WEBVTT 00:00:00.000 --> 00:00:02.379 Think about the assumptions we make every single 00:00:02.379 --> 00:00:06.440 day. Beat, we spent decades thinking AI was just 00:00:06.440 --> 00:00:08.679 a chat bot. Yeah, just a simple digital box. 00:00:08.839 --> 00:00:10.500 Right, you type a prompt, you get text back. 00:00:10.679 --> 00:00:12.939 But what happens when that chat bot finally stands 00:00:12.939 --> 00:00:15.900 up? It starts acting like a digital employee. 00:00:16.260 --> 00:00:19.589 Exactly. or when it autonomously solves 80 -year 00:00:19.589 --> 00:00:22.629 -old math problems. These are puzzles that baffled 00:00:22.629 --> 00:00:25.929 the absolute brightest human geniuses. It completely 00:00:25.929 --> 00:00:28.010 changes our relationship with this technology. 00:00:28.269 --> 00:00:29.969 I mean, the ground is shifting right beneath 00:00:29.969 --> 00:00:32.189 our feet. Welcome to this deep dive for you today. 00:00:32.700 --> 00:00:35.640 We are looking at a stack of incredibly fresh 00:00:35.640 --> 00:00:38.320 sources. Really exciting stuff. They map out 00:00:38.320 --> 00:00:41.539 the next era of artificial intelligence. We are 00:00:41.539 --> 00:00:44.020 going to trace the massive hardware shift happening 00:00:44.020 --> 00:00:47.479 right now. We'll explore the mind bending billions 00:00:47.479 --> 00:00:50.280 trading hands between top labs. We'll examine 00:00:50.280 --> 00:00:52.659 the tools turning AI into proactive workers. 00:00:52.880 --> 00:00:54.640 And finally, we'll look at a historic breakthrough 00:00:54.640 --> 00:00:56.939 in autonomous reasoning. It is a fascinating 00:00:56.939 --> 00:00:59.820 blueprint of what comes next. So to truly understand 00:00:59.820 --> 00:01:02.630 how AI is getting smarter. We have to start physically. 00:01:02.829 --> 00:01:05.150 Right at the foundation. Yeah. We have to look 00:01:05.150 --> 00:01:07.670 at the underlying hardware. And we really need 00:01:07.670 --> 00:01:09.590 to follow the money moving behind it. Right. 00:01:09.670 --> 00:01:11.590 And the numbers here are absolutely staggering. 00:01:12.090 --> 00:01:15.010 NVIDIA just posted another record -breaking financial 00:01:15.010 --> 00:01:17.230 quarter. What are we looking at? They reported 00:01:17.230 --> 00:01:22.250 $81 .6 billion in revenue. That is up 20 % from 00:01:22.250 --> 00:01:24.629 the previous quarter. That is a massive leap 00:01:24.629 --> 00:01:27.540 in just... Three months. Yeah, it really is. 00:01:27.620 --> 00:01:31.579 Their data center business alone hit $75 .2 billion. 00:01:31.859 --> 00:01:34.920 Wow. That shows infrastructure demand is still 00:01:34.920 --> 00:01:37.680 incredibly strong. The physical foundation is 00:01:37.680 --> 00:01:39.920 still being rapidly built. I want to pause on 00:01:39.920 --> 00:01:43.659 the financial flexing here. They approved $80 00:01:43.659 --> 00:01:46.879 billion in share buybacks. What does that signal 00:01:46.879 --> 00:01:49.980 to the broader tech market? Well, it signals 00:01:49.980 --> 00:01:52.840 massive internal confidence in their future growth. 00:01:53.180 --> 00:01:55.060 They are telling the market they have plenty 00:01:55.060 --> 00:01:57.290 of cash. Right. But what's even more interesting 00:01:57.290 --> 00:01:59.109 is their venture capital strategy. They hold 00:01:59.109 --> 00:02:02.090 $43 billion in private startup stakes right now. 00:02:02.230 --> 00:02:06.329 Exactly. They bought $18 .5 billion this quarter 00:02:06.329 --> 00:02:09.210 alone. That is wild. They aren't just selling 00:02:09.210 --> 00:02:11.870 chips to random companies anymore. They are actively 00:02:11.870 --> 00:02:15.289 funding the exact startups that buy their hardware. 00:02:15.490 --> 00:02:18.050 It creates this massive self -sustaining financial 00:02:18.050 --> 00:02:21.219 flywheel. Yeah, it does. That's brilliant, but 00:02:21.219 --> 00:02:23.699 it can't accelerate forever, right? No, probably 00:02:23.699 --> 00:02:27.099 not. Forecasts predict $91 billion for the next 00:02:27.099 --> 00:02:29.759 quarter, but their overall growth rate may slow 00:02:29.759 --> 00:02:34.099 to 12%. Plus, their China export revenue remains 00:02:34.099 --> 00:02:37.039 highly uncertain right now. That brings us to 00:02:37.039 --> 00:02:40.060 the biggest pivot in their strategy. Jensen Huang 00:02:40.060 --> 00:02:43.979 just announced a new $200 billion market. This 00:02:43.979 --> 00:02:46.539 is huge. It centers entirely around a new CPU 00:02:46.539 --> 00:02:49.590 called Vera. This is a fundamental shift in computing 00:02:49.590 --> 00:02:52.729 architecture. Vera is built specifically for 00:02:52.729 --> 00:02:55.530 agentic AI workflows. Let's define that term 00:02:55.530 --> 00:02:58.229 for everyone listening. Agentic AI is AI that 00:02:58.229 --> 00:03:00.310 takes actions independently to achieve a goal. 00:03:00.469 --> 00:03:03.110 Perfect definition. Beat. Think of this shift 00:03:03.110 --> 00:03:06.169 like a busy corporate office. The GPUs are the 00:03:06.169 --> 00:03:08.469 creative brainstormers in the room. Right. They 00:03:08.469 --> 00:03:11.680 generate huge ideas and process heavy data. Exactly. 00:03:11.860 --> 00:03:15.560 But CPUs like Vera are the actual project managers. 00:03:15.840 --> 00:03:18.520 They're the ones executing the complex daily 00:03:18.520 --> 00:03:21.860 tasks. That analogy is spot on. NVIDIA isn't 00:03:21.860 --> 00:03:24.699 just selling GPUs to researchers anymore. They 00:03:24.699 --> 00:03:28.159 want to power the full AI stack entirely. From 00:03:28.159 --> 00:03:31.240 massive data centers down to proactive autonomous 00:03:31.240 --> 00:03:34.419 agents. Huang says Vera has already generated 00:03:34.419 --> 00:03:38.099 $20 billion in sales. That is just in this current 00:03:38.099 --> 00:03:40.439 year alone. It proves the market demand. for 00:03:40.439 --> 00:03:43.800 agentic hardware is real. Future AI agents will 00:03:43.800 --> 00:03:46.780 deeply need these CPU -like systems. Why is that? 00:03:47.020 --> 00:03:49.319 Because they need them to use external tools 00:03:49.319 --> 00:03:52.080 efficiently. They have to run sequential tasks 00:03:52.080 --> 00:03:54.599 without constant human intervention. How does 00:03:54.599 --> 00:03:56.939 NVIDIA stepping into CPUs affect the rest of 00:03:56.939 --> 00:03:59.020 the industry? Well, they aren't just selling 00:03:59.020 --> 00:04:01.120 individual processing parts anymore. They want 00:04:01.120 --> 00:04:03.780 to own the entire operational architecture. Right. 00:04:03.860 --> 00:04:06.199 They are building the infrastructure for autonomous 00:04:06.199 --> 00:04:08.560 digital workers. So they own the brains and the 00:04:08.560 --> 00:04:10.879 hands now. Exactly. They are the foundation of 00:04:10.879 --> 00:04:12.939 the new economy. But if they are supplying the 00:04:12.939 --> 00:04:15.659 brains and the hands the energy required to power, 00:04:15.900 --> 00:04:18.920 that must be astronomical. Oh, it is. Who is 00:04:18.920 --> 00:04:22.300 actually footing that massive computing bill? 00:04:22.500 --> 00:04:25.439 The operational costs are becoming almost incomprehensible 00:04:25.439 --> 00:04:28.069 today. Let's look at Excel. for a very clear 00:04:28.069 --> 00:04:31.430 example. Okay. They burned $6 .4 billion last 00:04:31.430 --> 00:04:34.470 year. That was on just $3 .2 billion in revenue. 00:04:34.629 --> 00:04:36.910 The math there is incredibly brutal to look at. 00:04:36.990 --> 00:04:40.550 It is. And a recent SpaceX IPO filing revealed 00:04:40.550 --> 00:04:43.550 something even wilder. What was it? It detailed 00:04:43.550 --> 00:04:45.850 Grok's plan for a multiple trillion parameter 00:04:45.850 --> 00:04:48.699 model. A trillion parameter model has a trillion 00:04:48.699 --> 00:04:51.879 adjustable internal data connections. Yeah. That 00:04:51.879 --> 00:04:54.360 ambitious plan will make the compute bill substantially 00:04:54.360 --> 00:04:57.699 bigger. But here is the wildest twist in our 00:04:57.699 --> 00:04:59.360 sources. I know what you're going to say. Anthropic 00:04:59.360 --> 00:05:02.100 is reportedly paying XAI directly for compute 00:05:02.100 --> 00:05:06.079 power. They are paying $1 .5 billion to Sex Silence 00:05:06.079 --> 00:05:09.100 per month. Yeah. That is for access to Colossus 00:05:09.100 --> 00:05:13.500 Compute through May 2029. Whoa. Imagine scaling 00:05:13.500 --> 00:05:16.600 to a billion queries. The power required to run 00:05:16.600 --> 00:05:19.339 an entire data center just for one model. That 00:05:19.339 --> 00:05:21.639 is the reality of an industrial revolution scale. 00:05:21.839 --> 00:05:24.620 You are paying for miles of servers and gigawatts 00:05:24.620 --> 00:05:28.040 of power. It is a staggeringly expensive physical 00:05:28.040 --> 00:05:30.680 footprint. Yet despite these massive infrastructure 00:05:30.680 --> 00:05:34.019 costs, Anthropic expects its first profitable 00:05:34.019 --> 00:05:36.819 quarter very soon. Right. Their revenue is... 00:05:36.839 --> 00:05:40.420 jumping to around $10 .9 billion. That is a massive 00:05:40.420 --> 00:05:43.439 milestone for the entire tech industry. It proves 00:05:43.439 --> 00:05:45.399 the underlying business model can actually work. 00:05:45.560 --> 00:05:47.639 It dropped right alongside rumors of an open 00:05:47.639 --> 00:05:51.459 AI IPO. The market is maturing incredibly fast 00:05:51.459 --> 00:05:54.120 right now. Extremely fast. How does a company 00:05:54.120 --> 00:05:57.839 possibly survive burning billions a year? It's 00:05:57.839 --> 00:05:59.980 a massive infrastructure land grab right now. 00:06:00.060 --> 00:06:02.959 The sheer utility of the resulting models is 00:06:02.959 --> 00:06:05.399 the price. Right. That utility will eventually 00:06:05.399 --> 00:06:08.540 outweigh the massive upfront cost. We see this 00:06:08.540 --> 00:06:10.879 clearly with Anthropic finally hitting profitability. 00:06:11.139 --> 00:06:13.959 We must spend billions today to make trillions 00:06:13.959 --> 00:06:16.540 tomorrow. Exactly. The initial capital expenditure 00:06:16.540 --> 00:06:19.129 is the only way forward. So what incredible utility 00:06:19.129 --> 00:06:21.730 justifies paying a billion dollars monthly? It 00:06:21.730 --> 00:06:23.990 is the fundamental shift from generating text 00:06:23.990 --> 00:06:26.810 to taking action. We are looking at the toolkit 00:06:26.810 --> 00:06:29.550 for these new digital workers. The ecosystem 00:06:29.550 --> 00:06:32.949 of agentic tools is exploding right now. And 00:06:32.949 --> 00:06:36.290 it completely changes how we interact with software. 00:06:36.550 --> 00:06:40.329 EXA just raised $250 million. That is at a $2 00:06:40.329 --> 00:06:44.370 .2 billion valuation. Yeah. They are simply building 00:06:44.370 --> 00:06:47.949 a search engine for AI agents. This is a crucial 00:06:47.949 --> 00:06:51.129 piece of the puzzle. Agents don't browse websites 00:06:51.129 --> 00:06:53.009 like human beings do. Right, they can't see the 00:06:53.009 --> 00:06:56.649 screen. Exactly. They can't navigate flashy pop 00:06:56.649 --> 00:07:00.230 -ups or read layout designs. They need raw... 00:07:00.519 --> 00:07:02.660 perfectly structured data to operate effectively. 00:07:02.860 --> 00:07:05.379 So EGZA provides that exact structured data layer 00:07:05.379 --> 00:07:07.160 for them. Yeah. And developer demand for this 00:07:07.160 --> 00:07:09.899 is just huge. Google is also making massive updates 00:07:09.899 --> 00:07:12.800 to their tool ecosystem. Gemini Omni is generating 00:07:12.800 --> 00:07:15.720 entire new video scenes now. It's amazing. It 00:07:15.720 --> 00:07:18.120 uses mixed inputs to create high quality outputs. 00:07:18.399 --> 00:07:21.180 You can combine images, audio, video, and text 00:07:21.180 --> 00:07:24.279 seamlessly. Yeah. It is completely changing the 00:07:24.279 --> 00:07:26.199 creative production pipeline forever. Google 00:07:26.199 --> 00:07:28.300 AI Studio is taking this even further today. 00:07:28.480 --> 00:07:31.300 It turns simple text prompts into native Android 00:07:31.300 --> 00:07:33.779 apps. The traditional barrier to software coding 00:07:33.779 --> 00:07:37.160 is rapidly disappearing. Then there's anti -gravity 00:07:37.160 --> 00:07:41.620 2 .0 to discuss. It is turning AI into a real 00:07:41.620 --> 00:07:44.699 developer workspace. This one is wild. They feature 00:07:44.699 --> 00:07:47.939 parallel agents and background coding tasks seamlessly. 00:07:48.360 --> 00:07:50.300 We should definitely explain what parallel agents 00:07:50.300 --> 00:07:52.980 actually do. Parallel agents are multiple AI 00:07:52.980 --> 00:07:56.120 programs working on tasks simultaneously. Beat, 00:07:56.339 --> 00:07:59.279 I still wrestle with prompt drift myself. Oh, 00:07:59.379 --> 00:08:02.459 same here. That's when AI slowly loses focus 00:08:02.459 --> 00:08:05.360 on your original instructions. So the idea of 00:08:05.360 --> 00:08:08.519 parallel agents managing whole workflows feels 00:08:08.519 --> 00:08:11.639 wild. It is a lot for human beings to adapt to. 00:08:11.720 --> 00:08:14.740 We're moving from prompting to managing a digital 00:08:14.740 --> 00:08:17.439 staff. That is probably why Andrew Nehring released 00:08:17.439 --> 00:08:20.480 a free course. It is a two and a half hour masterclass 00:08:20.480 --> 00:08:22.720 on AI prompting. Very timely. It is designed 00:08:22.720 --> 00:08:25.120 to help humans keep up with these tools. Prompting 00:08:25.120 --> 00:08:27.000 is quickly becoming a necessary professional 00:08:27.000 --> 00:08:29.100 management skill. We also have to talk about 00:08:29.100 --> 00:08:31.800 proactive actions. Gemini Spark handles proactive 00:08:31.800 --> 00:08:34.440 actions across Google products natively. Right. 00:08:34.559 --> 00:08:36.960 It can manage tasks under your direction entirely 00:08:36.960 --> 00:08:39.820 in the background. But having proactive agents 00:08:39.820 --> 00:08:42.840 doing things. automatically raises massive concerns. 00:08:43.100 --> 00:08:45.480 You need serious security frameworks before you 00:08:45.480 --> 00:08:48.879 deploy them. Absolutely. OpenAI is adopting Google's 00:08:48.879 --> 00:08:52.179 SynthID watermarking for generated images. They 00:08:52.179 --> 00:08:54.399 are also building a public verification checker 00:08:54.399 --> 00:08:56.740 very soon. We really need to know what's real 00:08:56.740 --> 00:08:59.120 and what's AI. We are seeing movement on the 00:08:59.120 --> 00:09:01.519 regulatory front, too. The European Commission 00:09:01.519 --> 00:09:04.519 released draft guidance on high -risk AI. It 00:09:04.519 --> 00:09:07.240 clarifies exactly which systems fall under stricter 00:09:07.240 --> 00:09:09.419 rules. Now we have impending U .S. executive 00:09:09.419 --> 00:09:12.659 orders coming soon. OpenAI and Anthropic are 00:09:12.659 --> 00:09:15.139 discussing voluntary frameworks as well. They 00:09:15.139 --> 00:09:17.059 are talking about model sharing directly with 00:09:17.059 --> 00:09:19.700 the government. It's all moving so fast. So how 00:09:19.700 --> 00:09:23.639 do we prevent chaos if agents are acting proactively? 00:09:24.059 --> 00:09:26.559 We have to rely on strict verification systems. 00:09:26.899 --> 00:09:29.559 The strict necessity of Google SynthEye is a 00:09:29.559 --> 00:09:31.899 vital start. Yeah. We also need those incoming 00:09:31.899 --> 00:09:35.080 EU and US regulatory frameworks. Trust requires 00:09:35.080 --> 00:09:37.840 built -in digital watermarks and strict guardrails. 00:09:37.960 --> 00:09:39.870 Right. And without those guardrails, the... enterprise 00:09:39.870 --> 00:09:42.370 utility drops to zero. We've talked extensively 00:09:42.370 --> 00:09:45.169 about the physical hardware and the money. We 00:09:45.169 --> 00:09:48.009 explored the powerful new agentic tools being 00:09:48.009 --> 00:09:52.169 built. But the ultimate test of this era is pure 00:09:52.169 --> 00:09:55.029 reasoning. Can these machine learning models 00:09:55.029 --> 00:09:57.870 actually think deeply? This is the most profound 00:09:57.870 --> 00:10:01.120 question in the field today. OpenAI just announced 00:10:01.120 --> 00:10:04.659 a truly historic mathematical breakthrough. GPT 00:10:04.659 --> 00:10:07.759 -5 .5 autonomously created a completely original 00:10:07.759 --> 00:10:10.840 geometric proof. It's incredible. It disproves 00:10:10.840 --> 00:10:14.879 a famous geometry conjecture by Paul Erdos. He 00:10:14.879 --> 00:10:18.919 proposed this specific riddle back in 1946. This 00:10:18.919 --> 00:10:21.299 mathematical problem stood unsolved for nearly 00:10:21.299 --> 00:10:23.919 80 years. Mathematicians strongly believed the 00:10:23.919 --> 00:10:26.580 best solution would involve square grids. But 00:10:26.580 --> 00:10:29.419 the AI model found an entirely new geom... construction. 00:10:29.860 --> 00:10:31.980 Right. Crucially, this wasn't a specialized math 00:10:31.980 --> 00:10:34.379 tool at all. It was a general purpose reasoning 00:10:34.379 --> 00:10:36.899 model exploring the logic. It worked through 00:10:36.899 --> 00:10:39.480 the complex logical steps completely on its own. 00:10:39.559 --> 00:10:41.480 Wait, hold on. I have to push back a little bit. 00:10:41.539 --> 00:10:43.139 Okay, go ahead. Didn't this exact same thing 00:10:43.139 --> 00:10:46.200 happen late last year? OpenAI faced intense public 00:10:46.200 --> 00:10:49.100 skepticism back then. They did. People found 00:10:49.100 --> 00:10:51.740 the models were just surfacing existing human 00:10:51.740 --> 00:10:54.919 work. They weren't actually solving brand new 00:10:54.919 --> 00:10:57.879 geometric problems. Why should we believe this 00:10:57.879 --> 00:11:00.080 time is any different? Yeah, and that skepticism 00:11:00.080 --> 00:11:03.200 was completely justified back then. But this 00:11:03.200 --> 00:11:05.960 is exactly why this specific moment matters. 00:11:06.279 --> 00:11:09.720 How so? OpenAI came incredibly prepared for the 00:11:09.720 --> 00:11:12.379 public pushback this time. They brought the receipts? 00:11:12.639 --> 00:11:14.639 They absolutely did. They didn't just drop a 00:11:14.639 --> 00:11:17.580 press release. They subjected the model's output. 00:11:18.110 --> 00:11:20.950 to rigorous human peer review. They included 00:11:20.950 --> 00:11:23.549 supporting remarks from highly respected mathematicians 00:11:23.549 --> 00:11:26.389 today. People like Noga, Alon, and Melanie Wood 00:11:26.389 --> 00:11:29.070 reviewed the work. And Thomas Bloom, who runs 00:11:29.070 --> 00:11:32.570 the Erdin's Problems website itself. These experts 00:11:32.570 --> 00:11:35.090 validated that the logical steps were genuinely 00:11:35.090 --> 00:11:37.909 novel. This proves the model generated completely 00:11:37.909 --> 00:11:40.570 new mathematical logic. Wow. It didn't just memorize 00:11:40.570 --> 00:11:42.750 old data from the internet. Does this mean AI 00:11:42.750 --> 00:11:46.169 is now better at math than humans? It's not really 00:11:46.169 --> 00:11:48.129 about being quote unquote better than us. It 00:11:48.129 --> 00:11:50.809 is about AI being capable of genuine logical 00:11:50.809 --> 00:11:53.649 discovery. It works right alongside human experts 00:11:53.649 --> 00:11:57.090 to find objective truth. AI is a peer now, not 00:11:57.090 --> 00:11:59.350 just a calculator. Exactly. It fundamentally 00:11:59.350 --> 00:12:01.750 changes how scientific discovery will happen. 00:12:02.029 --> 00:12:04.850 Sponsor. Let's bring all these incredible pieces 00:12:04.850 --> 00:12:07.649 together for you. The era of the simple chatbot 00:12:07.649 --> 00:12:10.720 is truly over today. From NVIDIA's Verichips 00:12:10.720 --> 00:12:14.100 to billion -dollar compute infrastructure deals, 00:12:14.320 --> 00:12:18.120 the entire tech world is fundamentally re -architecting 00:12:18.120 --> 00:12:21.480 for proactive agents. We are arming general -purpose 00:12:21.480 --> 00:12:24.200 reasoning engines with massive hardware capabilities. 00:12:24.580 --> 00:12:26.980 They are smart enough to solve 80 -year -old 00:12:26.980 --> 00:12:29.320 math puzzles autonomously, and we are giving 00:12:29.320 --> 00:12:31.600 them tools to operate natively inside workflows. 00:12:31.919 --> 00:12:33.539 Right. They are quickly becoming our digital 00:12:33.539 --> 00:12:35.759 colleagues and active project managers. It is 00:12:35.759 --> 00:12:38.580 a profound shift in how humans interact with 00:12:38.580 --> 00:12:41.279 machines. I want to leave you with a final provocative 00:12:41.279 --> 00:12:44.480 thought today. If a general purpose AI model 00:12:44.480 --> 00:12:46.779 can look at an 80 -year -old geometry problem 00:12:46.779 --> 00:12:50.399 and find a completely new angle that human geniuses 00:12:50.399 --> 00:12:53.740 missed for decades, what longstanding assumptions 00:12:53.740 --> 00:12:56.299 in your own industry is an AI agent about to 00:12:56.299 --> 00:12:59.279 completely rewrite? It is a thrilling, slightly 00:12:59.279 --> 00:13:01.539 terrifying question to consider. Thank you for 00:13:01.539 --> 00:13:04.220 taking this deep dive with us today. ODTRO Music.