WEBVTT 00:00:00.000 --> 00:00:03.500 So here's the dilemma, right? Paying, say, $20 00:00:03.500 --> 00:00:05.860 a month for what's supposedly the absolute best 00:00:05.860 --> 00:00:09.539 AI or getting something that performs almost 00:00:09.539 --> 00:00:11.480 identically, maybe even better in some ways, 00:00:11.640 --> 00:00:15.580 but it's free and it's open source. Yeah, that's 00:00:15.580 --> 00:00:17.940 really the core tension in AI right now, isn't 00:00:17.940 --> 00:00:20.780 it? And this isn't just a small saving. It feels 00:00:20.780 --> 00:00:22.820 like a fundamental shift is happening. A new 00:00:22.820 --> 00:00:24.739 model has just shot up the rankings. Welcome 00:00:24.739 --> 00:00:26.600 to the Deep Dive. Today we're looking closely 00:00:26.600 --> 00:00:30.579 at Kimi K2 Thinking. This model has kind of quietly 00:00:30.579 --> 00:00:33.820 blown past almost every big name, closed source 00:00:33.820 --> 00:00:36.140 competitor in head -to -head tests. Absolutely. 00:00:36.280 --> 00:00:38.619 So our mission today is to really understand 00:00:38.619 --> 00:00:41.170 how. How did this open source model climb to 00:00:41.170 --> 00:00:43.530 number two globally? It's apparently just one 00:00:43.530 --> 00:00:46.270 point behind GPT -5, which is kind of wild. We're 00:00:46.270 --> 00:00:48.369 going to break down the tech, look at what it 00:00:48.369 --> 00:00:50.289 means for businesses, especially its reliability, 00:00:50.369 --> 00:00:52.170 because it's doing amazing things with financial 00:00:52.170 --> 00:00:55.210 analysis, expert coding, really complex stuff. 00:00:55.409 --> 00:00:57.710 Okay, let's unpack that shift, because it really 00:00:57.710 --> 00:00:59.990 does feel like another deep -seek moment, like 00:00:59.990 --> 00:01:02.270 the courses call it, where suddenly open source 00:01:02.270 --> 00:01:04.370 isn't just catching up, it's setting the standard. 00:01:04.629 --> 00:01:08.870 The ranking itself is genuinely staggering. Artificial 00:01:08.870 --> 00:01:12.030 analysis has Kimi K2 thinking ranked number two 00:01:12.030 --> 00:01:15.489 in the world. And we're not talking about it 00:01:15.489 --> 00:01:17.849 beating some niche models here. It's leaping 00:01:17.849 --> 00:01:20.370 over the giants. Like who? Let's name them. Okay, 00:01:20.430 --> 00:01:22.909 so it's outperforming Grok 4. That's XAI's big 00:01:22.909 --> 00:01:26.530 one. Claude 4 .5 Sonnet from Anthropic. And Gemini 00:01:26.530 --> 00:01:29.680 2 .5 Pro from Google. Just a few months back, 00:01:29.840 --> 00:01:32.739 the feeling was, you know, the absolute top tier 00:01:32.739 --> 00:01:35.319 that always be proprietary, always locked behind 00:01:35.319 --> 00:01:39.000 huge R &D budgets. Now you've got a zero cost 00:01:39.000 --> 00:01:41.140 option delivering intelligence that's knocking 00:01:41.140 --> 00:01:44.060 on GPT -5's door. That changes the whole equation 00:01:44.060 --> 00:01:45.640 if you're building things, right? It really does. 00:01:45.799 --> 00:01:48.280 And when you see that tiny gap, just one point 00:01:48.280 --> 00:01:51.659 behind GTT5, the open source part becomes the 00:01:51.659 --> 00:01:53.760 killer feature. You're getting, what, like 99 00:01:53.760 --> 00:01:56.420 % of the capability, but without the vendor risk. 00:01:56.599 --> 00:01:58.180 Exactly. And think about the practical side. 00:01:58.280 --> 00:02:00.120 If you're running a dev shop. or maybe handling 00:02:00.120 --> 00:02:02.239 really sensitive data. Now you can potentially 00:02:02.239 --> 00:02:04.180 run this model on your own servers, your private 00:02:04.180 --> 00:02:06.420 cloud. You keep complete control over your data. 00:02:06.480 --> 00:02:07.859 That's something you just don't get with most 00:02:07.859 --> 00:02:10.939 of the big cloud APIs. And the cost predictability 00:02:10.939 --> 00:02:13.759 must be huge. API fees can jump around, scale 00:02:13.759 --> 00:02:15.740 in ways you don't expect. Getting rid of that 00:02:15.740 --> 00:02:18.199 line item, that's got to feel good. Oh, totally. 00:02:18.280 --> 00:02:21.009 Imagine the budgeting relief. You know your server 00:02:21.009 --> 00:02:24.090 costs, roughly, but you ditch those massive, 00:02:24.189 --> 00:02:28.030 sometimes unpredictable API usage fees. It just 00:02:28.030 --> 00:02:30.729 shifts the power away from, you know, the handful 00:02:30.729 --> 00:02:32.750 of big tech companies controlling the best models. 00:02:32.870 --> 00:02:36.090 So zooming out, what would you say is the single 00:02:36.090 --> 00:02:39.969 biggest practical win for a business when a model 00:02:39.969 --> 00:02:42.349 this smart goes open source? I'd say it's freedom. 00:02:42.490 --> 00:02:44.449 Freedom from being locked into one vendor and 00:02:44.449 --> 00:02:47.169 those high kind of unpredictable API costs. Okay, 00:02:47.210 --> 00:02:50.310 let's pivot from the rankings to how it actually 00:02:50.310 --> 00:02:52.229 performs. Coding seems like a great place to 00:02:52.229 --> 00:02:53.810 start because that's often where these models 00:02:53.810 --> 00:02:55.930 show their limits. The source material kicks 00:02:55.930 --> 00:02:58.689 off with a wild challenge. Build a drag and drop 00:02:58.689 --> 00:03:01.129 website builder, kind of like Wix, but from a 00:03:01.129 --> 00:03:04.280 single prompt. That sounds ambitious. Yeah, it's 00:03:04.280 --> 00:03:06.819 a serious test of like structural reasoning. 00:03:07.020 --> 00:03:08.599 It's not just spitting out some static HTML. 00:03:09.000 --> 00:03:12.120 It needs to understand interaction dynamic elements. 00:03:12.340 --> 00:03:15.860 And KimiK2 delivered a fully functional editor, 00:03:16.060 --> 00:03:19.360 just one HTML file. It really seemed to grasp 00:03:19.360 --> 00:03:22.080 the underlying logic needed. So it had to plan 00:03:22.080 --> 00:03:23.939 out the JavaScript, right, the dragging, the 00:03:23.939 --> 00:03:26.620 dropping, handling the over events, plus the 00:03:26.620 --> 00:03:29.340 CSS for styling, and make it all work together 00:03:29.340 --> 00:03:32.439 smoothly. Yes. And the little details, too. It 00:03:32.439 --> 00:03:34.479 had working side panels, elements you could actually 00:03:34.479 --> 00:03:36.580 drag around, and importantly, a snap -to -grid 00:03:36.580 --> 00:03:38.840 system, you know, with the little red lines showing 00:03:38.840 --> 00:03:41.419 alignment, getting all that right in one shot 00:03:41.419 --> 00:03:44.960 from one prompt. That's really, really rare for 00:03:44.960 --> 00:03:46.919 this kind of complex application. Okay, the next 00:03:46.919 --> 00:03:49.560 test sounds even harder. The fluid dynamics simulation, 00:03:49.939 --> 00:03:51.659 I mean, that's straight up expert coding territory. 00:03:51.680 --> 00:03:55.219 You need physics, math, animation, all interactive. 00:03:55.560 --> 00:03:57.819 It's a real synthesis task. The model has to 00:03:57.819 --> 00:04:00.379 plan the physics simulation itself, managing 00:04:00.379 --> 00:04:02.939 particles. pressure, velocity, all that, and 00:04:02.939 --> 00:04:05.960 then translate that complex math into fast JavaScript 00:04:05.960 --> 00:04:08.960 that renders in real time on an HTML canvas. 00:04:09.180 --> 00:04:11.120 And the result was interactive. You could tweak 00:04:11.120 --> 00:04:15.080 sliders for viscosity, diffusion, and the fluid 00:04:15.080 --> 00:04:16.860 actually behaved like you'd expect. It looked 00:04:16.860 --> 00:04:19.220 realistic. Now, here's where it gets really telling, 00:04:19.279 --> 00:04:21.279 I think, about this open source parity idea. 00:04:21.519 --> 00:04:24.920 The sources point out that Grok 4, Cloud 4 .5, 00:04:25.240 --> 00:04:33.870 Gemini 2 .5 Pro... Run at all. Right. And Kini 00:04:33.870 --> 00:04:36.750 K2 and GPT -5 were the only two models tested 00:04:36.750 --> 00:04:38.730 that actually nailed it. That could build this 00:04:38.730 --> 00:04:41.649 complex physics -based interactive thing successfully. 00:04:41.889 --> 00:04:43.810 That's the takeaway you really need to absorb. 00:04:44.069 --> 00:04:46.069 Just imagine the complexity behind that. A system 00:04:46.069 --> 00:04:48.170 that gets the deep math of fluid physics and 00:04:48.170 --> 00:04:50.189 knows how to implement that efficiently in JavaScript, 00:04:50.410 --> 00:04:52.550 rendering it smoothly, all from a text prompt. 00:04:52.629 --> 00:04:54.410 That's something else. Yeah. It means we basically 00:04:54.410 --> 00:04:57.629 have two models now operating at that peak level 00:04:57.629 --> 00:05:00.170 for expert code generation. And one of them is 00:05:00.170 --> 00:05:03.060 free to use. So how does Kimi K2 passing that 00:05:03.060 --> 00:05:05.399 fluid dynamics test really change the game for 00:05:05.399 --> 00:05:08.319 evaluating model complexity? It proves Kimi K2 00:05:08.319 --> 00:05:11.100 isn't just good, it truly competes at the highest 00:05:11.100 --> 00:05:13.740 level of expert coding, even across different 00:05:13.740 --> 00:05:16.410 disciplines like physics and web tech. Before 00:05:16.410 --> 00:05:18.490 we get into maybe where it stumbles, there was 00:05:18.490 --> 00:05:21.230 another big win mentioned, right? The 3D geospatial 00:05:21.230 --> 00:05:24.230 visualization of Tokyo that also shows off its 00:05:24.230 --> 00:05:25.889 knowledge base. Yeah, and they did something 00:05:25.889 --> 00:05:27.490 important there. They turned off web search. 00:05:27.750 --> 00:05:30.629 So KimiK2 had to use early its internal baked 00:05:30.629 --> 00:05:33.430 -in knowledge. And it correctly placed neighborhoods 00:05:33.430 --> 00:05:37.129 like Shibuya and Asakusa on a 3D map. It added 00:05:37.129 --> 00:05:40.149 building extrusions, used Mapbox GLJS correctly, 00:05:40.389 --> 00:05:43.529 even added a day -night toggle. All from memory, 00:05:43.670 --> 00:05:45.449 essentially. That shows it's not just applying 00:05:45.449 --> 00:05:47.709 code patterns. It has actual world knowledge 00:05:47.709 --> 00:05:49.930 integrated pretty deeply. That's knowledge plus 00:05:49.930 --> 00:05:52.889 application skill. Absolutely. But, okay, to 00:05:52.889 --> 00:05:54.870 be balanced, we need to look at the edges. Where 00:05:54.870 --> 00:05:56.949 does GPT -5 still have that slight advantage? 00:05:57.329 --> 00:05:59.639 That brings us to the beehive simulation. Right, 00:05:59.699 --> 00:06:02.459 another super complex test. This needed specific 00:06:02.459 --> 00:06:05.439 biological knowledge about how bees build hives 00:06:05.439 --> 00:06:08.279 combined with tricky geometry, those hexagonal 00:06:08.279 --> 00:06:10.560 cells, forging patterns, interactive controls. 00:06:10.980 --> 00:06:14.139 And KimiK2 did build a simulation, which honestly 00:06:14.139 --> 00:06:16.220 is still impressive. You could see cells forming, 00:06:16.300 --> 00:06:19.790 bees moving around, but... The hexagonal alignment 00:06:19.790 --> 00:06:22.009 was off. Critically flawed, actually. The pattern 00:06:22.009 --> 00:06:24.689 wasn't regular like a real honeycomb. The hive 00:06:24.689 --> 00:06:27.350 grew kind of chaotically, not in that structured, 00:06:27.490 --> 00:06:30.290 layered way you see in nature. But GPT -5 got 00:06:30.290 --> 00:06:32.649 the geometry perfect. Stable, mathematically 00:06:32.649 --> 00:06:36.470 correct hexagons. Apparently, yes. GPT -5 nailed 00:06:36.470 --> 00:06:39.730 that part. Hmm. So if Chemiket -2 can handle 00:06:39.730 --> 00:06:42.810 the complex math of fluid dynamics, why would 00:06:42.810 --> 00:06:45.209 this specific geometric pattern trip it up? That 00:06:45.209 --> 00:06:47.370 seems counterintuitive. It's subtle, isn't it? 00:06:47.720 --> 00:06:49.579 I think it gets into the nuance of these models. 00:06:49.779 --> 00:06:52.500 Fluid dynamics, while complex, is heavily based 00:06:52.500 --> 00:06:55.439 on known equations. You apply the formulas. Achieving 00:06:55.439 --> 00:06:57.560 perfect geometric precision in a complex simulation 00:06:57.560 --> 00:07:00.139 like the beehive, that seems to need a different 00:07:00.139 --> 00:07:02.540 kind of extreme attention to detail. To coordinate 00:07:02.540 --> 00:07:05.300 systems, object relationships, maybe it's just 00:07:05.300 --> 00:07:07.730 harder to specify perfectly in a prompt. You 00:07:07.730 --> 00:07:09.629 know, I still wrestle with this myself sometimes. 00:07:09.810 --> 00:07:13.009 I find myself expecting absolute, almost scientific 00:07:13.009 --> 00:07:15.329 perfection from these single prompt outputs, 00:07:15.610 --> 00:07:18.470 even when they're incredibly complex. It's easy 00:07:18.470 --> 00:07:20.569 to forget they're working from learned patterns, 00:07:20.689 --> 00:07:23.209 not some fundamental understanding of mathematical 00:07:23.209 --> 00:07:25.750 truth, you know. That's the vulnerable admission, 00:07:26.009 --> 00:07:27.689 right? We all kind of do that. And these small 00:07:27.689 --> 00:07:30.569 failures, like the beehive alignment, they're 00:07:30.569 --> 00:07:33.350 useful. They show us precisely where the current 00:07:33.350 --> 00:07:36.350 limits are, and where careful prompting and maybe 00:07:36.350 --> 00:07:39.050 multi -step generation are still key. So why 00:07:39.050 --> 00:07:41.069 is understanding these little stumbles, like 00:07:41.069 --> 00:07:43.850 the beehive example, just as important as celebrating 00:07:43.850 --> 00:07:46.910 the big wins? It highlights where GPT -5 still 00:07:46.910 --> 00:07:49.310 holds an edge, particularly in tasks needing 00:07:49.310 --> 00:07:52.269 extreme geometric precision and intricate detail. 00:07:52.550 --> 00:07:54.810 Okay, really interesting. Let's take a quick 00:07:54.810 --> 00:07:57.009 pause here. When we come back, we'll dig into 00:07:57.009 --> 00:07:59.470 what might be the ultimate test for professional 00:07:59.470 --> 00:08:02.459 use. Can you actually trust it? We're talking 00:08:02.459 --> 00:08:05.800 reliability, zero hallucination, especially in 00:08:05.800 --> 00:08:08.319 high stakes areas like finance and scientific 00:08:08.319 --> 00:08:10.629 research. Welcome back to the Deep Dive. We're 00:08:10.629 --> 00:08:13.350 talking about the open source model Kimi K2 Thinking. 00:08:13.629 --> 00:08:16.269 So for any business, any receptor listening, 00:08:16.550 --> 00:08:19.310 reliability is paramount, right? A cool demo 00:08:19.310 --> 00:08:21.529 is one thing, but if the output isn't accurate, 00:08:21.750 --> 00:08:24.529 it's useless, maybe even dangerous. Let's get 00:08:24.529 --> 00:08:26.410 into that financial analysis use case mentioned 00:08:26.410 --> 00:08:28.350 in the sources. Yeah, this sounds like a killer 00:08:28.350 --> 00:08:31.149 app scenario because it tests really deep reasoning 00:08:31.149 --> 00:08:34.110 across multiple dense documents. So they fed 00:08:34.110 --> 00:08:37.269 it Q4 Financial Reports Think Thick PDS from 00:08:37.269 --> 00:08:40.980 Google, NVIDIA. Amazon. And the task was compare 00:08:40.980 --> 00:08:43.620 them, create charts, pull out key insights. That's 00:08:43.620 --> 00:08:45.899 tough. It's not just summarizing one doc. It 00:08:45.899 --> 00:08:47.879 has to find the same metrics across different 00:08:47.879 --> 00:08:49.799 report structures, different accounting styles, 00:08:49.919 --> 00:08:51.899 and pull exact numbers correctly from all of 00:08:51.899 --> 00:08:53.820 them. And the results. According to the source 00:08:53.820 --> 00:08:57.820 material, the accuracy was shocking. It nailed 00:08:57.820 --> 00:09:01.559 YouTube ads revenue. $10 .5 billion. Correct. 00:09:01.840 --> 00:09:04.259 It correctly pulled out NVIDIA's absolutely insane 00:09:04.259 --> 00:09:08.769 12 ,264 % year -over -year growth. Correct. The 00:09:08.769 --> 00:09:11.070 claim is the numbers were 100 % right across 00:09:11.070 --> 00:09:13.929 all three of these super dense reports. I mean, 00:09:13.929 --> 00:09:16.330 that level of precision, synthesizing hundreds 00:09:16.330 --> 00:09:19.190 of pages, that could save a financial analyst 00:09:19.190 --> 00:09:21.889 days of manual grunt work. OK, so if it builds 00:09:21.889 --> 00:09:23.809 trust in finance, what about really specialized 00:09:23.809 --> 00:09:26.389 science? It was tested on researching Alexander 00:09:26.389 --> 00:09:28.970 disease, a rare neurological disorder. Right. 00:09:29.049 --> 00:09:30.909 And here it apparently used its thinking and 00:09:30.909 --> 00:09:33.330 search modes, which sound agentic. Agentic basically 00:09:33.330 --> 00:09:35.190 means the model doesn't just respond. It can 00:09:35.190 --> 00:09:37.289 plan and execute steps like a human researcher. 00:09:37.320 --> 00:09:39.220 Okay, I need to search for papers, read them, 00:09:39.360 --> 00:09:41.860 synthesize findings, structure a report. And 00:09:41.860 --> 00:09:43.820 the quality of that final report after a process, 00:09:43.940 --> 00:09:46.440 what, 48 different research results? The claim 00:09:46.440 --> 00:09:49.220 is publication quality. It apparently generated 00:09:49.220 --> 00:09:51.460 detailed flowcharts mapping out the molecular 00:09:51.460 --> 00:09:54.279 pathophysiology, a clear diagnostic pathway. 00:09:54.840 --> 00:09:57.779 And crucially, it included a timely update about 00:09:57.779 --> 00:10:00.340 an expected FDA filing for a potential treatment 00:10:00.340 --> 00:10:03.519 in Q1 2026. That's not just summarizing old info. 00:10:03.600 --> 00:10:05.379 That's pulling cutting -edge, relevant details. 00:10:05.659 --> 00:10:09.120 Super sophisticated synthesis. Wow. Okay, and 00:10:09.120 --> 00:10:12.080 then the ultimate acid test for trust, the hallucination 00:10:12.080 --> 00:10:14.659 trap. They asked about stable diffusion 5, which 00:10:14.659 --> 00:10:17.600 doesn't exist. Exactly. This is a classic failure 00:10:17.600 --> 00:10:20.240 point for LLMs. They often just confidently invent 00:10:20.240 --> 00:10:22.240 plausible -sounding details about things that 00:10:22.240 --> 00:10:26.120 aren't real. perfectly. It didn't invent anything 00:10:26.120 --> 00:10:29.240 about SD5. Instead, it correctly stated it doesn't 00:10:29.240 --> 00:10:31.379 exist and provided accurate info on the actual 00:10:31.379 --> 00:10:34.480 current version, SD3 .5. That kind of reliability, 00:10:34.840 --> 00:10:37.259 refusing to just make stuff up, that's absolutely 00:10:37.259 --> 00:10:38.860 critical if you're going to use this in a professional 00:10:38.860 --> 00:10:41.279 setting. And the sources also mentioned quick 00:10:41.279 --> 00:10:43.259 hits like successfully creating an interactive 00:10:43.259 --> 00:10:46.779 gut bacteria taxonomy tree, quite niche, and 00:10:46.779 --> 00:10:48.960 an interactive physics course that perfectly 00:10:48.960 --> 00:10:51.840 modeled kinematics. So the picture emerging is 00:10:51.840 --> 00:10:54.179 one of reliable, sophisticated, and importantly 00:10:54.179 --> 00:10:57.460 trustworthy performance in complex domains. So 00:10:57.460 --> 00:10:59.940 given that stellar performance in finance, science, 00:10:59.960 --> 00:11:02.259 and the hallucination test, what's the biggest 00:11:02.259 --> 00:11:04.629 hurdle left for companies wanting to... adopt 00:11:04.629 --> 00:11:07.750 Kimi K2? Probably scaling its deployment, right? 00:11:07.830 --> 00:11:09.789 And integrating it smoothly into their existing 00:11:09.789 --> 00:11:12.490 workflows and tech infrastructure. So let's try 00:11:12.490 --> 00:11:14.990 to wrap this up. What Kimi K2 seems to represent, 00:11:15.250 --> 00:11:17.750 it feels like a really significant, maybe permanent 00:11:17.750 --> 00:11:20.549 shift in the AI power balance. It's clearly a 00:11:20.549 --> 00:11:22.710 powerhouse. It offers capability that's right 00:11:22.710 --> 00:11:24.970 up there near GPT -5, but it's free, it can be 00:11:24.970 --> 00:11:27.309 run privately, and it's proven capable of generating 00:11:27.309 --> 00:11:30.190 complex working apps and highly accurate analysis 00:11:30.190 --> 00:11:33.090 in demanding fields. Yeah, the strategic takeaway 00:11:33.090 --> 00:11:34.950 for anyone... When listening, developers, business 00:11:34.950 --> 00:11:37.590 leaders is pretty clear, I think. Open source 00:11:37.590 --> 00:11:40.389 has definitively closed the quality gap with 00:11:40.389 --> 00:11:42.750 the top proprietary models. You might no longer 00:11:42.750 --> 00:11:44.730 have to choose between the absolute best performance 00:11:44.730 --> 00:11:48.029 and having control over your data, your cost, 00:11:48.070 --> 00:11:50.649 your infrastructure. That choice is changing. 00:11:50.830 --> 00:11:53.490 And the source material hints that part two is 00:11:53.490 --> 00:11:55.429 going to dive into the tech specs specifically. 00:11:55.549 --> 00:11:59.149 It's one trillion parameter mixture of experts 00:11:59.149 --> 00:12:03.960 or Moe architecture. That Moe approach is probably 00:12:03.960 --> 00:12:06.679 key to how it achieves this performance while 00:12:06.679 --> 00:12:09.639 staying, well, manageable enough to be open sourced. 00:12:09.720 --> 00:12:12.279 Right. So maybe here's a final thought for you 00:12:12.279 --> 00:12:14.899 to chew on after this deep dive. If a free open 00:12:14.899 --> 00:12:17.159 source model can already do this, what happens 00:12:17.159 --> 00:12:19.080 next? What happens when the cost of the hardware 00:12:19.080 --> 00:12:21.259 needed to run a model like this drops low enough 00:12:21.259 --> 00:12:23.460 that basically every small team, every consultant, 00:12:23.519 --> 00:12:26.460 every startup can have their own private, powerful, 00:12:26.779 --> 00:12:30.120 maybe even custom -tuned AI? That feels like 00:12:30.120 --> 00:12:32.279 the next wave of disruption coming. Definitely 00:12:32.279 --> 00:12:33.840 something to think about. Keep digging.