58.700 --> 00:09:01.519 send sensitive client data to external servers. 00:09:01.759 --> 00:09:04.620 Local filters act as an impenetrable algorithmic 00:09:04.620 --> 00:09:07.440 firewall. They ensure proprietary data never 00:09:07.440 --> 00:09:10.279 leaves your secure internal network. So local 00:09:10.279 --> 00:09:13.120 privacy filters are essentially the new antivirus 00:09:13.120 --> 00:09:15.379 software. That is a perfect way to look at it 00:09:15.379 --> 00:09:17.700 structurally. It is foundational security for 00:09:17.700 --> 00:09:21.120 the modern AI -driven workflow. If AI can rewrite 00:09:21.120 --> 00:09:24.000 our code and master physical sports, what does 00:09:24.000 --> 00:09:26.600 it actually think is coming next for us? Yeah, 00:09:26.700 --> 00:09:29.779 that is the big question. We turn to the 2026 00:09:29.779 --> 00:09:33.639 Momentum 100 list to find out. This is wildly 00:09:33.639 --> 00:09:35.240 different from your traditional tech prediction 00:09:35.240 --> 00:09:38.299 lists. Usually places like MIT or Stanford rely 00:09:38.299 --> 00:09:41.100 on expert human voting. They get a room full 00:09:41.100 --> 00:09:43.460 of highly credentialed specialists to guess trends. 00:09:43.799 --> 00:09:46.460 But human experts are inherently biased toward 00:09:46.460 --> 00:09:49.120 their own academic specialties. This new list 00:09:49.120 --> 00:09:52.820 is purely data -driven, objective, and incredibly 00:09:52.820 --> 00:09:57.080 fascinating. A system called Cosmos 1 .0 simply 00:09:57.080 --> 00:10:00.299 mined vast amounts of Wikipedia data. It analyzed 00:10:00.299 --> 00:10:02.580 edit age and page view momentum to predict our 00:10:02.580 --> 00:10:04.980 future. And it outperformed those human specialists 00:10:04.980 --> 00:10:07.440 by a truly significant margin. It's amazing. 00:10:07.620 --> 00:10:10.960 Cosmos 1 .0 analyzed over 23 ,000 technology 00:10:10.960 --> 00:10:14.419 -adjacent concepts using Wikipedia2Vec. Whoa. 00:10:14.759 --> 00:10:17.759 Imagine mapping 23 ,000 concepts into 100 dimensions. 00:10:18.159 --> 00:10:20.159 Let's break down those 100 -dimensional embeddings 00:10:20.159 --> 00:10:22.100 clearly for you. Think of it as a mathematical 00:10:22.100 --> 00:10:24.179 map showing how different ideas connect together. 00:10:24.620 --> 00:10:27.179 Exactly. It looks at how mathematically close 00:10:27.179 --> 00:10:30.039 different articles actually are. Right. Frequent 00:10:30.039 --> 00:10:32.679 Wikipedia edits by researchers indicate incoming 00:10:32.679 --> 00:10:35.980 patents and venture capital. It is a highly accurate 00:10:35.980 --> 00:10:38.679 leading indicator of human attention. It's like 00:10:38.679 --> 00:10:41.480 stacking Lego blocks of beta. Proceed the big 00:10:41.480 --> 00:10:44.899 picture. Yeah. You map out The beta points, and 00:10:44.899 --> 00:10:47.179 suddenly you see the gravitational pull between 00:10:47.179 --> 00:10:51.940 seemingly unrelated systems. If soft robotics 00:10:51.940 --> 00:10:54.440 suddenly links heavily to automated surgery, 00:10:54.879 --> 00:10:58.220 the system detects that momentum before humans 00:10:58.220 --> 00:11:01.240 even notice. The AI didn't just pick the most 00:11:01.240 --> 00:11:04.080 globally popular terms either. It identified 00:11:04.080 --> 00:11:06.639 the actual architectural building blocks of the 00:11:06.639 --> 00:11:09.559 next industrial era. Reinforcement learning took 00:11:09.559 --> 00:11:11.840 the absolute number one spot on the list. It 00:11:11.840 --> 00:11:14.659 is the algorithmic brain. behind complex sequential 00:11:14.659 --> 00:11:16.899 decision making. We use it for everything from 00:11:16.899 --> 00:11:19.299 advanced drug design to drone racing. Yeah. The 00:11:19.299 --> 00:11:21.840 AI learns by ruthless trial and error at massive 00:11:21.840 --> 00:11:24.019 scale. Blockchain came in at number two on the 00:11:24.019 --> 00:11:26.820 list, but it is moving far beyond simple cryptocurrency 00:11:26.820 --> 00:11:29.980 trading now. AI flagged its immense momentum 00:11:29.980 --> 00:11:33.480 in secure swarm learning networks. Let's explore 00:11:33.480 --> 00:11:35.539 how swarm learning actually works in practice. 00:11:35.759 --> 00:11:38.919 It allows competing hospitals to train AI on 00:11:38.919 --> 00:11:41.669 medical data securely. Right. They don't even 00:11:41.669 --> 00:11:43.370 have to share their highly sensitive patient 00:11:43.370 --> 00:11:46.070 files. The deep learning happens locally on their 00:11:46.070 --> 00:11:49.070 own secure servers. And only the mathematical 00:11:49.070 --> 00:11:51.370 insights are shared across the broader network. 00:11:51.629 --> 00:11:54.289 We are also seeing a massive undeniable surge 00:11:54.289 --> 00:11:57.950 in hardware trends. 3D printing and soft robotics 00:11:57.950 --> 00:12:00.370 are the fastest growing physical categories. 00:12:00.769 --> 00:12:02.970 The physical manufacturing front is finally catching 00:12:02.970 --> 00:12:05.909 up to the software. Augmented reality is finally 00:12:05.909 --> 00:12:08.399 crossing the true momentum threshold too. It 00:12:08.399 --> 00:12:10.759 is officially moving from a novelty gimmick to 00:12:10.759 --> 00:12:13.580 mass utility. The underlying data shows people 00:12:13.580 --> 00:12:16.039 are actually searching for practical daily applications 00:12:16.039 --> 00:12:19.059 now. Beat. But I have to question the real world 00:12:19.059 --> 00:12:22.100 value of this data. Does being just three months 00:12:22.100 --> 00:12:25.100 early on a tech trend really matter? In the ruthless 00:12:25.100 --> 00:12:27.460 corporate world, three months is an absolute 00:12:27.460 --> 00:12:29.799 eternity. It gives you the necessary time to 00:12:29.799 --> 00:12:32.340 file patents and acquire startups. You can secure 00:12:32.340 --> 00:12:34.899 specialized engineering talent before your competitors 00:12:34.899 --> 00:12:37.840 even wake up. In tech, three months early is 00:12:37.840 --> 00:12:40.159 the difference between leading and losing. Exactly. 00:12:40.320 --> 00:12:43.299 It is the ultimate predictive competitive advantage 00:12:43.299 --> 00:12:46.019 in global business. Let's step back and look 00:12:46.019 --> 00:12:48.639 at the whole picture today. We are crossing a 00:12:48.639 --> 00:12:51.519 very distinct irreversible threshold right now. 00:12:51.769 --> 00:12:55.210 We really are. GPT -5 .5 is actively becoming 00:12:55.210 --> 00:12:58.570 the ultimate super app for our daily work. Google's 00:12:58.570 --> 00:13:01.970 core code base has become 75 % automated by machines. 00:13:02.529 --> 00:13:05.549 And AI is even predicting our future just by 00:13:05.549 --> 00:13:08.049 reading human encyclopedias. It is completely 00:13:08.049 --> 00:13:10.429 incredible to watch it all unfold in real time. 00:13:10.629 --> 00:13:13.549 AI is no longer just a passive digital tool you 00:13:13.549 --> 00:13:16.129 use occasionally. No, it's not. It is an active, 00:13:16.210 --> 00:13:19.009 autonomous participant in our shared physical 00:13:19.009 --> 00:13:21.320 reality. It is making independent... decisions 00:13:21.320 --> 00:13:23.679 and building the infrastructure around us. Don't 00:13:23.679 --> 00:13:25.100 forget to check our show notes for you today. 00:13:25.179 --> 00:13:28.059 We linked the Stanford AI lectures and that crazy 00:13:28.059 --> 00:13:31.519 ping pong match video. You can also find the 00:13:31.519 --> 00:13:34.620 full fascinating Momentum 100 list there. Cosmos 00:13:34.620 --> 00:13:37.519 1 .0 predicted the next industrial era just by 00:13:37.519 --> 00:13:40.500 reading human Wikipedia pages. But what is it 00:13:40.500 --> 00:13:42.059 going to predict when it starts analyzing the 00:13:42.059 --> 00:13:44.200 75 % of the Internet's code that AI is currently 00:13:44.200 --> 00:13:47.220 writing for us? Two sec silence. Something for 00:13:47.220 --> 00:13:48.899 you to chew on today. OT Road Music.