WEBVTT 00:00:00.000 --> 00:00:02.020 So here's something kind of paradoxical to chew 00:00:02.020 --> 00:00:05.360 on. Imagine running an AI model. It knows as 00:00:05.360 --> 00:00:07.320 much as like a massive one trillion parameter 00:00:07.320 --> 00:00:11.019 machine, but it runs with the speed, the latency, 00:00:11.140 --> 00:00:14.380 the cost of a model that's, what, 30 times smaller? 00:00:14.560 --> 00:00:17.980 Only 32 billion parameters active? That kind 00:00:17.980 --> 00:00:20.359 of efficiency. Right. It's genuinely stunning 00:00:20.359 --> 00:00:22.120 from an engineering perspective. Right. And this 00:00:22.120 --> 00:00:24.440 is in vaporware, not some research paper dream. 00:00:24.559 --> 00:00:28.600 It's Kimi K2 thinking built by Moonshot AI. And 00:00:28.600 --> 00:00:30.980 it's the open source challenger that's, well, 00:00:31.079 --> 00:00:33.420 it's really shaking things up because independent 00:00:33.420 --> 00:00:35.719 leaderboards, they now show it sitting globally 00:00:35.719 --> 00:00:39.100 at number two, just right behind GPT -5. Welcome 00:00:39.100 --> 00:00:41.140 to the deep dive. Today we're taking a close 00:00:41.140 --> 00:00:42.679 look, a measured look, at the source material 00:00:42.679 --> 00:00:45.520 we have on KimiK2 and, you know, what this really 00:00:45.520 --> 00:00:47.679 represents for the whole AI landscape. It feels 00:00:47.679 --> 00:00:49.939 like a pretty big shift. Yeah, our mission today 00:00:49.939 --> 00:00:52.039 is kind of threefold. First, we're going to really 00:00:52.039 --> 00:00:55.340 dig into the architecture, the engine under the 00:00:55.340 --> 00:00:58.179 hood, right? What makes KimiK2 so powerful but 00:00:58.179 --> 00:01:01.000 also so efficient, so affordable? Second, we'll 00:01:01.000 --> 00:01:03.560 look at the benchmarks. The proof, essentially, 00:01:03.740 --> 00:01:05.939 that open source isn't just catching up anymore. 00:01:06.060 --> 00:01:08.140 In some ways, it's actually pulling ahead of 00:01:08.140 --> 00:01:10.900 the big, close source players. And finally, we 00:01:10.900 --> 00:01:13.040 need to unpack what this means, the implications 00:01:13.040 --> 00:01:17.239 for things like data privacy, cost control, corporate 00:01:17.239 --> 00:01:19.519 independence even. It's pretty revolutionary 00:01:19.519 --> 00:01:21.420 stuff. Okay, let's start with that core tech 00:01:21.420 --> 00:01:24.019 then. The big question is how. How does Moonshot 00:01:24.019 --> 00:01:27.819 AI... get this like massive power, but keep the 00:01:27.819 --> 00:01:30.299 efficiency so incredibly high. It really comes 00:01:30.299 --> 00:01:31.959 down to the architecture. They're using what's 00:01:31.959 --> 00:01:34.500 called a mixture of experts model and Moe. So 00:01:34.500 --> 00:01:36.819 instead of one giant brain that like lights up 00:01:36.819 --> 00:01:38.599 everywhere for every query, think of it more 00:01:38.599 --> 00:01:40.620 like a huge team, a team of very specialized 00:01:40.620 --> 00:01:43.599 experts. Ah, okay. So if I ask a really specific 00:01:43.599 --> 00:01:46.299 complex question, maybe something technical about, 00:01:46.379 --> 00:01:48.519 I don't know, protein folding, the system is 00:01:48.519 --> 00:01:50.840 smart enough to route my query only to the few 00:01:50.840 --> 00:01:53.340 experts who actually know about protein folding. 00:01:54.299 --> 00:01:56.739 And the rest kind of stay quiet. Exactly. That's 00:01:56.739 --> 00:01:59.849 the core idea. In simple terms, Moe is a vast 00:01:59.849 --> 00:02:02.390 network that only lights up key sections. It 00:02:02.390 --> 00:02:04.930 achieves what they call sparsity. And that sparsity, 00:02:05.030 --> 00:02:06.569 the fact that most of the network isn't being 00:02:06.569 --> 00:02:09.469 used for any given task, that's the key to the 00:02:09.469 --> 00:02:11.409 efficiency numbers. So the total model size, 00:02:11.669 --> 00:02:15.349 the whole knowledge base is indeed that massive 00:02:15.349 --> 00:02:17.969 1 trillion parameters. Right up there with the 00:02:17.969 --> 00:02:19.770 biggest models out there. But, and this is the 00:02:19.770 --> 00:02:22.310 kicker for any single request you make, the activated 00:02:22.310 --> 00:02:24.469 parameters, the parts that actually spin up and 00:02:24.469 --> 00:02:28.520 generate the answer, only 32 billion. 3 .2 % 00:02:28.520 --> 00:02:31.080 sparsity. So you're getting this like incredibly 00:02:31.080 --> 00:02:34.819 deep knowledge base, but the speed, the cost, 00:02:34.960 --> 00:02:37.000 it feels like you're running a much, much smaller 00:02:37.000 --> 00:02:39.620 model. That's a huge strategic advantage, isn't 00:02:39.620 --> 00:02:41.539 it? Just from an engineering and cost perspective, 00:02:41.659 --> 00:02:43.419 it completely changes the deployment economics. 00:02:43.840 --> 00:02:46.439 Yeah. If you only need to power up 3 .2 % of 00:02:46.439 --> 00:02:49.599 the network, you're drastically cutting down 00:02:49.599 --> 00:02:52.319 the compute needed, the energy draw, the chip 00:02:52.319 --> 00:02:54.240 requirements for inference. Oh, absolutely. Think 00:02:54.240 --> 00:02:57.330 about traditional dense models. Every single 00:02:57.330 --> 00:03:00.389 parameter is involved every single time. That 00:03:00.389 --> 00:03:03.969 takes immense continuous power. But Kimi K2's 00:03:03.969 --> 00:03:06.330 Moe design, it means companies can scale up how 00:03:06.330 --> 00:03:09.669 many queries they handle without needing exponentially 00:03:09.669 --> 00:03:12.789 more hardware. It makes top -tier AI suddenly 00:03:12.789 --> 00:03:15.110 much more accessible, runnable on, let's say, 00:03:15.210 --> 00:03:17.770 less extreme hardware clusters. And beyond just 00:03:17.770 --> 00:03:19.930 the efficiency, the capabilities themselves sound 00:03:19.930 --> 00:03:22.590 pretty formidable. Context length is 256 ,000 00:03:22.590 --> 00:03:25.030 tokens. That's definitely big enough to handle 00:03:25.030 --> 00:03:27.509 whole code bases. Right. Or long legal docs, 00:03:27.650 --> 00:03:29.830 years of financial reports, maybe all in one 00:03:29.830 --> 00:03:31.800 go. For sure. And then there's this other metric, 00:03:31.900 --> 00:03:35.539 the agentic capability one, sequential tool calls. 00:03:35.759 --> 00:03:39.379 The sources say Kimi K2 can handle 200, maybe 00:03:39.379 --> 00:03:42.319 300 plus tool calls in a row without a human 00:03:42.319 --> 00:03:44.560 stepping in. That means you can give it a really 00:03:44.560 --> 00:03:47.199 complex multi -step plan, something like analyze 00:03:47.199 --> 00:03:49.819 sentiment for these five stocks, pull relevant 00:03:49.819 --> 00:03:52.659 news from the last quarter, draft a summary email, 00:03:52.879 --> 00:03:55.520 schedule the meeting, and it can just go and 00:03:55.520 --> 00:03:58.319 execute almost that entire workflow autonomously. 00:03:58.460 --> 00:04:00.900 Okay, that's impressive. 300 sequential steps 00:04:00.900 --> 00:04:04.620 without intervention. So given that capability, 00:04:04.919 --> 00:04:07.460 what's the biggest hurdle people actually face 00:04:07.460 --> 00:04:09.259 when they try to use this for really complex 00:04:09.259 --> 00:04:12.159 autonomous agent tasks? Where does it still struggle? 00:04:12.460 --> 00:04:14.460 Yeah, that's a good question, honestly. I still 00:04:14.460 --> 00:04:16.800 wrestle with prompt drift myself sometimes, keeping 00:04:16.800 --> 00:04:19.160 agents locked onto the original goal across really 00:04:19.160 --> 00:04:21.000 long workflows. It's kind of like, you know, 00:04:21.019 --> 00:04:23.600 giving someone a 10 -step task. By step 7 or 00:04:23.600 --> 00:04:25.180 8, they might just slightly misunderstand the 00:04:25.180 --> 00:04:27.000 original intent because of tiny errors building 00:04:27.000 --> 00:04:29.319 up. Mm -hmm, that makes sense. A bit of cumulative 00:04:29.319 --> 00:04:31.339 error. Okay, that vulnerability is important 00:04:31.339 --> 00:04:34.519 context. Now let's shift to the proof. The numbers, 00:04:34.699 --> 00:04:37.540 the benchmarks that seem to back up this claim 00:04:37.540 --> 00:04:40.120 that open source is really competing at the top 00:04:40.120 --> 00:04:41.600 level now. Right. This is where the story gets 00:04:41.600 --> 00:04:44.060 really interesting. So for general reasoning, 00:04:44.160 --> 00:04:46.319 and especially in these agentic benchmarks testing 00:04:46.319 --> 00:04:48.759 how well it uses tools and follows multi -step 00:04:48.759 --> 00:04:51.600 plans, Kimi K2 is consistently beating several 00:04:51.600 --> 00:04:54.540 top closed source models. It just seems better 00:04:54.540 --> 00:04:56.420 at actually applying its knowledge through tools. 00:04:56.579 --> 00:04:59.519 And coding. The results there sound pretty clear 00:04:59.519 --> 00:05:01.839 cut. The sources mentioned competitive programming 00:05:01.839 --> 00:05:04.360 challenges. And Kini K2 apparently got the highest 00:05:04.360 --> 00:05:08.459 score, decisively beating Claude 4 .5. That points 00:05:08.459 --> 00:05:10.800 to really strong logic and cogeneration. Yeah, 00:05:10.819 --> 00:05:13.839 really strong. And even in high -level academic 00:05:13.839 --> 00:05:17.800 stuff. They use this tough benchmark, GPQA, Diamond, 00:05:18.019 --> 00:05:20.740 graduate -level science questions. Kimi K2 is 00:05:20.740 --> 00:05:23.100 a beast there, too. Beats Cloud 4 .5, again, 00:05:23.180 --> 00:05:25.000 pretty handily. It's only slightly behind Grok 00:05:25.000 --> 00:05:28.300 4, scoring like 87 .5%. Really impressive. But 00:05:28.300 --> 00:05:30.259 the big headline, the thing that really feels 00:05:30.259 --> 00:05:33.319 like a turning point, is that independent global 00:05:33.319 --> 00:05:37.680 leaderboard ranking. For so long... The top spots, 00:05:37.899 --> 00:05:40.420 maybe the top five, were all proprietary models, 00:05:40.639 --> 00:05:43.819 closed systems. Now you look at the list, and 00:05:43.819 --> 00:05:46.040 okay, GPT -5 high is number one, yes, proprietary. 00:05:46.319 --> 00:05:48.779 But right there at number two, globally, it's 00:05:48.779 --> 00:05:51.699 Kimi K2 thinking. An open source model, setting 00:05:51.699 --> 00:05:54.180 higher than Grot 4, higher than Claude 4 .5, 00:05:54.240 --> 00:05:57.560 higher than Gemini 2 .5 Pro. It's huge validation 00:05:57.560 --> 00:06:00.420 for this whole Moe approach. And, you know, if 00:06:00.420 --> 00:06:03.420 performing at that level wasn't enough. We absolutely 00:06:03.420 --> 00:06:05.500 have to talk about the cost. Right. Because Kimi 00:06:05.500 --> 00:06:07.660 K2 is apparently way cheaper than the other top 00:06:07.660 --> 00:06:09.800 models. The numbers suggest it's roughly three 00:06:09.800 --> 00:06:12.600 times cheaper than GPT -5. Yeah, around 3x cheaper 00:06:12.600 --> 00:06:14.939 than GPT -5. But the gap gets even wider compared 00:06:14.939 --> 00:06:16.620 to some others. We're talking like a six -fold 00:06:16.620 --> 00:06:19.899 cost advantage over a Quad 4 .5 and Grok 4. Six 00:06:19.899 --> 00:06:21.939 times cheaper. And remember, you can potentially 00:06:21.939 --> 00:06:23.699 run it on less expensive hardware because of 00:06:23.699 --> 00:06:26.180 that sparsity. So the TCO, the total cost of 00:06:26.180 --> 00:06:28.819 ownership for a business, just plummets. So that 00:06:28.819 --> 00:06:31.079 leads to the practical question. Does that... 00:06:31.449 --> 00:06:35.170 huge cost saving, maybe six times cheaper. Does 00:06:35.170 --> 00:06:38.149 that outweigh the perhaps very small performance 00:06:38.149 --> 00:06:41.129 difference compared to the absolute number one 00:06:41.129 --> 00:06:45.529 GPT -5 for most everyday business uses? I think 00:06:45.529 --> 00:06:48.350 for most professional use cases, yes, the massive 00:06:48.350 --> 00:06:50.829 cost savings are an easy choice. The value proposition 00:06:50.829 --> 00:06:53.889 is just incredibly strong. Okay. This leads us 00:06:53.889 --> 00:06:56.170 nicely into the more philosophical side, which 00:06:56.170 --> 00:06:58.050 might actually be the most critical part of this 00:06:58.050 --> 00:07:00.769 whole story. Moonshot AI didn't just build this 00:07:00.769 --> 00:07:03.269 incredibly powerful model. They decided to release 00:07:03.269 --> 00:07:05.589 it with open weights. We should probably clarify 00:07:05.589 --> 00:07:07.509 what that means exactly. Right. Open weights 00:07:07.509 --> 00:07:10.480 means the core of the model. The actual train 00:07:10.480 --> 00:07:12.860 parameters, that one trillion number is freely 00:07:12.860 --> 00:07:14.779 downloadable. Anyone can grab it and run it. 00:07:14.839 --> 00:07:17.000 It's not quite fully open source, like, say, 00:07:17.100 --> 00:07:19.139 Linux, where maybe all the training data and 00:07:19.139 --> 00:07:21.399 code are also open. But the model's intelligence, 00:07:21.639 --> 00:07:24.000 its brain is out there. And we should probably 00:07:24.000 --> 00:07:25.899 stress who this is really for. This isn't something 00:07:25.899 --> 00:07:27.899 you just, you know, download to your laptop on 00:07:27.899 --> 00:07:30.139 a whim. That model file is apparently around 00:07:30.139 --> 00:07:33.920 600 gigabytes. You need a serious GPU cluster. 00:07:33.959 --> 00:07:36.220 Right. We're talking. potentially hundreds of 00:07:36.220 --> 00:07:38.420 thousands of dollars in hardware to run it well. 00:07:38.579 --> 00:07:41.980 Exactly. It's aimed squarely at companies, serious 00:07:41.980 --> 00:07:44.540 research labs, developers building significant 00:07:44.540 --> 00:07:47.459 applications. And it's a very strategic move, 00:07:47.540 --> 00:07:50.480 really. It's positioned as a direct answer to 00:07:50.480 --> 00:07:52.519 what some people call the closed source problem 00:07:52.519 --> 00:07:56.540 that's kind of dominated high -end AI until now. 00:07:56.680 --> 00:07:59.480 And that problem. That basically means being 00:07:59.480 --> 00:08:02.600 totally reliant on a few big, mostly American 00:08:02.600 --> 00:08:06.019 AI labs, which creates some real strategic vulnerabilities 00:08:06.019 --> 00:08:08.379 for businesses in other countries. Absolutely. 00:08:08.439 --> 00:08:11.000 You've got serious vendor lock -in. Your entire 00:08:11.000 --> 00:08:13.980 AI capability could depend on one company's pricing 00:08:13.980 --> 00:08:16.699 whims or sudden policy shifts. And critically, 00:08:16.920 --> 00:08:19.600 the data privacy issue is huge. To use those 00:08:19.600 --> 00:08:21.480 closed models, you have to constantly send your 00:08:21.480 --> 00:08:23.720 sensitive proprietary data out to their servers 00:08:23.720 --> 00:08:26.259 for processing. There's often very little transparency 00:08:26.259 --> 00:08:29.170 into how it's used. secured, it's complete dependence. 00:08:29.569 --> 00:08:31.550 So the open source approach, particularly with 00:08:31.550 --> 00:08:34.330 a powerful model like Kini K2 available with 00:08:34.330 --> 00:08:37.250 open weights, it just flips that whole dynamic. 00:08:37.269 --> 00:08:39.470 It offers data sovereignty. Total data control. 00:08:39.610 --> 00:08:42.070 Because you download the model, you run it on 00:08:42.070 --> 00:08:44.370 your hardware inside your secure perimeter, your 00:08:44.370 --> 00:08:46.690 confidential health data, your financial projections. 00:08:47.419 --> 00:08:50.340 They never leave your infrastructure. Plus, you 00:08:50.340 --> 00:08:53.179 get full control over customization. You can 00:08:53.179 --> 00:08:55.460 fine tune it deeply for your specific industry 00:08:55.460 --> 00:08:58.639 or tasks. And crucially, you gain independence 00:08:58.639 --> 00:09:01.860 from external pricing, external rules, external 00:09:01.860 --> 00:09:04.279 regulators. It must have taken a massive commitment, 00:09:04.320 --> 00:09:07.159 though, for Moonshot AI to spend what must have 00:09:07.159 --> 00:09:09.980 been astronomical sums developing something this 00:09:09.980 --> 00:09:12.820 close to GPT -5 level. And then essentially just... 00:09:13.370 --> 00:09:15.490 Give the engine away for free. It really is like 00:09:15.490 --> 00:09:17.750 a gift to the global research and development 00:09:17.750 --> 00:09:20.250 community. It just turbocharges innovation everywhere 00:09:20.250 --> 00:09:22.710 because now everyone can access and build on 00:09:22.710 --> 00:09:24.509 a truly state -of -the -art foundation model. 00:09:24.830 --> 00:09:28.169 Whoa. Imagine scaling to a billion queries without 00:09:28.169 --> 00:09:31.330 reliance on external companies. That's just pure 00:09:31.330 --> 00:09:34.269 innovation fuel for countless startups, for corporate 00:09:34.269 --> 00:09:36.919 R &D labs everywhere. And, you know, it puts 00:09:36.919 --> 00:09:39.580 immediate, intense pressure back on the closed 00:09:39.580 --> 00:09:43.159 labs. They now have to really justify those high 00:09:43.159 --> 00:09:45.799 subscription fees when something this good is 00:09:45.799 --> 00:09:48.320 available for free, provided you have the hardware. 00:09:48.519 --> 00:09:50.460 So putting aside the competitive angle for a 00:09:50.460 --> 00:09:52.860 second, what do you think is the single biggest 00:09:52.860 --> 00:09:55.919 benefit for global research when a foundational 00:09:55.919 --> 00:09:59.990 model this powerful is made open like this? Accelerated 00:09:59.990 --> 00:10:02.169 research happens because everyone can now build 00:10:02.169 --> 00:10:03.690 directly on a state -of -the -art foundation. 00:10:03.929 --> 00:10:06.590 It just raises the baseline for the entire field 00:10:06.590 --> 00:10:09.009 almost overnight. Okay, so let's get practical. 00:10:09.529 --> 00:10:11.889 How can someone listening right now actually 00:10:11.889 --> 00:10:14.049 access this power? Sound like there are basically 00:10:14.049 --> 00:10:16.649 two main ways. Option one, the simplest path 00:10:16.649 --> 00:10:18.950 is just using their online interface, right? 00:10:20.279 --> 00:10:22.480 Yeah, but the key there is you have to remember 00:10:22.480 --> 00:10:25.139 to enable thinking mode in the settings to actually 00:10:25.139 --> 00:10:28.320 access the Kimi K2 power. And that online version 00:10:28.320 --> 00:10:30.679 already has some pretty capable agent modes built 00:10:30.679 --> 00:10:32.960 in, like OK Computer for coding, which sounds 00:10:32.960 --> 00:10:35.419 quite autonomous, and Researcher for digging 00:10:35.419 --> 00:10:37.519 through and summarizing data. And then option 00:10:37.519 --> 00:10:40.700 two is for the, let's say, power users. the enterprises, 00:10:41.000 --> 00:10:43.879 the researchers needing maximum control. That's 00:10:43.879 --> 00:10:46.220 the local on -premise deployment, downloading 00:10:46.220 --> 00:10:48.379 the weights from Hugging Face. That's the road 00:10:48.379 --> 00:10:50.200 if you're dealing with highly sensitive data 00:10:50.200 --> 00:10:53.289 like health or finance. and need absolute 100 00:10:53.289 --> 00:10:56.889 % control. Exactly. So let's kind of sum up where 00:10:56.889 --> 00:11:00.149 Kimi K2 really shines. Its superpowers, if you 00:11:00.149 --> 00:11:03.169 will. First, definitely coding and software development. 00:11:03.389 --> 00:11:06.070 The reasoning ability for code seems top -notch. 00:11:06.129 --> 00:11:09.429 Second, building AI agents. That high number 00:11:09.429 --> 00:11:12.210 of sequential tool calls makes it ideal for complex 00:11:12.210 --> 00:11:15.419 autonomous tasks. Third, probably high level 00:11:15.419 --> 00:11:18.200 scientific and financial research. It seems exceptionally 00:11:18.200 --> 00:11:20.379 good at pulling together and reasoning over complex 00:11:20.379 --> 00:11:22.779 technical information. And of course, anytime 00:11:22.779 --> 00:11:25.399 cost is a major factor or when data privacy is 00:11:25.399 --> 00:11:27.440 absolutely mandatory. But we should be fair and 00:11:27.440 --> 00:11:29.919 note where the alternatives might still have 00:11:29.919 --> 00:11:32.279 a slight advantage. GPT -5, for instance. Yeah. 00:11:32.360 --> 00:11:34.379 The sources still suggest it might be a bit better 00:11:34.379 --> 00:11:36.740 at extremely complex, maybe more creative or 00:11:36.740 --> 00:11:38.720 abstract tasks. Right. They mentioned things 00:11:38.720 --> 00:11:41.240 like the beehive simulation example, scenarios 00:11:41.240 --> 00:11:43.460 needing really complex, maybe edge case physics. 00:11:43.500 --> 00:11:46.659 understanding, Kimi K2 apparently struggled a 00:11:46.659 --> 00:11:49.360 bit more there. So it suggests that while the 00:11:49.360 --> 00:11:52.139 general knowledge is vast, maybe the absolute 00:11:52.139 --> 00:11:55.360 peak of abstract complex reasoning isn't quite 00:11:55.360 --> 00:11:58.909 at the GPT -5 level yet. minor gaps. And context 00:11:58.909 --> 00:12:01.309 window length is another one. If you absolutely 00:12:01.309 --> 00:12:04.429 need the biggest possible window, say you're 00:12:04.429 --> 00:12:06.870 feeding it a massive 900 page book or dozens 00:12:06.870 --> 00:12:11.070 of dense PDFs at once, Gemini 2 .5 Pro's 1 million 00:12:11.070 --> 00:12:13.490 token window is still the leader there, right? 00:12:13.629 --> 00:12:15.789 Correct. So, yeah, the tradeoff seems pretty 00:12:15.789 --> 00:12:18.730 clear. You might accept a tiny, maybe negligible 00:12:18.730 --> 00:12:21.389 hit in absolute top end edge case quality. But 00:12:21.389 --> 00:12:23.350 in return, you get enormous gains in openness, 00:12:23.629 --> 00:12:25.789 total data privacy if you run it locally and 00:12:25.789 --> 00:12:27.889 that, you know, potentially. Six full cost efficiency. 00:12:28.129 --> 00:12:30.389 The final question on readiness. Considering 00:12:30.389 --> 00:12:32.409 those known weaknesses like edge case physics, 00:12:32.509 --> 00:12:34.870 some minor layout bugs mentioned, is Kimi K2 00:12:34.870 --> 00:12:36.769 truly ready for mission critical high stakes 00:12:36.769 --> 00:12:39.590 business use today? I'd say yes, especially for 00:12:39.590 --> 00:12:42.809 privacy critical work. The ability to have complete 00:12:42.809 --> 00:12:45.610 control over your data often outweighs those 00:12:45.610 --> 00:12:49.269 known minor bugs. That independence is frequently 00:12:49.269 --> 00:12:52.220 the deciding factor. Okay, let's try and bring 00:12:52.220 --> 00:12:54.480 this all together then, the core thesis of our 00:12:54.480 --> 00:12:57.340 deep talk today. It seems to me that Kimi K2's 00:12:57.340 --> 00:13:01.000 success really proves that open source AI isn't 00:13:01.000 --> 00:13:03.100 just playing catch -up anymore. It's actually 00:13:03.100 --> 00:13:06.080 leading in some really important areas. And it's 00:13:06.080 --> 00:13:09.100 firmly established itself as a peer, performance 00:13:09.100 --> 00:13:12.179 -wise, to the big proprietary models. It really 00:13:12.179 --> 00:13:14.559 feels like a historic shift. Think back just 00:13:14.559 --> 00:13:17.419 a few months. The top five models on those leaderboards, 00:13:17.600 --> 00:13:20.399 all closed systems. Innovation locked behind. 00:13:20.490 --> 00:13:23.289 paywalls. Now, the number two spot globally is 00:13:23.289 --> 00:13:25.129 held by an open source model that puts incredible 00:13:25.129 --> 00:13:27.429 pressure on big tech. They have to innovate faster, 00:13:27.590 --> 00:13:29.809 sure, but they also probably have to bring prices 00:13:29.809 --> 00:13:32.389 down for everyone needing AI. So for you, the 00:13:32.389 --> 00:13:34.629 listener, the takeaway is that you now have a 00:13:34.629 --> 00:13:37.850 real meaningful choice. It's not just about convenience 00:13:37.850 --> 00:13:39.850 versus nothing. It's convenience versus this 00:13:39.850 --> 00:13:42.470 potent combination of power, privacy, and cost 00:13:42.470 --> 00:13:45.220 control if you go the open way route. Yeah. So 00:13:45.220 --> 00:13:47.000 maybe here's a final thought to leave you with 00:13:47.000 --> 00:13:50.100 something to mull over. If a top tier open source 00:13:50.100 --> 00:13:53.159 model can basically match performance in most 00:13:53.159 --> 00:13:56.259 areas while potentially costing six times less 00:13:56.259 --> 00:13:59.679 to operate, what really is the long term sustainable 00:13:59.679 --> 00:14:02.519 value proposition for the closed models? How 00:14:02.519 --> 00:14:04.419 are they going to justify that significant price 00:14:04.419 --> 00:14:06.620 premium going forward? Definitely something to 00:14:06.620 --> 00:14:08.279 think about. We encourage you to check out the 00:14:08.279 --> 00:14:10.179 sources, look at the leaderboards and consider 00:14:10.179 --> 00:14:13.259 what this shift means for your own work or research. 00:14:13.799 --> 00:14:15.179 Thanks for joining us on this deep dive.