WEBVTT 00:00:00.000 --> 00:00:03.319 The fundamental belief that bigger models are 00:00:03.319 --> 00:00:07.320 always smarter models. It's hitting a wall. It 00:00:07.320 --> 00:00:10.599 is. And for the pioneers of AI and this realization, 00:00:10.980 --> 00:00:13.240 it seems to change everything about the path 00:00:13.240 --> 00:00:15.859 toward true superintelligence. It's a seismic 00:00:15.859 --> 00:00:18.320 shift, really. And this critique, you know, it 00:00:18.320 --> 00:00:20.280 isn't coming from some academic paper. Right. 00:00:20.359 --> 00:00:22.699 This is straight from Ilya Sutskiver, a co -founder 00:00:22.699 --> 00:00:25.719 of OpenAI. And that signals a huge, expensive 00:00:25.719 --> 00:00:28.120 change in how the industry is going to approach 00:00:28.120 --> 00:00:31.609 AI. Welcome back to the Deep Dive. Today, we're 00:00:31.609 --> 00:00:35.429 unpacking a really dense digest of AI developments 00:00:35.429 --> 00:00:38.109 that shows the tech is at a critical inflection 00:00:38.109 --> 00:00:40.630 point. We're stepping away from the day -to -day 00:00:40.630 --> 00:00:42.689 noise. We are, and we're focusing on the big 00:00:42.689 --> 00:00:44.670 strategy shifts. And our mission for you today 00:00:44.670 --> 00:00:48.070 is to give you a shortcut to being, well... Instantly 00:00:48.070 --> 00:00:51.509 informed on three massive connected ideas that 00:00:51.509 --> 00:00:53.210 are going to shape how you see the market. And 00:00:53.210 --> 00:00:55.710 how you use these tools. So our deep dive is 00:00:55.710 --> 00:00:58.009 in three parts. First, we're tackling why the 00:00:58.009 --> 00:00:59.869 age of scaling is ending. This whole idea of 00:00:59.869 --> 00:01:02.109 just throwing more compute and data at the problem. 00:01:02.250 --> 00:01:05.170 Exactly. According to Sutzkever and his new company, 00:01:05.310 --> 00:01:08.049 SSI. Then we pivot straight to the practical. 00:01:08.250 --> 00:01:10.609 We've got some really effective actionable tips 00:01:10.609 --> 00:01:13.030 for you. Things like immediate hacks for better 00:01:13.030 --> 00:01:15.340 prompting. And we'll look at the latest real 00:01:15.340 --> 00:01:17.340 -world performance benchmarks. And then part 00:01:17.340 --> 00:01:21.379 three, pure geopolitics. This is a big one. We 00:01:21.379 --> 00:01:24.879 are breaking down America's new $200 billion 00:01:24.879 --> 00:01:29.459 Genesis mission. A massive state -sponsored plan 00:01:29.459 --> 00:01:34.260 to basically double the nation's scientific output 00:01:34.260 --> 00:01:37.379 using AI. There's a lot of ground to cover, but 00:01:37.379 --> 00:01:38.760 we're going to make sure you walk away with the 00:01:38.760 --> 00:01:40.879 most important insights. Okay, let's get into 00:01:40.879 --> 00:01:43.180 it. This shift in philosophy, we have to start 00:01:43.180 --> 00:01:45.540 with Ilya Sutskever. He was absolutely central 00:01:45.540 --> 00:01:48.420 to GPT -4. Central. And now he's left to launch 00:01:48.420 --> 00:01:51.700 Safe Super Intelligence Inc., SSI. And this is 00:01:51.700 --> 00:01:54.459 such a huge statement. Setskiver's core assessment 00:01:54.459 --> 00:01:58.280 is that the period from, say, 2020 to 2025. The 00:01:58.280 --> 00:02:01.280 period that gave us GPT -4, Claude 3, Gemini. 00:02:01.400 --> 00:02:03.959 Right. That whole era was just defined by the 00:02:03.959 --> 00:02:06.620 mantra, bigger models, bigger clusters. Everything. 00:02:06.819 --> 00:02:09.259 And the data is now showing we're hitting diminishing 00:02:09.259 --> 00:02:11.539 returns on that. So all that massive investment. 00:02:11.840 --> 00:02:14.379 Yeah. You know, just throwing more data and more 00:02:14.379 --> 00:02:16.560 compute at the existing architecture, it's just. 00:02:16.680 --> 00:02:19.139 not giving that proportional leap in capability 00:02:19.139 --> 00:02:22.379 anymore so the architecture itself is the bottleneck 00:02:22.379 --> 00:02:25.139 i love the analogy and the sources the lego blocks 00:02:25.139 --> 00:02:27.300 yeah the lego blocks we're not just stacking 00:02:27.300 --> 00:02:30.639 more blocks of data To make a real leap, we need 00:02:30.639 --> 00:02:32.719 to change the design of the blocks themselves. 00:02:33.060 --> 00:02:35.379 That's it. Exactly. These current transformer 00:02:35.379 --> 00:02:37.939 models, I mean, they're brilliant at pattern 00:02:37.939 --> 00:02:41.460 recognition. Incredible. But they lack the architectural 00:02:41.460 --> 00:02:45.800 basis for complex emergent reasoning, that spontaneous 00:02:45.800 --> 00:02:48.080 critical thinking. It's almost like they have 00:02:48.080 --> 00:02:49.939 perfect memory, but they can't really have a 00:02:49.939 --> 00:02:51.930 novel thought. That's a great way to put it. 00:02:51.969 --> 00:02:54.610 So what's needed is entirely new fundamental 00:02:54.610 --> 00:02:57.550 research. And the sources point to three areas 00:02:57.550 --> 00:03:00.469 SSI is focusing on for that next leap. Right. 00:03:00.550 --> 00:03:03.110 They need new learning architectures, first of 00:03:03.110 --> 00:03:06.389 all, then better algorithms and safety frameworks 00:03:06.389 --> 00:03:08.729 that are baked in from the start. Not just bolted 00:03:08.729 --> 00:03:11.610 on at the end. Not bolted on. And finally, fundamentally 00:03:11.610 --> 00:03:14.629 new forms of internal reasoning, sort of like 00:03:14.629 --> 00:03:18.370 how humans connect ideas that seem unrelated. 00:03:18.509 --> 00:03:21.939 That sounds like a... I mean a... profound technological 00:03:21.939 --> 00:03:25.939 risk. Yeah. Why is Sutzkever so convinced that 00:03:25.939 --> 00:03:29.340 just sheer scale won't get us to superintelligence? 00:03:29.560 --> 00:03:31.719 It's deeper than just, you know, the limitations 00:03:31.719 --> 00:03:34.500 of the attention mechanism. It's about that mechanism's 00:03:34.500 --> 00:03:38.340 reliance on correlation, not causation. Oh, OK. 00:03:38.439 --> 00:03:40.400 The current models, they struggle to model the 00:03:40.400 --> 00:03:44.039 real world in a continuous causal way. You can 00:03:44.039 --> 00:03:46.800 scale it infinitely and it'll still make weird 00:03:46.800 --> 00:03:49.180 factual errors because it's built on probability, 00:03:49.439 --> 00:03:52.939 not logic. So SSI is rejecting the industry standard. 00:03:53.139 --> 00:03:55.360 Completely. They're operating like a secret research 00:03:55.360 --> 00:03:58.780 lab. The model is Bell Labs from the mid -20th 00:03:58.780 --> 00:04:00.639 century. And they turned down an acquisition 00:04:00.639 --> 00:04:03.599 offer from Meta. They did, which signals they're 00:04:03.599 --> 00:04:05.740 serious about this long game approach. They are 00:04:05.740 --> 00:04:07.639 deliberately stepping out of the race to just 00:04:07.639 --> 00:04:10.139 build bigger models. It takes some serious conviction 00:04:10.139 --> 00:04:12.180 to walk away from that compute race. Yeah. The 00:04:12.180 --> 00:04:14.479 scale we're talking about is almost, it's hard 00:04:14.479 --> 00:04:18.000 to comprehend. Whoa. Yeah. Imagine having access 00:04:18.000 --> 00:04:20.759 to the compute to scale up to a billion queries 00:04:20.759 --> 00:04:24.220 only to realize that money is better spent on 00:04:24.220 --> 00:04:26.300 a totally different approach. That decision. 00:04:27.079 --> 00:04:29.079 It's hard to wrap your head around. It really 00:04:29.079 --> 00:04:31.220 is. It suggests the financial incentives just 00:04:31.220 --> 00:04:33.259 don't align with the scientific goal anymore. 00:04:33.500 --> 00:04:36.879 So the core question driving him is, is the future 00:04:36.879 --> 00:04:39.620 in massive scale or in foundational research? 00:04:40.189 --> 00:04:42.050 And for him, it's foundational research. It's 00:04:42.050 --> 00:04:45.569 about novel algorithms over pure size. That makes 00:04:45.569 --> 00:04:49.189 perfect sense. Okay, let's shift gears now from 00:04:49.189 --> 00:04:51.740 the theoretical architecture of tomorrow. to 00:04:51.740 --> 00:04:54.220 what you can use right now absolutely while the 00:04:54.220 --> 00:04:56.259 titans fight over architecture you can get way 00:04:56.259 --> 00:04:59.399 better outputs today there was this list of 21 00:04:59.399 --> 00:05:03.060 idea hacks for chat gpt but one technique really 00:05:03.060 --> 00:05:05.480 stood out and this is a key technique that what 00:05:05.480 --> 00:05:08.180 99 of users are missing pretty much it's called 00:05:08.180 --> 00:05:10.459 prompts distillation okay so it's the process 00:05:10.459 --> 00:05:14.259 of taking your huge complex sometimes rambling 00:05:14.259 --> 00:05:16.480 prompts we've all written them five paragraphs 00:05:16.480 --> 00:05:18.959 of context exactly and you ruthlessly refine 00:05:18.959 --> 00:05:21.759 them down into these hyper focused instruction 00:05:21.759 --> 00:05:24.379 sets. And that refinement improves the output 00:05:24.379 --> 00:05:27.160 because it just removes all the noise. It forces 00:05:27.160 --> 00:05:29.720 the model into specific constraints. That's it. 00:05:29.819 --> 00:05:32.560 So instead of typing, write me a blog post about 00:05:32.560 --> 00:05:34.779 LLMs in the stock market, make it optimistic, 00:05:34.860 --> 00:05:37.660 but with risks. Which is so vague. So vague, 00:05:37.699 --> 00:05:40.310 you distill it too. Roll. Financial analyst. 00:05:40.550 --> 00:05:44.189 Task. Analyze LLM impact on NASDAQ Q futures. 00:05:44.470 --> 00:05:47.209 Output. Three data -supported forecast points. 00:05:47.649 --> 00:05:52.689 Constraint. 350 words. Tone. Measured. Wow. That 00:05:52.689 --> 00:05:55.009 is immediately applicable. You define the role, 00:05:55.069 --> 00:05:57.819 the task, the exact constraints. And it's free 00:05:57.819 --> 00:06:00.160 discipline. We also saw some really interesting 00:06:00.160 --> 00:06:02.740 viral hacks in the creative space. You mean the 00:06:02.740 --> 00:06:05.660 ability to pull the exact prompt from an image 00:06:05.660 --> 00:06:08.959 online? Yes. To literally steal a viral ad style 00:06:08.959 --> 00:06:11.160 instantly. That feels like a crucial tool for 00:06:11.160 --> 00:06:13.360 marketers now. It is. The image -to -text models 00:06:13.360 --> 00:06:15.420 are getting so good they can reverse engineer 00:06:15.420 --> 00:06:18.060 the style, not just the objects. You see a look 00:06:18.060 --> 00:06:19.879 you like, you can grab the blueprint. Saves months 00:06:19.879 --> 00:06:22.420 of testing. Huge accelerant. And we can't forget 00:06:22.420 --> 00:06:24.420 the high bar set on the technical side, like 00:06:24.420 --> 00:06:26.870 that viral YC demo from Carpathy on... building 00:06:26.870 --> 00:06:29.870 apps just by prompting. Still a classic. A masterclass 00:06:29.870 --> 00:06:32.350 in optimization. It's remarkable how fast these 00:06:32.350 --> 00:06:35.569 models are learning, too. Which brings us to 00:06:35.569 --> 00:06:38.550 the benchmarks. They're getting scary. I saw 00:06:38.550 --> 00:06:42.829 this one. Claude Opus 4 .5. It took Anthropic's 00:06:42.829 --> 00:06:45.529 real performance engineer take -home exam. The 00:06:45.529 --> 00:06:48.029 actual exam they give to human applicants. And 00:06:48.029 --> 00:06:50.629 it beat every single human who ever applied for 00:06:50.629 --> 00:06:53.959 the job. Not just passed. surpassed them. That 00:06:53.959 --> 00:06:56.060 is genuinely staggering. Yeah. But I do have 00:06:56.060 --> 00:06:58.920 to wonder, is a take home exam really comparable 00:06:58.920 --> 00:07:01.800 to, you know, dynamic real world engineering? 00:07:01.959 --> 00:07:04.670 That's the core tension, right? The models are 00:07:04.670 --> 00:07:07.370 optimized for benchmarks like exams. The performance 00:07:07.370 --> 00:07:09.470 is real, but it doesn't fully answer the question 00:07:09.470 --> 00:07:11.930 of their skill with totally unstructured corporate 00:07:11.930 --> 00:07:14.290 problems. Right. It strengthens the scale argument 00:07:14.290 --> 00:07:16.350 for now, but Sutskever would say that gap is 00:07:16.350 --> 00:07:18.509 still there. Exactly. And the competition in 00:07:18.509 --> 00:07:21.050 the rankings is just intense. Gemini 3, Claude 00:07:21.050 --> 00:07:25.529 4 .5, GPT 5 .1, Grok 4 .1. They're leapfrogging 00:07:25.529 --> 00:07:27.550 each other every week. And then there's the pure 00:07:27.550 --> 00:07:30.420 spectacle, the Musk challenge. Oh, yeah. Grock 00:07:30.420 --> 00:07:33.060 5 is slated to play against Faker and T1. The 00:07:33.060 --> 00:07:35.420 League of Legends world champions. That is the 00:07:35.420 --> 00:07:38.519 ultimate man versus machine test in a super complex 00:07:38.519 --> 00:07:40.639 environment. It's going to be a wild thing to 00:07:40.639 --> 00:07:43.720 watch. But I have to admit, on a personal note, 00:07:43.839 --> 00:07:46.639 I still wrestle with Prompt Drift myself. Oh, 00:07:46.639 --> 00:07:48.779 for sure. When I'm trying those complex hacks, 00:07:48.920 --> 00:07:51.779 just maintaining consistency session to session 00:07:51.779 --> 00:07:55.120 is tough. It takes real discipline to stay distilled. 00:07:55.160 --> 00:07:57.920 It absolutely requires vigilance. Which brings 00:07:57.920 --> 00:08:00.459 us, I think, to the serious ethical and legal 00:08:00.459 --> 00:08:02.680 backdrop here. Because the human consequences 00:08:02.680 --> 00:08:05.480 are becoming very clear. We saw a mention of 00:08:05.480 --> 00:08:07.779 a really difficult legal filing about accountability. 00:08:08.139 --> 00:08:10.839 We did. And we have to note, just neutrally, 00:08:10.939 --> 00:08:14.079 this ongoing legal matter with OpenAI. Their 00:08:14.079 --> 00:08:16.600 defense strategy regarding the tragic death of... 00:08:16.589 --> 00:08:18.910 of a 16 -year -old who died by suicide after 00:08:18.910 --> 00:08:21.470 manipulating chat GPT. They blame the user. They 00:08:21.470 --> 00:08:23.949 blame the user. And it just raises these profound, 00:08:24.189 --> 00:08:26.709 immediate questions about model control responsibility. 00:08:27.430 --> 00:08:30.490 The gap between capability and liability is huge. 00:08:30.790 --> 00:08:32.889 The pace of capability is so fast, the legal 00:08:32.889 --> 00:08:36.029 guardrails are moving so slow. Exactly. And meanwhile, 00:08:36.289 --> 00:08:39.129 the investment keeps flooding in. We saw that 00:08:39.129 --> 00:08:42.070 Carl Reina's Bobab Ventures raised, what, $12 00:08:42.070 --> 00:08:45.210 .9 million for robotics startups? The money is 00:08:45.210 --> 00:08:47.330 still flowing, even with all the ethical questions. 00:08:47.610 --> 00:08:50.110 Right. So considering Claude's engineering skills 00:08:50.110 --> 00:08:52.549 on one hand and these serious legal issues on 00:08:52.549 --> 00:08:56.190 the other, how quickly is the AI ethics conversation 00:08:56.190 --> 00:09:00.269 really evolving? Capability is outpacing the 00:09:00.269 --> 00:09:04.179 ethical and legal guardrails dramatically. The 00:09:04.179 --> 00:09:07.559 consequences of that gap are they're no longer 00:09:07.559 --> 00:09:09.720 hypothetical. That feels like the defining tension 00:09:09.720 --> 00:09:12.580 of this year. It really does. OK, let's pivot 00:09:12.580 --> 00:09:14.779 to our final segment. This massive government 00:09:14.779 --> 00:09:17.779 response to all of this. The U .S. Genesis mission. 00:09:18.120 --> 00:09:19.940 This is a geopolitical development we absolutely 00:09:19.940 --> 00:09:22.299 have to focus on. The Genesis mission is the 00:09:22.299 --> 00:09:24.659 code name for a new executive order. And they're 00:09:24.659 --> 00:09:26.679 framing it as a scientific Manhattan project. 00:09:26.960 --> 00:09:29.639 They are. The goal is incredibly ambitious. Use 00:09:29.639 --> 00:09:32.220 AI to double America's scientific productivity 00:09:32.220 --> 00:09:35.100 fast. So national security through scientific 00:09:35.100 --> 00:09:37.039 dominance. What are the key components they're 00:09:37.039 --> 00:09:39.639 building to do that? The heart of it is the American 00:09:39.639 --> 00:09:42.639 science and security platform. It's led by Energy 00:09:42.639 --> 00:09:45.230 Secretary Chris Wright. And it's leveraging the 00:09:45.230 --> 00:09:48.350 DOE's supercomputers and quantum processors, 00:09:48.590 --> 00:09:51.590 the engine. Right. But the truly game -changing 00:09:51.590 --> 00:09:54.309 part is the operational layer. It's almost sci 00:09:54.309 --> 00:09:58.090 -fi. The robotic labs. Fully AI -controlled robotic 00:09:58.090 --> 00:10:01.129 labs. These aren't just automated systems. These 00:10:01.129 --> 00:10:04.629 labs are designed to plan, run, and analyze their 00:10:04.629 --> 00:10:07.370 own experiments based on the AI's hypotheses. 00:10:07.990 --> 00:10:10.110 The goal is to remove the human bottleneck from 00:10:10.110 --> 00:10:12.929 discovery. Totally. So these labs could run thousands 00:10:12.929 --> 00:10:15.620 of experiments. overnight to find, say, a new 00:10:15.620 --> 00:10:17.940 battery chemistry, all without a human touching 00:10:17.940 --> 00:10:20.779 anything. And they're targeting these grand challenges 00:10:20.779 --> 00:10:23.360 like clean energy and advanced materials. And 00:10:23.360 --> 00:10:26.320 the fuel for it all is data, over $200 billion 00:10:26.320 --> 00:10:29.620 worth of secure proprietary data sets. So tell 00:10:29.620 --> 00:10:31.500 us more about those data sets. What kind of knowledge 00:10:31.500 --> 00:10:33.690 is that? Why was it unavailable before? This 00:10:33.690 --> 00:10:35.830 is the really high value stuff. Government held 00:10:35.830 --> 00:10:38.350 data, often classified research from national 00:10:38.350 --> 00:10:40.990 labs, defense agencies. It covers everything 00:10:40.990 --> 00:10:44.169 from climate modeling to material science. It 00:10:44.169 --> 00:10:46.470 was all siloed before because of security concerns. 00:10:46.710 --> 00:10:49.610 But Genesis is consolidating it all into one 00:10:49.610 --> 00:10:52.129 secure platform. This isn't just about making 00:10:52.129 --> 00:10:54.309 research faster. This is about establishing national 00:10:54.309 --> 00:10:57.450 control over a strategic knowledge base. And 00:10:57.450 --> 00:10:59.789 that security extends to manufacturing. They're 00:10:59.789 --> 00:11:02.070 building digital twins. Full simulations. Of 00:11:02.070 --> 00:11:04.600 complex. supply chains and factories so they 00:11:04.600 --> 00:11:07.700 can model changes in real time, protect infrastructure. 00:11:07.779 --> 00:11:10.519 The whole effort is being coordinated at a very 00:11:10.519 --> 00:11:13.440 high level. Michael Kratzios, former U .S. CTO, 00:11:13.440 --> 00:11:15.899 is leading it. And they're building on partnerships 00:11:15.899 --> 00:11:20.559 with OpenAI, Google, Palantir. All the big players. 00:11:20.759 --> 00:11:23.519 The timeline is what gets me. It is incredibly 00:11:23.519 --> 00:11:25.779 aggressive. It underscores the urgency. They 00:11:25.779 --> 00:11:28.259 gave themselves 60 days to list more than 20 00:11:28.259 --> 00:11:31.379 grand challenges to attack first. 60 days to 00:11:31.379 --> 00:11:33.840 define the research agenda for a $200 billion 00:11:33.840 --> 00:11:37.299 project. It's rapid. Then 90 days to inventory 00:11:37.299 --> 00:11:39.620 all the compute resources in the country, public 00:11:39.620 --> 00:11:42.840 and private. And then, and this is the big one, 00:11:42.919 --> 00:11:45.899 270 days. Less than a year. To prove it works 00:11:45.899 --> 00:11:49.539 with one real scientific use case, they are moving 00:11:49.539 --> 00:11:53.279 at wartime speed. So what's the core existential 00:11:53.279 --> 00:11:56.779 motivation behind building this massive proprietary 00:11:56.779 --> 00:11:59.700 government platform instead of just relying on 00:11:59.700 --> 00:12:02.230 the private sector? The U .S. is securing its 00:12:02.230 --> 00:12:04.190 scientific future by establishing proprietary 00:12:04.190 --> 00:12:07.129 control over the most valuable data, compute, 00:12:07.330 --> 00:12:09.669 and research infrastructure. So we've covered 00:12:09.669 --> 00:12:12.409 two massive, fundamentally different pivots today. 00:12:12.549 --> 00:12:14.929 On one hand, you have this deep, quiet research 00:12:14.929 --> 00:12:18.769 shift away from scaling led by Sutzkever. A kind 00:12:18.769 --> 00:12:20.690 of foundational revolution. And on the other, 00:12:20.830 --> 00:12:23.850 you have this monumental government mobilization. 00:12:23.929 --> 00:12:26.970 The Genesis Mission, a state -sponsored sprint. 00:12:27.500 --> 00:12:29.919 to industrialize science itself. And both of 00:12:29.919 --> 00:12:32.120 these shifts show that AI progress is moving 00:12:32.120 --> 00:12:34.519 way beyond simple metrics like parameter count. 00:12:34.720 --> 00:12:37.580 It's an exciting and frankly, a little unnerving 00:12:37.580 --> 00:12:39.399 time to be paying attention. Absolutely. The 00:12:39.399 --> 00:12:41.539 age of building bigger is giving way to the age 00:12:41.539 --> 00:12:43.600 of building smarter. Whether that's through pure 00:12:43.600 --> 00:12:46.700 research or massive state -directed mobilization. 00:12:46.919 --> 00:12:48.639 So here's a final thought for you to chew on. 00:12:49.299 --> 00:12:52.539 Which strategy is more likely to yield the next 00:12:52.539 --> 00:12:57.240 truly big breakthrough? Is it the focused, quiet 00:12:57.240 --> 00:13:01.039 Bell Labs approach of SSI, which risks falling 00:13:01.039 --> 00:13:04.100 behind on compute? Or is it the state -sponsored, 00:13:04.100 --> 00:13:06.340 high -resource Manhattan Project of the Genesis 00:13:06.340 --> 00:13:09.039 mission, which, you know, risks being too rigid 00:13:09.039 --> 00:13:10.840 and top -down? Something to think about. And 00:13:10.840 --> 00:13:13.179 while you think about that, definitely go try 00:13:13.179 --> 00:13:15.740 some of those prompt distillation hacks. That 00:13:15.740 --> 00:13:17.940 skill alone will boost your productivity today. 00:13:18.080 --> 00:13:20.320 And keep an eye on these geopolitical investments. 00:13:20.539 --> 00:13:22.440 They're going to define global science for the 00:13:22.440 --> 00:13:24.870 next decade. Thank you for sharing your sources 00:13:24.870 --> 00:13:26.649 with us for this deep dive. We'll be tracking 00:13:26.649 --> 00:13:28.769 all of it. Until next time. Stay curious.