WEBVTT 00:00:00.000 --> 00:00:03.899 Imagine this. You take a sentence and about 60 00:00:03.899 --> 00:00:06.540 seconds later, you have a full website, not just 00:00:06.540 --> 00:00:10.400 text, but like playable video, user logins, the 00:00:10.400 --> 00:00:12.679 works. Yeah. Or creating, you know, cinematic 00:00:12.679 --> 00:00:14.660 video clips just like that. And then imagine 00:00:14.660 --> 00:00:18.120 getting all of that for free. That's really the 00:00:18.120 --> 00:00:20.460 heart of it. That's the quiet revolution happening 00:00:20.460 --> 00:00:23.899 in AI right now. Welcome to the Deep Dive. We're 00:00:23.899 --> 00:00:27.059 here to unpack complex topics and really pull 00:00:27.059 --> 00:00:28.980 out the key insights. And today, yeah, we're 00:00:28.980 --> 00:00:31.719 diving into what feels like a huge shift in artificial 00:00:31.719 --> 00:00:34.259 intelligence. For years, it felt like the best 00:00:34.259 --> 00:00:37.439 AI tools were locked away, right? Expensive subscriptions, 00:00:37.939 --> 00:00:40.619 complicated access. Totally, like an exclusive 00:00:40.619 --> 00:00:44.079 club. But now, especially with some big moves 00:00:44.079 --> 00:00:46.719 from Chinese tech firms, that whole model is 00:00:46.719 --> 00:00:48.880 getting flipped on its head. Flipped how? They're 00:00:48.880 --> 00:00:51.359 basically giving away truly world -class AI. 00:00:52.030 --> 00:00:54.810 For everyone. OK, so our mission today is to 00:00:54.810 --> 00:00:57.070 really get under the hood of this change. We 00:00:57.070 --> 00:01:00.270 need to explore this great AI divide, these like 00:01:00.270 --> 00:01:02.390 fundamentally different philosophies. Right. 00:01:02.469 --> 00:01:04.590 And we'll look at three specific examples that 00:01:04.590 --> 00:01:07.329 are already making waves. Minimax M1 for websites, 00:01:07.650 --> 00:01:13.250 CDance 1 .0 for video and KimiDev 72B for code 00:01:13.250 --> 00:01:16.319 fixing. A superhuman code doctor, I think you 00:01:16.319 --> 00:01:18.760 called it. Yeah, something like that. It's about 00:01:18.760 --> 00:01:22.560 showing how these tools give, well, superpowers 00:01:22.560 --> 00:01:25.379 to individuals, to small businesses. I don't 00:01:25.379 --> 00:01:28.000 like the playing field. Exactly. In ways that 00:01:28.000 --> 00:01:30.200 honestly seemed impossible just a little while 00:01:30.200 --> 00:01:32.680 ago. All right, let's unpack that core idea first. 00:01:32.900 --> 00:01:36.200 Why? Why give away something that costs? Presumably 00:01:36.200 --> 00:01:38.180 hundreds of millions to develop. It just seems 00:01:38.180 --> 00:01:40.840 to fly in the face of, you know, normal business 00:01:40.840 --> 00:01:43.120 strategy. It does seem counterintuitive from 00:01:43.120 --> 00:01:45.420 a Western perspective. Yeah. That model usually 00:01:45.420 --> 00:01:48.480 focuses on high value users, subscription fees, 00:01:48.719 --> 00:01:50.900 maximizing revenue from each customer willing 00:01:50.900 --> 00:01:53.099 to pay top dollar. Like you said, the exclusive 00:01:53.099 --> 00:01:55.819 club. Exactly. But the Eastern model, especially 00:01:55.819 --> 00:01:58.810 these big Chinese tech companies. It's playing 00:01:58.810 --> 00:02:01.790 a different game entirely. Their goal is domination 00:02:01.790 --> 00:02:04.209 through democratization. Domination through democratization. 00:02:04.390 --> 00:02:06.709 Okay. So they make these advanced tools free. 00:02:06.890 --> 00:02:10.210 Why? To get everyone using them. Massive adoption. 00:02:10.590 --> 00:02:12.770 Huge market penetration. And that feeds back 00:02:12.770 --> 00:02:15.210 into their systems. Precisely. It creates this 00:02:15.210 --> 00:02:17.650 enormous data feedback loop. They learn faster. 00:02:17.770 --> 00:02:19.530 Their models improve quicker because they have 00:02:19.530 --> 00:02:22.349 so much more usage data coming in. The end game 00:02:22.349 --> 00:02:25.289 isn't immediate profit per user. It's ecosystem 00:02:25.289 --> 00:02:28.770 control. Yes. become the foundation the free 00:02:28.770 --> 00:02:32.430 default and often the best choice then you essentially 00:02:32.430 --> 00:02:35.710 own the ecosystem so one sells access the other 00:02:35.710 --> 00:02:38.509 builds market share by giving power away that's 00:02:38.509 --> 00:02:40.310 a great way to put it control through distribution 00:02:40.310 --> 00:02:42.909 essentially okay that makes sense and it really 00:02:42.909 --> 00:02:45.229 sets the stage for these tools it's not theoretical 00:02:45.229 --> 00:02:48.379 anymore Let's talk Minimax M1, the instant website 00:02:48.379 --> 00:02:50.560 architect. Yeah, this one is a perfect example. 00:02:50.780 --> 00:02:53.259 I read this example that just blew my mind. Someone 00:02:53.259 --> 00:02:56.460 prompted it. Create a clone of the Netflix website 00:02:56.460 --> 00:02:59.240 complete with playable video trailers. And it 00:02:59.240 --> 00:03:01.300 wasn't just a mock -up. Nope. Fully interactive. 00:03:02.000 --> 00:03:04.340 Navigation, profiles, even the video trailers 00:03:04.340 --> 00:03:06.159 playing when you hover over them. Just like the 00:03:06.159 --> 00:03:08.379 real site. In about a minute. And this isn't 00:03:08.379 --> 00:03:10.500 just a slightly smarter chatbot spitting out 00:03:10.500 --> 00:03:13.139 HTML. This comes from what they call an open 00:03:13.139 --> 00:03:16.020 weight, large scale, hybrid reasoning model. 00:03:16.219 --> 00:03:19.639 Okay. Open weight, hybrid reasoning. Break that 00:03:19.639 --> 00:03:21.319 down a bit. What does that actually mean for 00:03:21.319 --> 00:03:24.560 someone using it? Well, open weight means the 00:03:24.560 --> 00:03:27.659 model's parameters. It's sort of brain. are publicly 00:03:27.659 --> 00:03:30.180 available. Anyone can use and build on it. Yeah. 00:03:30.280 --> 00:03:33.280 Under a very permissive license, Apache 2 .0, 00:03:33.360 --> 00:03:35.560 which is key for commercial use. Okay. So it's 00:03:35.560 --> 00:03:39.419 truly open. Right. And hybrid reasoning. Think 00:03:39.419 --> 00:03:40.960 of it like having different specialized tools. 00:03:41.139 --> 00:03:43.900 It combines different AI techniques. Plus, it 00:03:43.900 --> 00:03:46.360 has a massive context window, up to a million 00:03:46.360 --> 00:03:49.259 tokens. A million tokens. It's like processing 00:03:49.259 --> 00:03:51.340 multiple novels worth of information at once. 00:03:51.439 --> 00:03:53.219 Right. Pretty much. A token is just a small piece 00:03:53.219 --> 00:03:55.409 of text or code. So it can understand really 00:03:55.409 --> 00:03:57.830 complex requests like an entire website structure 00:03:57.830 --> 00:04:00.349 and content all in one go. And it's efficient, 00:04:00.449 --> 00:04:02.469 too, you mentioned. Extremely. It has something 00:04:02.469 --> 00:04:05.150 like 450 billion parameters total, which sounds 00:04:05.150 --> 00:04:08.270 huge. But for any given task, it only activates 00:04:08.270 --> 00:04:10.969 the relevant ones, maybe 45 billion. Like calling 00:04:10.969 --> 00:04:13.129 in only the specialists you need for a specific 00:04:13.129 --> 00:04:16.069 job. Exactly. Saves a ton of computing power. 00:04:16.230 --> 00:04:18.790 Makes it incredibly fast. So that efficiency 00:04:18.790 --> 00:04:21.180 and the open model. That's how they can offer 00:04:21.180 --> 00:04:23.720 it. How do people actually start using Minimax 00:04:23.720 --> 00:04:27.019 M1 today? It's pretty simple. Go to their website, 00:04:27.160 --> 00:04:29.000 sign up for a free account. They usually give 00:04:29.000 --> 00:04:30.939 you starting credits, maybe a thousand or so. 00:04:31.079 --> 00:04:34.000 Okay. I'd say start simple. Like create a portfolio 00:04:34.000 --> 00:04:37.120 website for a photographer, see what it does. 00:04:37.259 --> 00:04:39.660 Then you can iterate, maybe ask for changes. 00:04:39.899 --> 00:04:42.040 And then move to more complex stuff. Yeah. Then 00:04:42.040 --> 00:04:43.939 try something like that. Interactive dashboard 00:04:43.939 --> 00:04:47.000 for project deadlines example. Just build up 00:04:47.000 --> 00:04:49.439 complexity. What's the biggest hurdle people 00:04:49.439 --> 00:04:53.339 face when they first try it? Honestly, just believing 00:04:53.339 --> 00:04:55.860 it. Overcoming that initial thought of, this 00:04:55.860 --> 00:04:58.279 sounds way too good to be true. Huh, right. The 00:04:58.279 --> 00:05:00.600 psychological barrier. Because we're so used 00:05:00.600 --> 00:05:03.579 to powerful tech costing a lot. Exactly. There's 00:05:03.579 --> 00:05:06.920 this ingrained skepticism about free when it 00:05:06.920 --> 00:05:09.699 comes to high -end AI. It's about shifting your 00:05:09.699 --> 00:05:12.779 mindset from scarcity to, well, abundance in 00:05:12.779 --> 00:05:15.879 this space. That's a profound shift. Okay, let's 00:05:15.879 --> 00:05:20.329 talk video. For years, AI video felt clunky, 00:05:20.449 --> 00:05:23.889 short clips, weird motion. Yeah, definitely hit 00:05:23.889 --> 00:05:26.550 or miss, often more miss. But then ByteDance, 00:05:26.589 --> 00:05:30.009 the TikTok people, quietly dropped CDance 1 .0, 00:05:30.069 --> 00:05:32.689 and now it's apparently topping the leaderboards. 00:05:32.910 --> 00:05:36.069 It really is. The speed and quality are just 00:05:36.069 --> 00:05:39.610 remarkable. You can get a really nice 5 -second 00:05:39.610 --> 00:05:42.910 1080p video clip in about 40 seconds on just 00:05:42.910 --> 00:05:45.709 one regular graphics card. That's fast. But what 00:05:45.709 --> 00:05:48.069 makes it stand out beyond speed? The cinematic 00:05:48.069 --> 00:05:50.810 understanding. It can generate multi -shot sequences 00:05:50.810 --> 00:05:53.250 from a single prompt. You know, like start with 00:05:53.250 --> 00:05:55.350 a wide shot, then cut to a close -up, maybe add 00:05:55.350 --> 00:05:57.389 a tracking shot. It gets it. How does it learn 00:05:57.389 --> 00:05:59.269 that? That feels like a director's skill. It's 00:05:59.269 --> 00:06:01.449 down to the training. They use this multi -dimensional 00:06:01.449 --> 00:06:03.370 reward system, so it wasn't just learning follow 00:06:03.370 --> 00:06:05.290 the prompt or make it look pretty. It learned 00:06:05.290 --> 00:06:07.829 all those things together. Visuals, motion, following 00:06:07.829 --> 00:06:10.509 instructions, even scene composition. All at 00:06:10.509 --> 00:06:13.000 once. That seems to be the key difference. It 00:06:13.000 --> 00:06:15.360 understands how shops should connect, how motion 00:06:15.360 --> 00:06:17.300 should look natural within a scene. It's not 00:06:17.300 --> 00:06:19.800 just stitching images. Okay, so for someone wanting 00:06:19.800 --> 00:06:23.339 to use this, maybe a small business making ads 00:06:23.339 --> 00:06:26.240 or a creator, how do they access C -Dance? You'd 00:06:26.240 --> 00:06:28.019 usually go through a platform that hosts it, 00:06:28.060 --> 00:06:30.220 like Wavespeed AI is one example. And how would 00:06:30.220 --> 00:06:32.860 you start? Again, start simple. Text to video. 00:06:33.019 --> 00:06:35.040 A sharp knife cutting through a copper gear. 00:06:35.199 --> 00:06:37.860 Something clear. Okay. Then maybe try image to 00:06:37.860 --> 00:06:45.720 video. Upload a picture of, say, Gotcha. And 00:06:45.720 --> 00:06:47.699 the multi -shot. Yeah. Then describe a sequence. 00:06:48.920 --> 00:06:52.860 Scene one. Detective enters a dark room. Scene 00:06:52.860 --> 00:06:55.899 two. Clothes up on clues on a table. Scene three. 00:06:56.139 --> 00:06:58.600 Detective picks up an object. Looks thoughtful. 00:06:58.759 --> 00:07:00.959 It understands that structure. What is it fundamentally 00:07:00.959 --> 00:07:03.439 doing better to make the video look so real? 00:07:03.680 --> 00:07:06.420 It seems to grasp motion and physics better. 00:07:06.829 --> 00:07:08.910 It's not just generating plausible frames. It's 00:07:08.910 --> 00:07:10.889 generating plausible movement between frames. 00:07:11.069 --> 00:07:13.670 That makes sense. Physics is hard. Okay, on to 00:07:13.670 --> 00:07:17.310 the third one. KimiDev72B, the superhuman code 00:07:17.310 --> 00:07:20.110 doctor. This hits close to home for anyone who 00:07:20.110 --> 00:07:23.550 codes. Oh, yeah. Debugging is the worst sometimes. 00:07:23.810 --> 00:07:27.029 Tell me about it. I still wrestle with prompt 00:07:27.029 --> 00:07:29.689 drift myself when I'm debugging my own small 00:07:29.689 --> 00:07:32.790 projects, trying one thing, then another, getting 00:07:32.790 --> 00:07:36.170 lost. The idea of an AI cutting through that. 00:07:36.740 --> 00:07:40.800 Huge. And KimiDev from Moonshot AI seems to do 00:07:40.800 --> 00:07:43.879 just that. It's free, and it's outperforming 00:07:43.879 --> 00:07:47.079 paid tools on standard tests. How good is it? 00:07:47.139 --> 00:07:48.399 Like, what are the numbers? Well, there's this 00:07:48.399 --> 00:07:51.439 benchmark called SWBench. It uses real -world 00:07:51.439 --> 00:07:54.420 broken code from GitHub projects. The previous 00:07:54.420 --> 00:07:57.699 best free models scored around 40%. Yeah. Top 00:07:57.699 --> 00:08:00.800 paid ones were maybe 50%, 55%. Okay. And KimiDev? 00:08:00.920 --> 00:08:03.339 Scored 60 .4%. That's not just a little better. 00:08:03.420 --> 00:08:05.980 That's a massive jump. Like a 50 % improvement 00:08:05.980 --> 00:08:08.480 over the prior best free model. Wow. Okay. How 00:08:08.480 --> 00:08:10.279 does it actually work? What's the magic? They 00:08:10.279 --> 00:08:12.259 call it a two -brain approach. Yeah. Kind of 00:08:12.259 --> 00:08:14.680 cool. One brain is the bug fixer. It analyzes 00:08:14.680 --> 00:08:17.060 the code, pinpoints the problem with real proficient. 00:08:17.160 --> 00:08:18.779 It finds the needle in the haystack. Pretty much. 00:08:18.899 --> 00:08:21.399 Then the second brain is a test writer. It automatically 00:08:21.399 --> 00:08:23.040 writes new tests to make sure the fix actually 00:08:23.040 --> 00:08:24.920 worked and didn't break something else. So it 00:08:24.920 --> 00:08:27.199 fixes the bug and writes the unit test to prove 00:08:27.199 --> 00:08:29.939 it. Exactly. Built -in quality check. That's 00:08:29.939 --> 00:08:32.440 really clever. How do developers actually use 00:08:32.440 --> 00:08:34.419 this? Can you just point it at your messy code? 00:08:34.720 --> 00:08:36.919 You can often run it locally, which is a big 00:08:36.919 --> 00:08:38.860 plus for privacy. You'd find it on places like 00:08:38.860 --> 00:08:42.159 GitHub or Hugging Face. Right, keep your proprietary 00:08:42.159 --> 00:08:44.799 code on your machine. Yeah, you give it the bug 00:08:44.799 --> 00:08:47.840 report, the code base, it'll propose a fix, and 00:08:47.840 --> 00:08:50.279 the tests. Definitely start with simpler bugs 00:08:50.279 --> 00:08:52.559 first. And you still need to review it, obviously. 00:08:52.899 --> 00:08:54.539 Well, absolutely. It's an incredibly powerful 00:08:54.539 --> 00:08:57.440 assistant, but human oversight is still crucial. 00:08:57.870 --> 00:09:00.629 You review the fix, review the tests, then merge 00:09:00.629 --> 00:09:03.029 it. So KimiDev isn't replacing developers tomorrow. 00:09:03.289 --> 00:09:05.649 No, no, it's augmenting them, making them way 00:09:05.649 --> 00:09:07.409 more efficient at one of the most time -consuming 00:09:07.409 --> 00:09:10.129 parts of the job. Sponsor. Okay, let's pull back 00:09:10.129 --> 00:09:12.049 a bit. We've looked at Minimax for websites, 00:09:12.509 --> 00:09:16.210 CDance for video, KimiDev for code. What's the 00:09:16.210 --> 00:09:18.769 big picture here? What does this all really mean? 00:09:19.090 --> 00:09:21.409 I think it signals a really fundamental shift, 00:09:21.649 --> 00:09:25.269 maybe even the end of an era for expensive, closed 00:09:25.269 --> 00:09:28.679 -off AI tools. These capabilities are becoming 00:09:28.679 --> 00:09:31.379 accessible to everyone, often for free. Which 00:09:31.379 --> 00:09:34.200 is huge for me. Individuals, small businesses. 00:09:34.750 --> 00:09:36.990 startups, anyone who previously couldn't afford 00:09:36.990 --> 00:09:41.149 access to this level of AI power. Suddenly, you 00:09:41.149 --> 00:09:43.750 have these incredible tools at your fingertips. 00:09:44.049 --> 00:09:46.610 The playing field really is leveling out. Completely. 00:09:46.830 --> 00:09:48.789 Yeah. It's less about your budget now and more 00:09:48.789 --> 00:09:51.289 about your ideas, your creativity, your speed 00:09:51.289 --> 00:09:54.350 in using these tools. And what about the companies 00:09:54.350 --> 00:09:56.029 that have been charging those high subscription 00:09:56.029 --> 00:09:58.149 fees? Well, the pressure is definitely on them 00:09:58.149 --> 00:10:01.210 now. It's much harder to justify charging a lot 00:10:01.210 --> 00:10:03.389 when there are free alternatives that are just 00:10:03.389 --> 00:10:05.710 as good or sometimes even better. So they'll 00:10:05.710 --> 00:10:08.389 need to find new ways to add value. Exactly. 00:10:08.809 --> 00:10:11.129 Maybe through better user experience, specialized 00:10:11.129 --> 00:10:13.990 industry models, integrations, enterprise level 00:10:13.990 --> 00:10:17.429 support, things beyond just the raw AI capability 00:10:17.429 --> 00:10:20.110 itself. Looking across all three tools, there's 00:10:20.110 --> 00:10:22.289 a common thread that strikes me. Speed. Absolutely. 00:10:22.850 --> 00:10:25.330 Websites in minutes, videos in seconds, code 00:10:25.330 --> 00:10:27.470 fixed almost instantly. That changes the whole 00:10:27.470 --> 00:10:29.710 pace of creation and innovation, doesn't it? 00:10:29.929 --> 00:10:32.370 It creates a massive competitive advantage. If 00:10:32.370 --> 00:10:34.549 you can iterate, experiment and deploy that much 00:10:34.549 --> 00:10:38.269 faster, you can learn and adapt far quicker than 00:10:38.269 --> 00:10:40.789 competitors stuck on slower cycles. So the takeaway 00:10:40.789 --> 00:10:43.950 seems to be this AI revolution isn't some future 00:10:43.950 --> 00:10:45.909 thing we're waiting for. Not at all. It's here 00:10:45.909 --> 00:10:49.090 now. And it's being driven by a different philosophy, 00:10:49.309 --> 00:10:52.190 this democratization idea. And it's available 00:10:52.190 --> 00:10:55.830 to you today for free. Take these insights. Think 00:10:55.830 --> 00:10:58.309 about how they could change how you work, what 00:10:58.309 --> 00:11:00.600 you create. Yeah, it's not just about doing the 00:11:00.600 --> 00:11:03.159 same things faster. It's about what entirely 00:11:03.159 --> 00:11:05.639 new things become possible now. The only real 00:11:05.639 --> 00:11:07.080 question left is what are you going to build 00:11:07.080 --> 00:11:09.779 with it? Thanks for joining us on this deep dive. 00:11:09.980 --> 00:11:12.279 We hope this leaves you feeling truly well informed.