So we got this huge stack of source material today, but there was one metric that just immediately jumped out at me. Let me guess. The math competition one. Yeah. An open source model achieved a, well... Pretty stunning 96 % gold medal performance on the AME 2025 math competition. Right. And what's so fascinating there isn't just the raw performance. It's the strategy. The lab, DeepSeek, they didn't just win. No. They immediately turned around and made the blueprints available to everyone.
Yeah. I mean, that just fundamentally shifts the playing field instantly. Welcome to the Deep Dive. Our whole mission here is to take this firehose of AI news and research and, you know, really distill it down to what actually matters for you. And today's stack shows some major, major acceleration, not just small steps. Exactly. We're looking for the signal and all that noise. And today we're going to focus on three areas that really feel like a seismic shift. First,
we have to talk about that new... Benchmark from DeepSeek v3 .2, especially its Olympia tier capabilities. Second, we'll hit the AI flashpoints, all the rapid fire tools and infrastructure changes that are happening like right now. And finally, we're going to dive into an internal study from Anthropic about how the AI co -worker is already changing how expert engineers get their work done. OK, let's start with that new benchmark, DeepSeek v3 .2 and this high performance model they're
calling Specialer. The sources had a great analogy for it. Yeah, I like that one. They said if GPT -5 and Gemini 3 Pro are these custom -built, super -exclusive Tesla roadsters, DeepSeek just rolled out a high -speed electric bullet train for everyone. And they made the tracks free. Yeah. I think we need to define what Olympiad tier actually means because it's not just, you know, memorizing facts. Right. These tests, like the IMO or the AIM, they require really novel
problem -solving. The model has to synthesize concepts. reason deeply, and apply strategies it has never seen before. These are not your average college entry exams. These are competitive, almost research -level tests. And the results from special are just staggering. It hit 96 % on AUN 2025. And even better, it got 99 .2 % on the Harvard -MIT Mathematics Tournament, HMMT 2025. 99 .2. Yeah, and that's stated as the best
of any reasoning model currently out there. It also grabbed gold status across all the big competitive programming events. IMO, CMO, IOI. the ICPC finals. So when we talk about AI achieving human level competence, this is a whole different level. Exactly. This is a basic task. This is proving world class deep expertise in abstract reasoning. And this is where the story gets really, really
interesting. The economics. Yeah. So even if special needs answers are longer, maybe two or three times longer than the competition, it's still running five to 10 times cheaper overall. And that is the ultimate democratizing force, right? Yeah. A five to 10 times cost reduction means something that used to need a huge R &D budget. can now be accessed by a startup, a university. Yeah. Even just one person. Whoa. Imagine scaling that 5 to 10x cost reduction to a billion queries.
Like across an entire global education system. That changes absolutely everything. It does. And the open source strategy is what seals the deal. They published everything, which is so rare when you have performance this high. Everything. Yeah, the training data methodology, their fine -tuning techniques, even reports on where the model still fails. V3 .2 is live on API, and they're offering SpecialEye as a temporary endpoint
until December 15th, just for testing. They're basically forcing the whole world to speed up. So what are the immediate implications of having this high performance open source AI blueprint just out there for everyone? I mean, the speed of global adoption is just going to skyrocket. The cost barrier has been shattered. Lower cost and open blueprints will rapidly accelerate global AI adoption. Yep. Okay, let's shift gears. Let's look at the rapid -fire highlights, the flashpoints
that are defining the landscape right now. We can probably group these, maybe starting with creativity and content. Sounds good. Yeah. And we've definitely hit peak competition in video generation. The tools are moving past simple prompts into real cinematic control. Runway's Gen 4 .5 just dropped, and its feature list is pretty critical. We're talking full camera control. Which means you can dictate. The perspective, the angle, the movement without needing to be
a pro editor. Right. It also has near -perfect physics and can orchestrate multiple separate elements in one scene. That is huge. Full camera control means a creator can get cinematic shots right out of the box. And critically, the sources say 4 .5 officially beat rivals like Sora 2 Pro and VO3 in quality. We're also seeing this massive convergence happening. Klingon 1 is on the market now, and it handles both video creation and editing
in one interface. No more handling. between different tools and other platforms are consolidating power too. OpenAI, for instance, is combining over 50 image models into one place. So you don't even have to choose the best model for a certain style anymore. Exactly. The platform just handles it. Now, moving over to the enterprise and infrastructure side, there's a really interesting signal. OpenAI reportedly declared an internal code red for chat GPT. Yeah, they're pausing new feature rollouts.
The internal memo explicitly said they're going back to basics. A code red? Usually means they're hitting scaling issues, right? Foundational reliability problems. Yeah, that's the sense I get. And you see this focus on stability across the industry. Look at NVIDIA. They just invested $2 billion in synopses specifically to use cloud computing to speed up product engineering. Right. The focus isn't on the next flashy consumer feature. No, it's on making the underlying engine reliable
under massive global pressure. Okay. Okay, switching to practical applications. Two things really stood out. One was this specialized deep research prompt that can instantly spit out a 10 -page deep dive on pretty much any company. Super useful. For sure. And then we're seeing AI move into governance, but subtly. Robot traffic cops are now directing traffic in Hangzhou. It's a low -stakes way to get people used to machine governance,
you know. And for knowledge workers, Kimi dropped something they're calling free agentic slides. Okay, define agentic slides for us. So agentic just means the model can perform multiple steps on its own. In this case, it's an AI that creates complete, editable, exportable PowerPoint presentations with unlimited images start to finish. So it takes away the most annoying part of a lot of
corporate jobs. Pretty much. So considering this code red pivot and all the intense competition, where do you think the primary focus shifts for the major AI labs now? I think the focus is definitely shifting from adding new features to just solidifying the core model's reliability and safety. So less about new features, more about stable, reliable model performance. Exactly. That's the bottleneck now. And that focus on reliability actually leads
us perfectly into our last segment. the anthropic internal report right this study gives us this rare really high resolution look into the daily lives of their own engineers using advanced ai in this case claude it's based on 132 employee surveys and 53 interviews and the productivity gains are just i mean they're astonishing they found that their engineers are now using claude in 60 of their daily work 60 yeah And this translates
directly to a 50 % productivity boost. That's two to three times the jump they saw just a year ago. A 50 % boost is massive. It really shows how quickly this technology is moving from just being a tool to being more of a co -worker. Six months ago, Claude needed a nudge after maybe 10 independent actions. Now it's handling 20 or more sequential actions before a human needs to step in. And that capability, of course, presents a double -edged sword, which the report points
out. Yeah. On the plus side, Claude makes these specialist engineers more full stack. They can suddenly work on database management or front -end design, things way outside their core skills. That sounds empowering, but what's the negative side? The concern raised in the interviews is that relying on the tool might actually make them worse at their core craft over time. If you outsource that mental muscle, it starts to
atrophy. I still wrestle with prompt drift myself, where you rely on the tool so much that your own ability to define the problem degrades a little. It's hard to justify taking the long way when the tool works so well. Yeah, that reliance challenge is universal. And the report notes that Claude is actively replacing colleague interactions. What do you mean? Like, the quick question you'd ask a coworker, the informal brainstorm. That's
now happening with the AI. So we're seeing this slow shift from simple task delegation to full work stream collaboration with the AI. So the AI isn't just replacing humans, which is always the big fear. It's unlocking what the report calls latent work. Exactly. All the necessary but messy, annoying stuff. Documentation, finding obscure edge cases, the work that was too expensive or time consuming to do before. Now it just gets
done. So if AI is replacing those quick colleague interactions, what happens to the natural, unscripted knowledge sharing that really defines a strong team culture? Well, that's the risk. That informal knowledge transfer, those lucky accidents that lead to innovation, they could get automated away or just lost if all communication becomes AI mediated. We risk losing informal knowledge transfer when collaboration becomes purely AI mediated. It's a key cultural risk we have to
watch. OK, so let's summarize the big findings from this deep dive. We've seen this incredible external competition from labs like DeepSeek. They're offering world class performance at a super low cost, which is forcing this radical democratization. And at the same time, we've looked at the internal transformation through that unprofit report, which shows how high level experts are fundamentally changing how they work. That 50 percent productivity spike is real and
it's happening right now. Right. And the tension is right there. Global AI power is getting cheaper and more open every single day. But human work, while it's more leveraged, is also becoming potentially more reliant on the machine for basic functions, replacing both manual work and talking to our colleagues. And before we wrap, we just wanted to briefly acknowledge you, our listener. Right. We've seen a lot of interest in moving toward a weekly recap. You know, the best tools, top
news, smart takes. We hear you loud and clear that trying to process this volume of information daily is a real challenge. We are actively working on how to structure that for you. So here's the. final thought to leave you with based on everything we talked about today. The Anthropic Report shows AI unlocks latent work and raises productivity 50%. But if you become worse at your core craft because AI is handling 60 % of the work, how
do you define long -term career resilience? What specific human skill, one that can't be automated, will you intentionally sharpen this year to counter that trend?