WEBVTT 00:00:00.000 --> 00:00:02.680 You know, if you want to find the real practical 00:00:02.680 --> 00:00:07.000 frontier of AI today, you don't look at chatbots 00:00:07.000 --> 00:00:09.480 writing poetry. You look at something really 00:00:09.480 --> 00:00:12.419 mundane. You look at washing a greasy frying 00:00:12.419 --> 00:00:15.609 pan. Right. Or... And this is the ultimate test. 00:00:15.810 --> 00:00:18.910 You ask a robot to open one of those super thin 00:00:18.910 --> 00:00:22.550 plastic dark poop bags. Ah, the worst. It's the 00:00:22.550 --> 00:00:25.350 seemingly simple physical thing that's still 00:00:25.350 --> 00:00:28.230 the ultimate humbling challenge. It just shows 00:00:28.230 --> 00:00:30.870 you the gap between the digital world and the 00:00:30.870 --> 00:00:33.369 physical one. And welcome to the Deep Dive. We 00:00:33.369 --> 00:00:35.210 take your sources, your research, and we pull 00:00:35.210 --> 00:00:37.630 out what really matters. Today, it's all about 00:00:37.630 --> 00:00:40.420 applied intelligence. ai in the real world right 00:00:40.420 --> 00:00:42.899 now okay so we're going to unpack three key areas 00:00:42.899 --> 00:00:45.240 first we'll look at the latest robot olympics 00:00:45.240 --> 00:00:48.359 and uh a really big training data breakthrough 00:00:48.359 --> 00:00:50.479 okay then we'll get into some new utility tools 00:00:50.479 --> 00:00:52.520 that are really speeding up daily workflows for 00:00:52.520 --> 00:00:56.299 creators and finally a new open source ai agent 00:00:56.299 --> 00:00:59.920 that's well it's seriously challenging the big 00:00:59.920 --> 00:01:02.259 players okay let's unpack this so we've all seen 00:01:02.259 --> 00:01:05.400 those videos glossy demos exactly robots dancing 00:01:05.400 --> 00:01:08.489 moving boxes perfectly it looks incredible In 00:01:08.489 --> 00:01:10.670 a lab. But the real world is messy. That's it. 00:01:11.010 --> 00:01:14.909 Friction. Gravity. A self -closing door. That's 00:01:14.909 --> 00:01:17.090 where those demos just fall apart. And that's 00:01:17.090 --> 00:01:19.829 exactly why this new Robot Olympics from PI is 00:01:19.829 --> 00:01:22.989 so important. It forces them into these ridiculously 00:01:22.989 --> 00:01:26.329 hard real -world chores. And we're not talking 00:01:26.329 --> 00:01:28.469 about simple pick -and -place tasks. Think about 00:01:28.469 --> 00:01:32.430 the complexity here. The robot, .6, had to get 00:01:32.430 --> 00:01:34.730 through a door with a lever handle. Which requires 00:01:34.730 --> 00:01:37.230 that perfect sequence of pushing down and pulling. 00:01:37.349 --> 00:01:39.329 Right. And it had to insert a key into a lock. 00:01:39.430 --> 00:01:42.189 And then the big one, wash a greasy pan with 00:01:42.189 --> 00:01:45.120 soap. You've got fluid dynamics, tactile feedback. 00:01:45.640 --> 00:01:48.060 It's all incredibly hard for a machine. And the 00:01:48.060 --> 00:01:51.019 results really show where the rubber meets the 00:01:51.019 --> 00:01:53.099 road, or I guess where the gripper meets the 00:01:53.099 --> 00:01:55.519 fabric. It actually managed to turn a sock right 00:01:55.519 --> 00:01:57.980 side out, but then it completely failed trying 00:01:57.980 --> 00:02:00.319 to do a shirt sleeve. The geometry was just too 00:02:00.319 --> 00:02:02.879 complex. Exactly. And the sources said the single 00:02:02.879 --> 00:02:05.400 hardest task was opening that thin plastic bag. 00:02:05.560 --> 00:02:09.189 The dog poop bag. That one. The material is so 00:02:09.189 --> 00:02:12.050 thin and reflective, it literally blinded the 00:02:12.050 --> 00:02:14.870 camera sensors mid -move. Wow. It's a classic 00:02:14.870 --> 00:02:17.349 perception problem mixed with a dexterity problem. 00:02:17.490 --> 00:02:19.210 Yeah. It's a fascinating failure, really. But 00:02:19.210 --> 00:02:21.810 what's really fascinating here is the shift in 00:02:21.810 --> 00:02:24.409 how these robots are being trained. Yes. It used 00:02:24.409 --> 00:02:28.509 to be thousands of hours of expensive robot -specific 00:02:28.509 --> 00:02:31.090 demos. Yeah. You know, watching other robots 00:02:31.090 --> 00:02:33.689 do the tasks. We're moving past that. The breakthrough 00:02:33.689 --> 00:02:36.569 is that the fine -tuning was done on egocentric 00:02:36.569 --> 00:02:39.800 human video. OK, so define egocentric video for 00:02:39.800 --> 00:02:42.340 us. It's just first person footage like a GoPro 00:02:42.340 --> 00:02:44.800 on your head while you're doing the task, making 00:02:44.800 --> 00:02:46.879 coffee, folding laundry, whatever. So the robot 00:02:46.879 --> 00:02:49.520 is learning by watching us. Exactly. It's a cheap, 00:02:49.599 --> 00:02:52.439 scalable data source. And the robot isn't just 00:02:52.439 --> 00:02:55.360 learning how a robot moves. It's learning the 00:02:55.360 --> 00:02:58.000 intent behind human movements. So it generalizes 00:02:58.000 --> 00:03:00.840 way faster. This whole process is a little like 00:03:00.840 --> 00:03:02.860 learning to drive just by watching dash cam footage. 00:03:03.099 --> 00:03:05.020 And honestly, I still wrestle with prompt drift 00:03:05.020 --> 00:03:07.740 myself when I try new AI systems. So seeing a 00:03:07.740 --> 00:03:10.580 robot learn this complex physical skill is just 00:03:10.580 --> 00:03:13.639 it's incredible. So if these human videos are 00:03:13.639 --> 00:03:16.300 the key, what's the bottleneck? What's preventing 00:03:16.300 --> 00:03:18.780 these robots from being in every home folding 00:03:18.780 --> 00:03:21.580 our laundry right now? Repeatability across many 00:03:21.580 --> 00:03:25.759 different unpredictable real homes is still the 00:03:25.759 --> 00:03:29.210 primary missing piece. Right. So, yeah. The physical 00:03:29.210 --> 00:03:31.430 world is still a huge challenge. But the digital 00:03:31.430 --> 00:03:34.969 world, that's where AI is making these immediate 00:03:34.969 --> 00:03:37.810 massive efficiency gains. Right. Away from the 00:03:37.810 --> 00:03:39.770 hardware problem. We always focus on the huge 00:03:39.770 --> 00:03:42.710 models. But the daily utility stuff, that's where 00:03:42.710 --> 00:03:46.990 AI is saving people hours of tedious work. And 00:03:46.990 --> 00:03:49.210 the source has pointed to a few of these. Let's 00:03:49.210 --> 00:03:51.520 talk creative efficiency. They compare models 00:03:51.520 --> 00:03:54.439 like GPT Image 1 .5 with another one, Nano Banana 00:03:54.439 --> 00:03:57.460 Pro, for people making ads or thumbnails. And 00:03:57.460 --> 00:04:00.039 the huge win there is being able to edit photos 00:04:00.039 --> 00:04:02.919 directly inside something like ChatGPT. Okay, 00:04:02.960 --> 00:04:04.919 so you don't need a separate app. No more exporting, 00:04:05.020 --> 00:04:07.599 opening Photoshop, making a change, saving, re 00:04:07.599 --> 00:04:09.639 -uploading. You just type what you want. That 00:04:09.639 --> 00:04:11.979 avoidance of app switching, that's probably the 00:04:11.979 --> 00:04:13.599 biggest convenience win right there for anyone 00:04:13.599 --> 00:04:15.900 creating a lot of content. It really is. And 00:04:15.900 --> 00:04:18.319 beyond visuals, there are these document and 00:04:18.319 --> 00:04:21.569 video tools. Take one's called Typeless. Okay. 00:04:21.910 --> 00:04:24.649 It's designed to take your spoken words, you 00:04:24.649 --> 00:04:27.009 know, your messy brainstorm in a meeting, and 00:04:27.009 --> 00:04:29.629 instantly turn it into a polished document like 00:04:29.629 --> 00:04:32.129 it was perfectly typed. So it turns a chaotic 00:04:32.129 --> 00:04:35.569 brain dump into a clean memo instantly. Yeah. 00:04:35.649 --> 00:04:37.470 That's a massive time saver. And then there's 00:04:37.470 --> 00:04:40.449 EgoX. This one is really wild. It can take any 00:04:40.449 --> 00:04:42.930 third -person video, like a movie clip. Like 00:04:42.930 --> 00:04:44.870 from the Duck Knight or something. Exactly. And 00:04:44.870 --> 00:04:47.230 it converts it into a first -person. Point of 00:04:47.230 --> 00:04:49.829 view version. Wow. I can see that being huge 00:04:49.829 --> 00:04:52.290 for training or immersive storytelling. For sure. 00:04:52.389 --> 00:04:55.149 And we can't forget the visual Swiss Army knife, 00:04:55.389 --> 00:04:58.769 nano banana playground. Right. It does text to 00:04:58.769 --> 00:05:01.430 image, image editing. And this is key. It handles 00:05:01.430 --> 00:05:04.050 multiple aspect ratios really well. No more weird 00:05:04.050 --> 00:05:06.949 forced crops. The sources also mentioned a system 00:05:06.949 --> 00:05:09.589 for recycling old content, right? Like turning 00:05:09.589 --> 00:05:12.110 forgotten blog posts into new traffic. Yeah. 00:05:12.149 --> 00:05:14.550 Another example of AI just streamlining workflows. 00:05:15.199 --> 00:05:16.879 Making the most of what you already have without 00:05:16.879 --> 00:05:19.579 hiring more people. So out of all of these, which 00:05:19.579 --> 00:05:22.360 tool offers the biggest immediate time saver 00:05:22.360 --> 00:05:25.339 for just a general professional user? Using in 00:05:25.339 --> 00:05:28.079 -chat photo editing, avoiding app switching is 00:05:28.079 --> 00:05:30.399 truly the biggest convenience win. Okay, so the 00:05:30.399 --> 00:05:32.120 digital world is getting more efficient. But 00:05:32.120 --> 00:05:35.660 what about the really big, complex tasks? This 00:05:35.660 --> 00:05:37.620 is where it gets really interesting. There's 00:05:37.620 --> 00:05:41.660 this new open source model, Minimax M2 .1. Right. 00:05:41.800 --> 00:05:43.800 And developers are already calling it a clod 00:05:43.800 --> 00:05:46.689 killer. That's some high praise. But before we 00:05:46.689 --> 00:05:48.670 get into why, let's just quickly define what 00:05:48.670 --> 00:05:52.310 an AI agent is in simple terms. Right. Good idea. 00:05:52.389 --> 00:05:56.129 An AI agent is a program that plans, acts. and 00:05:56.129 --> 00:05:58.709 then corrects its own mistakes to finish complex 00:05:58.709 --> 00:06:01.970 tasks. Okay. Plans, acts, corrects. Got it. And 00:06:01.970 --> 00:06:04.550 M2 .1 was built from the ground up to do just 00:06:04.550 --> 00:06:07.050 that. The performance numbers really do back 00:06:07.050 --> 00:06:08.709 up the hype. Let's talk about those numbers. 00:06:08.810 --> 00:06:13.490 It scored 72 .5 % on something called SWE Multilingual. 00:06:13.810 --> 00:06:16.509 What does that percentage actually mean? So SWE 00:06:16.509 --> 00:06:19.350 Multilingual tests the model's ability to fix 00:06:19.350 --> 00:06:22.550 bugs in open source code autonomously. So that 00:06:22.550 --> 00:06:26.949 72 .5 % means it can find diagnose, and fix a 00:06:26.949 --> 00:06:29.610 huge number of real -world software bugs with 00:06:29.610 --> 00:06:32.629 zero human help. That actually puts it ahead 00:06:32.629 --> 00:06:34.610 of some big proprietary models like CloudSonic 00:06:34.610 --> 00:06:38.430 4 .5. And then there was the 88 .6 % on VibeBench. 00:06:38.470 --> 00:06:40.649 Right. And VibeBench measures how well the agent 00:06:40.649 --> 00:06:43.129 can interact with user interfaces. Like navigating 00:06:43.129 --> 00:06:46.029 a website or a phone app? Exactly. It's like 00:06:46.029 --> 00:06:48.410 having an AI that can beta test your software 00:06:48.410 --> 00:06:51.060 or make complex design changes on its own. Whoa. 00:06:51.540 --> 00:06:54.660 Imagine scaling this planning accuracy to handle 00:06:54.660 --> 00:06:57.459 a billion complex database queries autonomously. 00:06:57.660 --> 00:07:00.939 That level of reliability, that changes everything 00:07:00.939 --> 00:07:03.319 for big companies. It absolutely does. But with 00:07:03.319 --> 00:07:05.439 that kind of self -correcting ability, how much 00:07:05.439 --> 00:07:07.240 power does it take to run? Is this something 00:07:07.240 --> 00:07:09.759 that's even sustainable outside of a huge tech 00:07:09.759 --> 00:07:11.680 firm? That's the right question. And the key 00:07:11.680 --> 00:07:14.459 is it's agent -native design. Most older models, 00:07:14.600 --> 00:07:16.620 they just follow a list. Step one, step two. 00:07:16.720 --> 00:07:19.079 If anything changes, they fail. They just stop. 00:07:19.379 --> 00:07:21.980 M2 .1 is different. After every action it takes, 00:07:22.120 --> 00:07:24.339 it looks at the result and if it needs to, it 00:07:24.339 --> 00:07:26.500 revises its plan. It corrects its own mistakes 00:07:26.500 --> 00:07:29.899 on the fly. So it basically debugs itself. That's 00:07:29.899 --> 00:07:31.879 a great way to put it. And because it supports 00:07:31.879 --> 00:07:34.899 languages, developers actually use Rust, Java, 00:07:35.199 --> 00:07:39.079 Go. It can build real full stack apps. So it's 00:07:39.079 --> 00:07:41.220 not generating what they call AI spaghetti code. 00:07:41.439 --> 00:07:45.120 Exactly. It's clean, structured code. The sources 00:07:45.120 --> 00:07:47.339 say it's holding up incredibly well even against 00:07:47.339 --> 00:07:50.379 older top -tier models like Cloud II. So if M2 00:07:50.379 --> 00:07:53.540 .1 is open source and performs this well, what's 00:07:53.540 --> 00:07:55.879 the biggest barrier stopping everyone from just 00:07:55.879 --> 00:07:58.980 adopting it over the proprietary models? Accessibility 00:07:58.980 --> 00:08:01.740 and scaling infrastructure remain the main challenges 00:08:01.740 --> 00:08:05.439 for any open source alternative. Okay. So if 00:08:05.439 --> 00:08:07.339 we pull back and look at everything we've covered, 00:08:07.560 --> 00:08:09.839 there's a pretty clear narrative here connecting 00:08:09.839 --> 00:08:13.220 the physical and the digital. We went from robots 00:08:13.220 --> 00:08:16.879 learning to wash pans by watching human videos. 00:08:17.120 --> 00:08:19.959 That egocentric learning. Right. All the way 00:08:19.959 --> 00:08:22.459 to AI agents that can debug their own complex 00:08:22.459 --> 00:08:24.279 digital plans. And the theme that connects it 00:08:24.279 --> 00:08:26.800 all is leveraging human knowledge, our videos, 00:08:26.939 --> 00:08:29.600 our code, our old content, all to build systems 00:08:29.600 --> 00:08:32.100 that are more adaptive, more reliable. Yeah, 00:08:32.120 --> 00:08:34.299 the frontier isn't just raw intelligence anymore. 00:08:34.379 --> 00:08:36.740 It's autonomous adaptation. That's really well 00:08:36.740 --> 00:08:38.940 put. And if we connect this to the bigger picture. 00:08:39.720 --> 00:08:41.860 We've talked about these new ways to organize 00:08:41.860 --> 00:08:44.539 and create, all driven by agents that can revise 00:08:44.539 --> 00:08:48.000 their plans instantly. And that raises a pretty 00:08:48.000 --> 00:08:50.340 important question for us. I think if AI agents 00:08:50.340 --> 00:08:53.879 get this good at revising their plans mid -task, 00:08:54.039 --> 00:08:57.700 how much of our own mental workflow, you know, 00:08:57.740 --> 00:09:00.779 our human reliance on static linear plans. Yeah. 00:09:01.289 --> 00:09:03.350 How much will that need to adapt to keep pace? 00:09:03.490 --> 00:09:06.789 A good question. What if in this new age, failure 00:09:06.789 --> 00:09:09.649 is just the essential first step in a successful 00:09:09.649 --> 00:09:11.649 self -corrected plan? Something to mull over 00:09:11.649 --> 00:09:13.730 as you optimize your own workflows. Thank you 00:09:13.730 --> 00:09:15.429 for sharing your sources for this deep dive. 00:09:15.490 --> 00:09:17.389 We appreciate you joining us. Until next time, 00:09:17.409 --> 00:09:17.909 stay curious.