WEBVTT 00:00:00.000 --> 00:00:02.000 The biggest names in American AI, we're talking 00:00:02.000 --> 00:00:05.599 OpenAI, Google, Meta. They're basically invisible 00:00:05.599 --> 00:00:08.380 in the world's most used OpenAI models. They've 00:00:08.380 --> 00:00:11.080 just completely ceded the throne. It's an unprecedented 00:00:11.080 --> 00:00:14.000 statistical flip, and we've got the data. We're 00:00:14.000 --> 00:00:18.600 talking about 2 .2 billion global downloads in 00:00:18.600 --> 00:00:21.760 just one year. This isn't just a minor shift 00:00:21.760 --> 00:00:23.739 in the rankings, you know. It's a fundamental 00:00:23.739 --> 00:00:27.280 rebalancing of global AI leadership. The scale 00:00:27.280 --> 00:00:29.929 confirms it. Welcome back to the Deep Dive. Today, 00:00:29.989 --> 00:00:32.390 we're taking a calm, curious look at how quickly 00:00:32.390 --> 00:00:35.250 the foundational rules of the AI ecosystem are 00:00:35.250 --> 00:00:37.170 just being completely rewritten. The pace of 00:00:37.170 --> 00:00:40.009 change has never felt faster. That's right. And 00:00:40.009 --> 00:00:41.630 we're here to guide you through this. Our mission 00:00:41.630 --> 00:00:44.070 today is to really distill three critical shifts 00:00:44.070 --> 00:00:45.649 you need to understand right now. First, we're 00:00:45.649 --> 00:00:47.649 going to unpack the hard data that confirms this 00:00:47.649 --> 00:00:51.030 new dominant force in open source AI. And it's 00:00:51.030 --> 00:00:53.170 probably going to surprise you. Second, we have 00:00:53.170 --> 00:00:55.509 a practical playbook for staying visible in what 00:00:55.509 --> 00:00:57.689 we're now calling the generative engine optimization. 00:00:57.899 --> 00:01:01.640 or GEO era. And finally, we'll dive into a pure 00:01:01.640 --> 00:01:04.780 scientific breakthrough, an AI system that uses 00:01:04.780 --> 00:01:07.219 something called vibe -proving to solve mathematical 00:01:07.219 --> 00:01:09.540 problems that have been open for over 30 years. 00:01:09.700 --> 00:01:12.340 And it does it in hours, not decades. Okay, let's 00:01:12.340 --> 00:01:14.439 get into this monumental shift in the open source 00:01:14.439 --> 00:01:17.540 world because the data really is a shock. This 00:01:17.540 --> 00:01:19.980 recent study from MIT and Hugging Face tracked 00:01:19.980 --> 00:01:23.859 massive model usage and, well, it revealed something 00:01:23.859 --> 00:01:26.540 remarkable. It's the headline of the entire week. 00:01:26.989 --> 00:01:29.709 So the study tracked 2 .2 billion downloads on 00:01:29.709 --> 00:01:32.790 Hugging Face between August 2024 and August 2025. 00:01:33.510 --> 00:01:35.989 And when they sliced the data by the creator's 00:01:35.989 --> 00:01:38.909 origin, China officially surpassed the U .S. 00:01:38.930 --> 00:01:41.769 in the open AI ecosystem. Can you give us the 00:01:41.769 --> 00:01:43.709 exact numbers here? Sure. China now leads with 00:01:43.709 --> 00:01:46.730 17 .1 % of all those model downloads. The U .S. 00:01:46.730 --> 00:01:49.870 is just behind at 15 .8%. But the really deep 00:01:49.870 --> 00:01:51.930 insight is who makes up those percentages. Right. 00:01:52.010 --> 00:01:54.170 And what's absolutely staggering here is who 00:01:54.170 --> 00:02:04.500 is missing from the U .S. side. And here is the 00:02:04.500 --> 00:02:06.840 kicker, the thing that fundamentally changes 00:02:06.840 --> 00:02:10.680 how we view the global AI race. The U .S. leaders 00:02:10.680 --> 00:02:12.919 we talk about every single day, Google, Meta, 00:02:13.020 --> 00:02:18.439 OpenAI, have a 0 % presence in the 2024 to 2025 00:02:18.439 --> 00:02:21.300 top charts. They are literally invisible in the 00:02:21.300 --> 00:02:23.699 open marketplace. That feels so counterintuitive. 00:02:23.740 --> 00:02:25.919 We hear about them constantly. Only one notable 00:02:25.919 --> 00:02:28.460 U .S. contributor even makes the list, and that's 00:02:28.460 --> 00:02:31.659 Comfy, clocking in at 5 .4%. Right. And for listeners 00:02:31.659 --> 00:02:34.199 who might not know, Comfy isn't an LLM like DeepSeek 00:02:34.199 --> 00:02:37.719 or Quen. It's a popular U .S.-based workflow 00:02:37.719 --> 00:02:40.419 and tooling system. It's specific. for AI image 00:02:40.419 --> 00:02:42.120 generation. It just highlights that the American 00:02:42.120 --> 00:02:44.439 presence is in the utility layer, not the foundation 00:02:44.439 --> 00:02:46.800 model layer. So if you look at the strategic 00:02:46.800 --> 00:02:49.080 divergence, it becomes crystal clear what's happening. 00:02:49.319 --> 00:02:51.580 U .S. companies are focused almost entirely on 00:02:51.580 --> 00:02:54.900 closed APIs and high margin product monetization. 00:02:54.960 --> 00:02:56.560 They want you using their subscription services. 00:02:57.159 --> 00:02:59.800 Exactly. We see Meta kind of sitting on Lama 00:02:59.800 --> 00:03:02.860 3, carefully controlling its distribution. And 00:03:02.860 --> 00:03:05.539 OpenAI, which, you know, pioneered open source, 00:03:05.780 --> 00:03:08.280 it shifted away from open releases years ago. 00:03:08.780 --> 00:03:11.120 Google's best models are often locked behind 00:03:11.120 --> 00:03:13.379 cloud subscriptions. It's all about immediate 00:03:13.379 --> 00:03:16.080 financial returns and proprietary control. Meanwhile, 00:03:16.199 --> 00:03:18.500 the Chinese counter strategy is aggressively 00:03:18.500 --> 00:03:22.599 open. DeepSeek is shipping incredibly fast, highly 00:03:22.599 --> 00:03:25.460 multilingual models and releasing usable checkpoints 00:03:25.460 --> 00:03:28.280 almost weekly. Quinn is doing the same, improving 00:03:28.280 --> 00:03:31.360 its whole multilingual suite almost in real time. 00:03:31.740 --> 00:03:33.539 They're playing the long game. They're trying 00:03:33.539 --> 00:03:35.979 to establish their models as the default tool 00:03:35.979 --> 00:03:38.439 for developers all over the world. What used 00:03:38.439 --> 00:03:41.180 to be a U .S.-led open source movement, think 00:03:41.180 --> 00:03:43.780 of like the early days of Linux or TensorFlow, 00:03:44.180 --> 00:03:46.599 that's now becoming a China -led hardware plus 00:03:46.599 --> 00:03:48.939 weights game. Meaning the foundation models themselves, 00:03:49.199 --> 00:03:51.379 the weights, the multilingual support are now 00:03:51.379 --> 00:03:53.840 being set globally by companies optimized for 00:03:53.840 --> 00:03:56.180 fast open distribution. Which brings up a really 00:03:56.180 --> 00:03:58.620 important question. What is the long term, maybe 00:03:58.620 --> 00:04:02.280 even geopolitical impact of U .S. companies prioritizing 00:04:02.280 --> 00:04:05.219 closed monetization over leading the global open 00:04:05.219 --> 00:04:07.599 source standards. The global standard for open 00:04:07.599 --> 00:04:10.840 AI tools is rapidly becoming Chinese led technology. 00:04:11.479 --> 00:04:13.639 That brings us to the practical side of all this 00:04:13.639 --> 00:04:17.100 visibility. For decades, ranking on Google was 00:04:17.100 --> 00:04:20.319 kind of a blunt tool keyword stuffing generating 00:04:20.319 --> 00:04:23.660 backlinks. That playbook is completely dead now. 00:04:23.779 --> 00:04:26.660 That old SEO search engine optimization is just 00:04:26.660 --> 00:04:29.000 irrelevant in a generative future. We are now 00:04:29.000 --> 00:04:31.620 firmly in the era of generative engine optimization. 00:04:32.970 --> 00:04:35.170 Let's define GEO simply for everyone listening. 00:04:35.329 --> 00:04:38.029 Sure. GEO is just making your content easy for 00:04:38.029 --> 00:04:40.610 AI models to read, synthesize, and summarize 00:04:40.610 --> 00:04:42.709 accurately. You have to think of it as writing 00:04:42.709 --> 00:04:44.649 not just for human scanning, but for machine 00:04:44.649 --> 00:04:46.709 digestion. And here's where it gets really interesting 00:04:46.709 --> 00:04:49.149 because the big players like Google, Microsoft, 00:04:49.470 --> 00:04:51.189 Perplexity, they've all shared their core advice. 00:04:51.649 --> 00:04:56.279 AI search heavily favors content that is... fluent 00:04:56.279 --> 00:04:58.860 in machine -readable language. Which means structure 00:04:58.860 --> 00:05:01.019 is everything. You need strong brand clarity, 00:05:01.300 --> 00:05:04.000 sure, but more importantly, a predictable structure. 00:05:04.319 --> 00:05:06.500 Content has to be designed to be synthesized. 00:05:06.600 --> 00:05:08.899 If a model can't clearly parse your argument, 00:05:09.040 --> 00:05:11.439 your facts, your structure -like stacking Lego 00:05:11.439 --> 00:05:14.300 blocks of data, you just won't show up. And we're 00:05:14.300 --> 00:05:15.899 talking about more than just, you know, putting 00:05:15.899 --> 00:05:18.980 in some H1 tags, right? It's about clear semantic 00:05:18.980 --> 00:05:22.019 relationships, using structured data, getting 00:05:22.019 --> 00:05:25.399 rid of ambiguity. It's tough. Because the models 00:05:25.399 --> 00:05:27.600 are also trained on human language, which is 00:05:27.600 --> 00:05:30.339 messy by nature. I'll be honest, I still wrestle 00:05:30.339 --> 00:05:32.879 with prompt drift myself, you know, trying to 00:05:32.879 --> 00:05:34.680 keep my inputs perfectly clean and my outputs 00:05:34.680 --> 00:05:37.279 perfectly structured for these new rules. It's 00:05:37.279 --> 00:05:39.600 harder than it looks to do that consistently 00:05:39.600 --> 00:05:42.100 in your daily work. That vulnerability is real 00:05:42.100 --> 00:05:44.920 for everyone. The human brain is built for nuance, 00:05:45.000 --> 00:05:48.240 but machines are built for clarity. So given 00:05:48.240 --> 00:05:51.060 these new machine readability rules, how should 00:05:51.060 --> 00:05:54.040 we really rethink content creation from the ground 00:05:54.040 --> 00:05:56.339 up? Content needs to be structured and clear 00:05:56.339 --> 00:05:58.420 for machine synthesis, not just human scanning. 00:05:58.579 --> 00:06:00.220 All right, let's shift gears and look at some 00:06:00.220 --> 00:06:02.120 of the major news highlights hitting the ecosystem 00:06:02.120 --> 00:06:04.300 this week, starting with model spread and investment. 00:06:04.720 --> 00:06:07.079 What's fascinating here is just the sheer reach. 00:06:07.300 --> 00:06:10.819 Google's Gemini 3 Pro and the advanced Nano Banana 00:06:10.819 --> 00:06:14.060 Pro are now live in about 120 countries. They 00:06:14.060 --> 00:06:16.639 are expanding globally at light speed. OK, you 00:06:16.639 --> 00:06:18.959 have to tell us what Nano Banana Pro is, because 00:06:18.959 --> 00:06:21.639 that sounds like some internal jargon, not a 00:06:21.639 --> 00:06:24.120 public model name. It totally is. Nanobanana 00:06:24.120 --> 00:06:26.860 Pro is a highly advanced Google model. It's often 00:06:26.860 --> 00:06:28.819 used for these underlying visual and spatial 00:06:28.819 --> 00:06:31.579 tasks. So its practical function is unlocking 00:06:31.579 --> 00:06:34.439 character consistency in image generation, 3D 00:06:34.439 --> 00:06:37.980 translation. It handles the really finicky, detailed 00:06:37.980 --> 00:06:40.540 work behind the scenes. And for users who are 00:06:40.540 --> 00:06:42.779 interacting with the Pro tier, this ability to 00:06:42.779 --> 00:06:45.220 tap thinking with 3Pro right in the search bar 00:06:45.220 --> 00:06:48.399 is huge. It turns these complex queries into 00:06:48.399 --> 00:06:51.500 interactive, almost masterful visual summaries. 00:06:51.920 --> 00:06:54.519 That's real power for the user. And beyond the 00:06:54.519 --> 00:06:56.500 models, we saw a lot of investment signaling 00:06:56.500 --> 00:06:58.959 market confidence. Black Forest Labs just raised 00:06:58.959 --> 00:07:01.980 $300 million. That funding is going to support 00:07:01.980 --> 00:07:04.079 major tools and infrastructure that people use 00:07:04.079 --> 00:07:06.819 every day, including models like Grok, Adobe's 00:07:06.819 --> 00:07:09.959 platform integration, and Flux2's new 4K image 00:07:09.959 --> 00:07:11.980 model. The money's flowing to infrastructure. 00:07:12.399 --> 00:07:15.019 And if you connect that investment to the bigger 00:07:15.019 --> 00:07:17.120 picture, you start to see these foundational 00:07:17.120 --> 00:07:20.600 shifts in education. AI is quickly becoming the 00:07:20.600 --> 00:07:23.699 new it major. Enrollment at places like MIT is 00:07:23.699 --> 00:07:27.220 exploding for AI -specific tracks, and it's pushing 00:07:27.220 --> 00:07:28.779 out some of the traditional computer science 00:07:28.779 --> 00:07:31.399 programs. Yeah, and that educational shift is 00:07:31.399 --> 00:07:33.339 already playing out in the workforce, and it's 00:07:33.339 --> 00:07:36.060 happening fast. Accenture, the global consulting 00:07:36.060 --> 00:07:39.800 giant, just hard -rebranded 800 ,000 employee 00:07:39.800 --> 00:07:43.180 roles. They are now called re -inventors. That 00:07:43.180 --> 00:07:45.399 sends a pretty clear message, doesn't it? It's 00:07:45.399 --> 00:07:48.819 basically, learn J &AI or you're out. mandate 00:07:48.819 --> 00:07:50.959 for the whole consulting world. They are not 00:07:50.959 --> 00:07:53.040 waiting around. They're forcing massive media 00:07:53.040 --> 00:07:56.360 change. And on a necessary societal note, OpenAI 00:07:56.360 --> 00:07:59.819 dropped a $2 million grant program focused specifically 00:07:59.819 --> 00:08:02.579 on AI mental health research. It's a critical 00:08:02.579 --> 00:08:04.879 area recognizing we need to study the cognitive 00:08:04.879 --> 00:08:07.060 and emotional impact of these tools as quickly 00:08:07.060 --> 00:08:09.779 as we build them. So do these rapid workforce 00:08:09.779 --> 00:08:13.000 and education changes signal a foundational sort 00:08:13.000 --> 00:08:15.220 of rapid obsolescence of traditional white collar 00:08:15.220 --> 00:08:18.060 roles that you have to upskill instantly? Companies 00:08:18.060 --> 00:08:20.860 are redefining jobs instantly, forcing massive, 00:08:20.959 --> 00:08:23.800 immediate upskilling globally. Okay, let's turn 00:08:23.800 --> 00:08:26.019 our attention to one of the most exciting breakthroughs 00:08:26.019 --> 00:08:28.779 of the year. It's this blend of intuition and 00:08:28.779 --> 00:08:31.980 hard logic in mathematics. Harmonix Aristotle 00:08:31.980 --> 00:08:34.700 AI system cracked a problem that has been open 00:08:34.700 --> 00:08:37.779 for 30 years. This is truly foundational stuff. 00:08:38.159 --> 00:08:40.860 Aristotle solved the famous Erdos problem, hashtag 00:08:40.860 --> 00:08:44.059 124, which has been open since the 1990s. And 00:08:44.059 --> 00:08:46.500 the speed of the whole process is, well, it's 00:08:46.500 --> 00:08:48.759 what's so remarkable here. How fast are we talking? 00:08:49.019 --> 00:08:51.919 So Aristotle solved the entire problem, something 00:08:51.919 --> 00:08:54.220 that decades of human mathematicians couldn't 00:08:54.220 --> 00:08:57.120 crack in just six hours. But here's the critical 00:08:57.120 --> 00:08:59.580 detail about rigor that changes everything. It 00:08:59.580 --> 00:09:02.740 then verified the full proof using Lean, a formal 00:09:02.740 --> 00:09:05.620 proof checker, in only 60 seconds. Let's pause 00:09:05.620 --> 00:09:07.759 on Lean for a second. What is that exactly? Lean 00:09:07.759 --> 00:09:10.700 is basically a rigorous logical accountant. It's 00:09:10.700 --> 00:09:12.759 software designed to check every single step 00:09:12.759 --> 00:09:15.779 of a mathematical proof to ensure absolute formalized 00:09:15.779 --> 00:09:18.179 accuracy. There's no room for human error or 00:09:18.179 --> 00:09:22.440 any fuzzy steps. Whoa. Imagine scaling this intuitive 00:09:22.440 --> 00:09:26.259 and rigorously verifiable approach across all 00:09:26.259 --> 00:09:29.240 the unsolved problems in science and math. That 00:09:29.240 --> 00:09:32.039 capability just fundamentally changes the research 00:09:32.039 --> 00:09:35.370 timeline. That is a moment of wonder. And that 00:09:35.370 --> 00:09:37.669 speed and rigor is the essence of what Harmonic 00:09:37.669 --> 00:09:41.210 is calling vibe proving. It's a fascinating concept 00:09:41.210 --> 00:09:44.049 that bridges two worlds that are usually separate. 00:09:44.169 --> 00:09:46.309 Explain vibe proving for us. What does that name 00:09:46.309 --> 00:09:49.250 even mean? Vibe proving means the AI intuitively 00:09:49.250 --> 00:09:53.090 explores the huge solution space. It takes these 00:09:53.090 --> 00:09:55.950 high level creative leaps a human mathematician 00:09:55.950 --> 00:09:58.090 might take late at night when they finally see 00:09:58.090 --> 00:10:00.570 the answer. It gets the vibe of the solution. 00:10:00.889 --> 00:10:03.360 And then it checks its own work. Precisely. It 00:10:03.360 --> 00:10:05.840 immediately formalizes that intuitive idea with 00:10:05.840 --> 00:10:08.480 symbolic rigor to make sure every single step 00:10:08.480 --> 00:10:10.759 is checkable and provable by something like lean. 00:10:11.000 --> 00:10:13.960 So it blends this fluid discovery with hard proof. 00:10:14.179 --> 00:10:16.960 It compresses years of expert work into just 00:10:16.960 --> 00:10:20.100 hours with zero loss of accuracy. And this isn't 00:10:20.100 --> 00:10:22.820 just a theory. Aristotle already earned an International 00:10:22.820 --> 00:10:25.740 Math Olympiad gold medal. It's already proven 00:10:25.740 --> 00:10:28.480 itself against the best human minds. Yeah, it 00:10:28.480 --> 00:10:31.080 now instantly ranks alongside the math reasoning 00:10:31.080 --> 00:10:33.659 capabilities of the big players like OpenAI and 00:10:33.659 --> 00:10:36.679 Google DeepMind. The difference is the speed 00:10:36.679 --> 00:10:40.379 of verifiable novel discovery. What does vibe 00:10:40.379 --> 00:10:42.820 proving fundamentally change about the process 00:10:42.820 --> 00:10:45.139 of foundational mathematical research? It creates 00:10:45.139 --> 00:10:48.039 a blend of human -like creative intuition and 00:10:48.039 --> 00:10:51.000 machine -level absolute accuracy. That was a 00:10:51.000 --> 00:10:52.340 tremendous amount of ground we just covered. 00:10:52.480 --> 00:10:55.299 So let's synthesize the three core takeaways 00:10:55.299 --> 00:10:57.620 that you, the listener, really need to carry 00:10:57.620 --> 00:11:00.279 forward. First, remember this. The center of 00:11:00.279 --> 00:11:02.620 gravity for open AI leadership has decisively 00:11:02.620 --> 00:11:05.299 shifted to China. It's driven by these aggressive 00:11:05.299 --> 00:11:08.799 open strategies of DeepSeek and Quinn. The US 00:11:08.799 --> 00:11:11.120 giants are just absent from the global open source 00:11:11.120 --> 00:11:14.320 table. Second, content survival demands immediate 00:11:14.320 --> 00:11:17.019 adaptation to machine synthesis through generative 00:11:17.019 --> 00:11:19.759 engine optimization. If the machines can't read 00:11:19.759 --> 00:11:21.960 your structured content perfectly, you won't 00:11:21.960 --> 00:11:24.389 exist in the generative search future. And finally, 00:11:24.570 --> 00:11:27.149 AI systems are now capable of generating foundational, 00:11:27.490 --> 00:11:30.129 verifiable new mathematical knowledge in hours. 00:11:30.389 --> 00:11:32.830 This is going to accelerate scientific discovery 00:11:32.830 --> 00:11:35.330 at a rate we've never seen before. Given that 00:11:35.330 --> 00:11:37.250 Aristotle solved a 30 -year -old mathematical 00:11:37.250 --> 00:11:40.169 problem and then verified the entire solution 00:11:40.169 --> 00:11:43.450 in just 60 seconds, how quickly will AI systems 00:11:43.450 --> 00:11:45.909 move through the remaining known open problems 00:11:45.909 --> 00:11:49.190 in mathematics and science? That is the question 00:11:49.190 --> 00:11:51.269 we'll leave you with today. Keep exploring those 00:11:51.269 --> 00:11:53.070 concepts and we'll catch you on the next Deep 00:11:53.070 --> 00:11:53.429 Dive.