WEBVTT 00:00:00.000 --> 00:00:02.240 So we got this huge stack of source material 00:00:02.240 --> 00:00:05.839 today, but there was one metric that just immediately 00:00:05.839 --> 00:00:09.000 jumped out at me. Let me guess. The math competition 00:00:09.000 --> 00:00:12.099 one. Yeah. An open source model achieved a, well... 00:00:12.669 --> 00:00:16.370 Pretty stunning 96 % gold medal performance on 00:00:16.370 --> 00:00:20.070 the AME 2025 math competition. Right. And what's 00:00:20.070 --> 00:00:22.230 so fascinating there isn't just the raw performance. 00:00:22.370 --> 00:00:25.929 It's the strategy. The lab, DeepSeek, they didn't 00:00:25.929 --> 00:00:28.230 just win. No. They immediately turned around 00:00:28.230 --> 00:00:30.910 and made the blueprints available to everyone. 00:00:31.109 --> 00:00:32.750 Yeah. I mean, that just fundamentally shifts 00:00:32.750 --> 00:00:35.710 the playing field instantly. Welcome to the Deep 00:00:35.710 --> 00:00:37.869 Dive. Our whole mission here is to take this 00:00:37.869 --> 00:00:40.829 firehose of AI news and research and, you know, 00:00:40.829 --> 00:00:43.109 really distill it down to what actually matters 00:00:43.109 --> 00:00:45.969 for you. And today's stack shows some major, 00:00:46.049 --> 00:00:48.729 major acceleration, not just small steps. Exactly. 00:00:48.789 --> 00:00:50.450 We're looking for the signal and all that noise. 00:00:50.509 --> 00:00:52.390 And today we're going to focus on three areas 00:00:52.390 --> 00:00:54.890 that really feel like a seismic shift. First, 00:00:55.049 --> 00:00:57.380 we have to talk about that new... Benchmark from 00:00:57.380 --> 00:01:01.359 DeepSeek v3 .2, especially its Olympia tier capabilities. 00:01:01.920 --> 00:01:04.920 Second, we'll hit the AI flashpoints, all the 00:01:04.920 --> 00:01:07.159 rapid fire tools and infrastructure changes that 00:01:07.159 --> 00:01:09.939 are happening like right now. And finally, we're 00:01:09.939 --> 00:01:12.000 going to dive into an internal study from Anthropic 00:01:12.000 --> 00:01:14.739 about how the AI co -worker is already changing 00:01:14.739 --> 00:01:17.459 how expert engineers get their work done. OK, 00:01:17.540 --> 00:01:20.200 let's start with that new benchmark, DeepSeek 00:01:20.200 --> 00:01:23.439 v3 .2 and this high performance model they're 00:01:23.439 --> 00:01:26.219 calling Specialer. The sources had a great analogy 00:01:26.219 --> 00:01:28.780 for it. Yeah, I like that one. They said if GPT 00:01:28.780 --> 00:01:31.159 -5 and Gemini 3 Pro are these custom -built, 00:01:31.200 --> 00:01:34.819 super -exclusive Tesla roadsters, DeepSeek just 00:01:34.819 --> 00:01:36.900 rolled out a high -speed electric bullet train 00:01:36.900 --> 00:01:39.120 for everyone. And they made the tracks free. 00:01:39.359 --> 00:01:41.319 Yeah. I think we need to define what Olympiad 00:01:41.319 --> 00:01:43.480 tier actually means because it's not just, you 00:01:43.480 --> 00:01:46.019 know, memorizing facts. Right. These tests, like 00:01:46.019 --> 00:01:48.739 the IMO or the AIM, they require really novel 00:01:48.739 --> 00:01:51.219 problem -solving. The model has to synthesize 00:01:51.219 --> 00:01:54.150 concepts. reason deeply, and apply strategies 00:01:54.150 --> 00:01:56.230 it has never seen before. These are not your 00:01:56.230 --> 00:01:58.829 average college entry exams. These are competitive, 00:01:58.989 --> 00:02:02.069 almost research -level tests. And the results 00:02:02.069 --> 00:02:05.150 from special are just staggering. It hit 96 % 00:02:05.150 --> 00:02:08.610 on AUN 2025. And even better, it got 99 .2 % 00:02:08.610 --> 00:02:11.409 on the Harvard -MIT Mathematics Tournament, HMMT 00:02:11.409 --> 00:02:15.430 2025. 99 .2. Yeah, and that's stated as the best 00:02:15.430 --> 00:02:18.400 of any reasoning model currently out there. It 00:02:18.400 --> 00:02:21.240 also grabbed gold status across all the big competitive 00:02:21.240 --> 00:02:27.599 programming events. IMO, CMO, IOI. the ICPC finals. 00:02:27.860 --> 00:02:29.979 So when we talk about AI achieving human level 00:02:29.979 --> 00:02:32.180 competence, this is a whole different level. 00:02:32.360 --> 00:02:34.819 Exactly. This is a basic task. This is proving 00:02:34.819 --> 00:02:37.840 world class deep expertise in abstract reasoning. 00:02:38.020 --> 00:02:39.620 And this is where the story gets really, really 00:02:39.620 --> 00:02:42.139 interesting. The economics. Yeah. So even if 00:02:42.139 --> 00:02:44.139 special needs answers are longer, maybe two or 00:02:44.139 --> 00:02:46.379 three times longer than the competition, it's 00:02:46.379 --> 00:02:48.479 still running five to 10 times cheaper overall. 00:02:48.680 --> 00:02:50.479 And that is the ultimate democratizing force, 00:02:50.580 --> 00:02:52.560 right? Yeah. A five to 10 times cost reduction 00:02:52.560 --> 00:02:55.379 means something that used to need a huge R &D 00:02:55.379 --> 00:02:58.460 budget. can now be accessed by a startup, a university. 00:02:58.719 --> 00:03:01.900 Yeah. Even just one person. Whoa. Imagine scaling 00:03:01.900 --> 00:03:05.039 that 5 to 10x cost reduction to a billion queries. 00:03:05.199 --> 00:03:08.520 Like across an entire global education system. 00:03:08.599 --> 00:03:11.240 That changes absolutely everything. It does. 00:03:11.479 --> 00:03:13.259 And the open source strategy is what seals the 00:03:13.259 --> 00:03:16.159 deal. They published everything, which is so 00:03:16.159 --> 00:03:18.639 rare when you have performance this high. Everything. 00:03:18.680 --> 00:03:20.879 Yeah, the training data methodology, their fine 00:03:20.879 --> 00:03:22.919 -tuning techniques, even reports on where the 00:03:22.919 --> 00:03:28.259 model still fails. V3 .2 is live on API, and 00:03:28.259 --> 00:03:31.099 they're offering SpecialEye as a temporary endpoint 00:03:31.099 --> 00:03:34.460 until December 15th, just for testing. They're 00:03:34.460 --> 00:03:36.520 basically forcing the whole world to speed up. 00:03:36.599 --> 00:03:38.879 So what are the immediate implications of having 00:03:38.879 --> 00:03:41.219 this high performance open source AI blueprint 00:03:41.219 --> 00:03:43.900 just out there for everyone? I mean, the speed 00:03:43.900 --> 00:03:46.280 of global adoption is just going to skyrocket. 00:03:46.280 --> 00:03:48.599 The cost barrier has been shattered. Lower cost 00:03:48.599 --> 00:03:51.580 and open blueprints will rapidly accelerate global 00:03:51.580 --> 00:03:54.569 AI adoption. Yep. Okay, let's shift gears. Let's 00:03:54.569 --> 00:03:56.770 look at the rapid -fire highlights, the flashpoints 00:03:56.770 --> 00:03:58.530 that are defining the landscape right now. We 00:03:58.530 --> 00:04:00.430 can probably group these, maybe starting with 00:04:00.430 --> 00:04:02.969 creativity and content. Sounds good. Yeah. And 00:04:02.969 --> 00:04:05.050 we've definitely hit peak competition in video 00:04:05.050 --> 00:04:07.810 generation. The tools are moving past simple 00:04:07.810 --> 00:04:10.729 prompts into real cinematic control. Runway's 00:04:10.729 --> 00:04:14.030 Gen 4 .5 just dropped, and its feature list is 00:04:14.030 --> 00:04:16.870 pretty critical. We're talking full camera control. 00:04:17.110 --> 00:04:19.629 Which means you can dictate. The perspective, 00:04:19.850 --> 00:04:22.410 the angle, the movement without needing to be 00:04:22.410 --> 00:04:24.620 a pro editor. Right. It also has near -perfect 00:04:24.620 --> 00:04:27.180 physics and can orchestrate multiple separate 00:04:27.180 --> 00:04:30.720 elements in one scene. That is huge. Full camera 00:04:30.720 --> 00:04:32.879 control means a creator can get cinematic shots 00:04:32.879 --> 00:04:36.819 right out of the box. And critically, the sources 00:04:36.819 --> 00:04:41.079 say 4 .5 officially beat rivals like Sora 2 Pro 00:04:41.079 --> 00:04:44.759 and VO3 in quality. We're also seeing this massive 00:04:44.759 --> 00:04:47.639 convergence happening. Klingon 1 is on the market 00:04:47.639 --> 00:04:50.399 now, and it handles both video creation and editing 00:04:50.399 --> 00:04:53.339 in one interface. No more handling. between different 00:04:53.339 --> 00:04:55.519 tools and other platforms are consolidating power 00:04:55.519 --> 00:04:58.019 too. OpenAI, for instance, is combining over 00:04:58.019 --> 00:05:00.439 50 image models into one place. So you don't 00:05:00.439 --> 00:05:02.199 even have to choose the best model for a certain 00:05:02.199 --> 00:05:04.360 style anymore. Exactly. The platform just handles 00:05:04.360 --> 00:05:06.540 it. Now, moving over to the enterprise and infrastructure 00:05:06.540 --> 00:05:09.420 side, there's a really interesting signal. OpenAI 00:05:09.420 --> 00:05:12.319 reportedly declared an internal code red for 00:05:12.319 --> 00:05:15.800 chat GPT. Yeah, they're pausing new feature rollouts. 00:05:15.860 --> 00:05:18.639 The internal memo explicitly said they're going 00:05:18.639 --> 00:05:21.899 back to basics. A code red? Usually means they're 00:05:21.899 --> 00:05:24.839 hitting scaling issues, right? Foundational reliability 00:05:24.839 --> 00:05:27.139 problems. Yeah, that's the sense I get. And you 00:05:27.139 --> 00:05:29.560 see this focus on stability across the industry. 00:05:29.720 --> 00:05:32.600 Look at NVIDIA. They just invested $2 billion 00:05:32.600 --> 00:05:36.220 in synopses specifically to use cloud computing 00:05:36.220 --> 00:05:39.180 to speed up product engineering. Right. The focus 00:05:39.180 --> 00:05:42.680 isn't on the next flashy consumer feature. No, 00:05:42.699 --> 00:05:44.860 it's on making the underlying engine reliable 00:05:44.860 --> 00:05:48.639 under massive global pressure. Okay. Okay, switching 00:05:48.639 --> 00:05:50.920 to practical applications. Two things really 00:05:50.920 --> 00:05:53.899 stood out. One was this specialized deep research 00:05:53.899 --> 00:05:56.879 prompt that can instantly spit out a 10 -page 00:05:56.879 --> 00:05:59.420 deep dive on pretty much any company. Super useful. 00:05:59.600 --> 00:06:01.600 For sure. And then we're seeing AI move into 00:06:01.600 --> 00:06:05.399 governance, but subtly. Robot traffic cops are 00:06:05.399 --> 00:06:07.819 now directing traffic in Hangzhou. It's a low 00:06:07.819 --> 00:06:09.439 -stakes way to get people used to machine governance, 00:06:09.620 --> 00:06:12.040 you know. And for knowledge workers, Kimi dropped 00:06:12.040 --> 00:06:13.879 something they're calling free agentic slides. 00:06:14.199 --> 00:06:17.060 Okay, define agentic slides for us. So agentic 00:06:17.060 --> 00:06:19.079 just means the model can perform multiple steps 00:06:19.079 --> 00:06:22.360 on its own. In this case, it's an AI that creates 00:06:22.360 --> 00:06:26.000 complete, editable, exportable PowerPoint presentations 00:06:26.000 --> 00:06:29.180 with unlimited images start to finish. So it 00:06:29.180 --> 00:06:30.819 takes away the most annoying part of a lot of 00:06:30.819 --> 00:06:33.639 corporate jobs. Pretty much. So considering this 00:06:33.639 --> 00:06:35.779 code red pivot and all the intense competition, 00:06:36.160 --> 00:06:38.699 where do you think the primary focus shifts for 00:06:38.699 --> 00:06:41.560 the major AI labs now? I think the focus is definitely 00:06:41.560 --> 00:06:44.819 shifting from adding new features to just solidifying 00:06:44.819 --> 00:06:47.339 the core model's reliability and safety. So less 00:06:47.339 --> 00:06:49.899 about new features, more about stable, reliable 00:06:49.899 --> 00:06:52.699 model performance. Exactly. That's the bottleneck 00:06:52.699 --> 00:06:55.319 now. And that focus on reliability actually leads 00:06:55.319 --> 00:06:58.040 us perfectly into our last segment. the anthropic 00:06:58.040 --> 00:07:00.439 internal report right this study gives us this 00:07:00.439 --> 00:07:03.199 rare really high resolution look into the daily 00:07:03.199 --> 00:07:05.759 lives of their own engineers using advanced ai 00:07:05.759 --> 00:07:10.060 in this case claude it's based on 132 employee 00:07:10.060 --> 00:07:12.819 surveys and 53 interviews and the productivity 00:07:12.819 --> 00:07:14.819 gains are just i mean they're astonishing they 00:07:14.819 --> 00:07:16.980 found that their engineers are now using claude 00:07:16.980 --> 00:07:21.970 in 60 of their daily work 60 yeah And this translates 00:07:21.970 --> 00:07:25.769 directly to a 50 % productivity boost. That's 00:07:25.769 --> 00:07:27.889 two to three times the jump they saw just a year 00:07:27.889 --> 00:07:31.050 ago. A 50 % boost is massive. It really shows 00:07:31.050 --> 00:07:33.209 how quickly this technology is moving from just 00:07:33.209 --> 00:07:36.009 being a tool to being more of a co -worker. Six 00:07:36.009 --> 00:07:38.170 months ago, Claude needed a nudge after maybe 00:07:38.170 --> 00:07:41.490 10 independent actions. Now it's handling 20 00:07:41.490 --> 00:07:43.930 or more sequential actions before a human needs 00:07:43.930 --> 00:07:46.550 to step in. And that capability, of course, presents 00:07:46.550 --> 00:07:48.689 a double -edged sword, which the report points 00:07:48.689 --> 00:07:51.410 out. Yeah. On the plus side, Claude makes these 00:07:51.410 --> 00:07:54.569 specialist engineers more full stack. They can 00:07:54.569 --> 00:07:56.569 suddenly work on database management or front 00:07:56.569 --> 00:07:58.990 -end design, things way outside their core skills. 00:07:59.290 --> 00:08:01.569 That sounds empowering, but what's the negative 00:08:01.569 --> 00:08:04.560 side? The concern raised in the interviews is 00:08:04.560 --> 00:08:06.779 that relying on the tool might actually make 00:08:06.779 --> 00:08:09.800 them worse at their core craft over time. If 00:08:09.800 --> 00:08:12.379 you outsource that mental muscle, it starts to 00:08:12.379 --> 00:08:14.779 atrophy. I still wrestle with prompt drift myself, 00:08:15.079 --> 00:08:17.660 where you rely on the tool so much that your 00:08:17.660 --> 00:08:19.759 own ability to define the problem degrades a 00:08:19.759 --> 00:08:22.279 little. It's hard to justify taking the long 00:08:22.279 --> 00:08:24.660 way when the tool works so well. Yeah, that reliance 00:08:24.660 --> 00:08:27.240 challenge is universal. And the report notes 00:08:27.240 --> 00:08:30.139 that Claude is actively replacing colleague interactions. 00:08:30.860 --> 00:08:32.980 What do you mean? Like, the quick question you'd 00:08:32.980 --> 00:08:36.059 ask a coworker, the informal brainstorm. That's 00:08:36.059 --> 00:08:38.440 now happening with the AI. So we're seeing this 00:08:38.440 --> 00:08:42.860 slow shift from simple task delegation to full 00:08:42.860 --> 00:08:45.320 work stream collaboration with the AI. So the 00:08:45.320 --> 00:08:47.659 AI isn't just replacing humans, which is always 00:08:47.659 --> 00:08:49.460 the big fear. It's unlocking what the report 00:08:49.460 --> 00:08:52.860 calls latent work. Exactly. All the necessary 00:08:52.860 --> 00:08:56.639 but messy, annoying stuff. Documentation, finding 00:08:56.639 --> 00:08:59.600 obscure edge cases, the work that was too expensive 00:08:59.600 --> 00:09:02.379 or time consuming to do before. Now it just gets 00:09:02.379 --> 00:09:04.929 done. So if AI is replacing those quick colleague 00:09:04.929 --> 00:09:07.750 interactions, what happens to the natural, unscripted 00:09:07.750 --> 00:09:10.409 knowledge sharing that really defines a strong 00:09:10.409 --> 00:09:13.070 team culture? Well, that's the risk. That informal 00:09:13.070 --> 00:09:15.149 knowledge transfer, those lucky accidents that 00:09:15.149 --> 00:09:16.990 lead to innovation, they could get automated 00:09:16.990 --> 00:09:19.309 away or just lost if all communication becomes 00:09:19.309 --> 00:09:21.929 AI mediated. We risk losing informal knowledge 00:09:21.929 --> 00:09:24.429 transfer when collaboration becomes purely AI 00:09:24.429 --> 00:09:26.850 mediated. It's a key cultural risk we have to 00:09:26.850 --> 00:09:29.470 watch. OK, so let's summarize the big findings 00:09:29.470 --> 00:09:31.889 from this deep dive. We've seen this incredible 00:09:31.889 --> 00:09:34.169 external competition from labs like DeepSeek. 00:09:34.269 --> 00:09:36.529 They're offering world class performance at a 00:09:36.529 --> 00:09:39.950 super low cost, which is forcing this radical 00:09:39.950 --> 00:09:42.149 democratization. And at the same time, we've 00:09:42.149 --> 00:09:44.070 looked at the internal transformation through 00:09:44.070 --> 00:09:46.649 that unprofit report, which shows how high level 00:09:46.649 --> 00:09:48.830 experts are fundamentally changing how they work. 00:09:48.929 --> 00:09:52.169 That 50 percent productivity spike is real and 00:09:52.169 --> 00:09:54.210 it's happening right now. Right. And the tension 00:09:54.210 --> 00:09:57.600 is right there. Global AI power is getting cheaper 00:09:57.600 --> 00:10:01.730 and more open every single day. But human work, 00:10:01.789 --> 00:10:05.009 while it's more leveraged, is also becoming potentially 00:10:05.009 --> 00:10:07.149 more reliant on the machine for basic functions, 00:10:07.370 --> 00:10:09.649 replacing both manual work and talking to our 00:10:09.649 --> 00:10:12.529 colleagues. And before we wrap, we just wanted 00:10:12.529 --> 00:10:15.330 to briefly acknowledge you, our listener. Right. 00:10:15.389 --> 00:10:17.529 We've seen a lot of interest in moving toward 00:10:17.529 --> 00:10:19.649 a weekly recap. You know, the best tools, top 00:10:19.649 --> 00:10:22.549 news, smart takes. We hear you loud and clear 00:10:22.549 --> 00:10:24.690 that trying to process this volume of information 00:10:24.690 --> 00:10:27.850 daily is a real challenge. We are actively working 00:10:27.850 --> 00:10:30.009 on how to structure that for you. So here's the. 00:10:30.059 --> 00:10:32.120 final thought to leave you with based on everything 00:10:32.120 --> 00:10:34.440 we talked about today. The Anthropic Report shows 00:10:34.440 --> 00:10:37.419 AI unlocks latent work and raises productivity 00:10:37.419 --> 00:10:40.379 50%. But if you become worse at your core craft 00:10:40.379 --> 00:10:43.500 because AI is handling 60 % of the work, how 00:10:43.500 --> 00:10:45.659 do you define long -term career resilience? What 00:10:45.659 --> 00:10:47.879 specific human skill, one that can't be automated, 00:10:48.200 --> 00:10:50.500 will you intentionally sharpen this year to counter 00:10:50.500 --> 00:10:50.960 that trend?