WEBVTT 00:00:00.000 --> 00:00:01.919 OK, so let's get into this. We're looking at 00:00:01.919 --> 00:00:06.820 three really different stories about where AI 00:00:06.820 --> 00:00:10.400 is going right now. We have this new household 00:00:10.400 --> 00:00:13.480 robot that's designed to be, you know, cute on 00:00:13.480 --> 00:00:15.839 purpose. Right. Then you have these AI models 00:00:15.839 --> 00:00:18.039 just battling it out in these secret benchmarks. 00:00:18.320 --> 00:00:21.559 And on the other end, a group just dropped the 00:00:21.559 --> 00:00:24.620 entire blueprint for a major language model. 00:00:25.160 --> 00:00:28.739 True open source. And if you need proof of just 00:00:28.739 --> 00:00:31.660 how intense all of this is getting, get this. 00:00:32.219 --> 00:00:34.500 Meta is now asking the government for permission 00:00:34.500 --> 00:00:37.320 to trade electricity. Yeah, just to feed their 00:00:37.320 --> 00:00:40.039 AI, the stakes are getting incredibly high. Welcome 00:00:40.039 --> 00:00:43.000 to the deep dive. Our mission today is to unpack 00:00:43.000 --> 00:00:45.380 these sources for you and really figure out what 00:00:45.380 --> 00:00:47.539 matters most. We've got three main areas we're 00:00:47.539 --> 00:00:49.600 going to hit. First up, that dish doing robot 00:00:49.600 --> 00:00:52.659 and the surprisingly simple cheat code they found 00:00:52.659 --> 00:00:54.560 for training it. Then we'll get into the latest 00:00:54.560 --> 00:00:57.119 industry headlines. Who's winning the AI race 00:00:57.119 --> 00:00:59.259 and what does all that energy consumption really 00:00:59.259 --> 00:01:01.859 mean? And finally, we'll look at a really radical 00:01:01.859 --> 00:01:04.640 new definition of open source from the AI2 team 00:01:04.640 --> 00:01:06.920 that could change everything. Let's start in 00:01:06.920 --> 00:01:09.079 the home, in the kitchen, actually. Yeah, the 00:01:09.079 --> 00:01:11.629 cute robot. Sunday Robotics. just came out of 00:01:11.629 --> 00:01:14.310 stealth with their memo robot. And we've all 00:01:14.310 --> 00:01:16.549 been hearing that promise of a Rosie the robot 00:01:16.549 --> 00:01:19.569 for, what, decades now? Right. A real helper 00:01:19.569 --> 00:01:22.109 around the house. It's finally starting to feel 00:01:22.109 --> 00:01:24.090 a little more tangible, and I think it's because 00:01:24.090 --> 00:01:26.750 they solved the data problem in a really different 00:01:26.750 --> 00:01:29.930 way. The huge roadblock for home robots has always 00:01:29.930 --> 00:01:32.739 been the data. Right. Getting good data on how 00:01:32.739 --> 00:01:35.319 humans actually do things. Exactly. How do you 00:01:35.319 --> 00:01:39.060 open that specific drawer or pick up a mug without, 00:01:39.140 --> 00:01:42.459 you know, crushing it? It's all about dexterity. 00:01:42.500 --> 00:01:45.079 And traditionally that meant using a super expensive 00:01:45.079 --> 00:01:48.700 telop rig. Yeah. Where a skilled operator wears 00:01:48.700 --> 00:01:51.599 all this motion capture gear to control the robot 00:01:51.599 --> 00:01:54.659 remotely. It's precise, but it's also incredibly 00:01:54.659 --> 00:01:57.379 slow and costs a fortune per hour. So they just 00:01:57.379 --> 00:01:59.760 threw that whole idea out. Completely. They swapped 00:01:59.760 --> 00:02:03.819 it for the $200 skill capture gloves. $200. Yeah. 00:02:04.599 --> 00:02:06.840 Real people just put them on in their own homes 00:02:06.840 --> 00:02:10.110 and do chores. Live their lives. Pretty much. 00:02:10.210 --> 00:02:13.370 They're clearing tables, folding laundry, loading 00:02:13.370 --> 00:02:16.949 dishwashers, even pulling espresso shots. It 00:02:16.949 --> 00:02:20.009 captures pure human motor skills, but cheaply 00:02:20.009 --> 00:02:22.509 and at a massive scale. That's brilliant. You're 00:02:22.509 --> 00:02:24.710 just crowdsourcing the training data from the 00:02:24.710 --> 00:02:26.830 8 billion people who already know how to do the 00:02:26.830 --> 00:02:29.229 job. And the sources say this method generated 00:02:29.229 --> 00:02:32.729 10 million episodes of data. 10 million. That's 00:02:32.729 --> 00:02:35.990 just, that is a massive shortcut. It is. It just 00:02:35.990 --> 00:02:39.169 rapidly accelerates how you teach a robot complex 00:02:39.169 --> 00:02:42.009 home dexterity. So how good is it? Well, the 00:02:42.009 --> 00:02:44.050 results are pretty impressive. Memo can do a 00:02:44.050 --> 00:02:46.889 full table -to -dishwasher run. It's a sequence 00:02:46.889 --> 00:02:50.050 of 68 different dexterous moves. And the big 00:02:50.050 --> 00:02:52.490 question. The big question. Across more than 00:02:52.490 --> 00:02:55.909 20 live demos, they recorded zero broken wine 00:02:55.909 --> 00:02:58.729 glasses. Wow. Okay, that's the real test. That's 00:02:58.729 --> 00:03:00.469 the real test. That kind of delicate handling 00:03:00.469 --> 00:03:02.979 tells you the data quality is there. But wait, 00:03:03.080 --> 00:03:05.060 if the gloves are that cheap, isn't filtering 00:03:05.060 --> 00:03:07.990 all that messy crowdsourced data? Just an absolute 00:03:07.990 --> 00:03:09.889 nightmare. I mean, couldn't cheap data just lead 00:03:09.889 --> 00:03:12.449 to cheap Dex Purdy? That's the core challenge 00:03:12.449 --> 00:03:14.490 for sure. But look at the team. It's a bunch 00:03:14.490 --> 00:03:17.430 of Stanford PhDs and ex -Tesla FSD engineers. 00:03:17.770 --> 00:03:20.090 They lived through that whole millions of miles 00:03:20.090 --> 00:03:22.629 data collection problem. They know how to filter 00:03:22.629 --> 00:03:25.509 noise. They explicitly said they chose to leverage 00:03:25.509 --> 00:03:28.949 8 billion humans instead of millions of miles. 00:03:29.250 --> 00:03:31.370 OK, that background brings up some skepticism 00:03:31.370 --> 00:03:34.370 for me, though. Yeah. The FSD team was, let's 00:03:34.370 --> 00:03:37.219 be honest. Famous for overpromising on timelines? 00:03:37.599 --> 00:03:41.199 Mm -hmm. Does that affect how we should see their 00:03:41.199 --> 00:03:44.979 late 2026 ship date? It definitely raises an 00:03:44.979 --> 00:03:47.939 eyebrow, but I think the domain is more constrained 00:03:47.939 --> 00:03:50.500 here than full self -driving. And their other 00:03:50.500 --> 00:03:53.300 big choice is maybe just as important. The design. 00:03:53.500 --> 00:03:56.819 The design. The robot is soft. It's round. It 00:03:56.819 --> 00:03:59.379 wears a little hat. So cuteness is now a technical 00:03:59.379 --> 00:04:02.280 feature. Exactly. Because it builds trust. If 00:04:02.280 --> 00:04:04.939 this little non -threatening robot makes a mistake 00:04:04.939 --> 00:04:07.300 while it's learning, you're more likely to forgive 00:04:07.300 --> 00:04:09.479 it and let it keep trying. Instead of unplugging 00:04:09.479 --> 00:04:11.240 it and throwing it in the closet. Right. User 00:04:11.240 --> 00:04:13.219 empathy actually becomes part of the training 00:04:13.219 --> 00:04:16.879 loop. So if we boil this down, leveraging crowdsourced 00:04:16.879 --> 00:04:19.779 human experience. rapidly accelerates complex 00:04:19.779 --> 00:04:22.000 home dexterity. That's it. It just fundamentally 00:04:22.000 --> 00:04:24.439 changes the robotics timeline. Okay, let's shift 00:04:24.439 --> 00:04:28.279 from the kitchen counter to the wild world of 00:04:28.279 --> 00:04:30.300 AI rankings. Yeah, it's a high -stakes race. 00:04:30.420 --> 00:04:31.980 And these benchmarks, they used to be pretty 00:04:31.980 --> 00:04:34.800 stable. Not anymore. There's a new one, the Humane 00:04:34.800 --> 00:04:37.019 Benchmark, that just sent out some shockwaves. 00:04:37.199 --> 00:04:39.600 It's designed to test really modern stuff like 00:04:39.600 --> 00:04:42.100 advanced reasoning, handling images and text, 00:04:42.199 --> 00:04:46.060 and safety. And the big surprise? GPT came in 00:04:46.060 --> 00:04:48.759 at number eight. Number eight. I mean, that is 00:04:48.759 --> 00:04:51.720 shockingly low for a model that basically felt 00:04:51.720 --> 00:04:54.560 like the default for so long. If GPT is at eight, 00:04:54.639 --> 00:04:57.730 who's at the top? Top spot went to Gemini 2 .5 00:04:57.730 --> 00:05:00.329 Pro. After that's Tight Race. You've got Deep 00:05:00.329 --> 00:05:03.709 Seek, Mistroll, and then get this Grok 4 and 00:05:03.709 --> 00:05:06.970 Grok 3 took spots 4 and 5. So the performance 00:05:06.970 --> 00:05:09.990 gap is just closing incredibly fast. Faster than 00:05:09.990 --> 00:05:12.110 ever. And these technical leaps are leading to 00:05:12.110 --> 00:05:14.250 some pretty wild spectacle too. You mean the 00:05:14.250 --> 00:05:16.810 Gemini 3 Pro ad on the Las Vegas sphere? Yeah, 00:05:16.829 --> 00:05:19.069 that thing. People were genuinely arguing online 00:05:19.069 --> 00:05:21.370 about whether it was real footage or AI generated. 00:05:21.709 --> 00:05:23.509 Because the quality was just too good to tell. 00:05:24.379 --> 00:05:26.779 That line is getting very, very blurry. And it's 00:05:26.779 --> 00:05:29.769 not just visuals, it's speed. We just saw the 00:05:29.769 --> 00:05:32.509 first fully driverless race cars in Abu Dhabi. 00:05:32.529 --> 00:05:35.689 I saw that. Hitting 155 miles per hour, racing 00:05:35.689 --> 00:05:38.610 wheel to wheel. That's all algorithmic decision 00:05:38.610 --> 00:05:41.350 making at insane speeds. And then on the creative 00:05:41.350 --> 00:05:43.529 side, you have things like Google's Nano Banana 00:05:43.529 --> 00:05:46.649 Pro. Which is just doing crazy things with images. 00:05:47.410 --> 00:05:50.870 Fixing specific text inside a photo or blending, 00:05:50.970 --> 00:05:53.250 what, 14 different images together seamlessly. 00:05:53.610 --> 00:05:57.000 It's just... It's hard to even keep up. It really 00:05:57.000 --> 00:05:59.560 is. I have to say, even with all these powerful 00:05:59.560 --> 00:06:02.420 new tools, I still wrestle with prompt drift 00:06:02.420 --> 00:06:05.050 myself. Oh, really? Yeah. I'll find a model, 00:06:05.089 --> 00:06:07.350 gives me perfect output for like the first week 00:06:07.350 --> 00:06:09.629 it's out. Then a month later, the consistency 00:06:09.629 --> 00:06:12.410 just degrades and I'm back to reengineering my 00:06:12.410 --> 00:06:14.990 prompts all over again. Getting reliable output 00:06:14.990 --> 00:06:18.110 is still a huge challenge. That inconsistency 00:06:18.110 --> 00:06:20.550 makes the whole infrastructure problem even worse, 00:06:20.649 --> 00:06:22.649 doesn't it? We're talking about massive amounts 00:06:22.649 --> 00:06:24.930 of energy to run these things. Which brings us 00:06:24.930 --> 00:06:27.490 back to meta. Bloomberg is reporting that because 00:06:27.490 --> 00:06:29.449 the grid is struggling to keep up with their 00:06:29.449 --> 00:06:32.410 AI hunger, they've asked the feds for permission 00:06:32.410 --> 00:06:34.829 to trade electricity. They're trying to become 00:06:34.829 --> 00:06:37.170 an energy company. Essentially, yeah. A utility 00:06:37.170 --> 00:06:39.410 company just to service their own data centers. 00:06:39.730 --> 00:06:42.170 They're so power hungry, they're trying to shape 00:06:42.170 --> 00:06:44.769 national energy strategy just to keep the lights 00:06:44.769 --> 00:06:47.209 on for their models. Wow. So if we connect the 00:06:47.209 --> 00:06:50.569 dots here, the hidden cost of scaling these huge 00:06:50.569 --> 00:06:54.129 models is that AI is now powerful enough to dictate 00:06:54.129 --> 00:06:57.529 national energy infrastructure strategy. It changes 00:06:57.529 --> 00:07:00.449 the entire planning horizon for the energy sector. 00:07:00.670 --> 00:07:03.480 We hear the term open source. A lot in AI. And 00:07:03.480 --> 00:07:05.800 for a while now, that's usually just meant one 00:07:05.800 --> 00:07:08.660 thing, sharing the model weights. Right. And 00:07:08.660 --> 00:07:10.500 we should probably define that. The weights are 00:07:10.500 --> 00:07:12.860 just the final set of numbers that make the model 00:07:12.860 --> 00:07:15.360 work. It's the finished product, but not the 00:07:15.360 --> 00:07:18.259 recipe. Exactly. So if you're a researcher trying 00:07:18.259 --> 00:07:21.620 to figure out why a model is biased, just having 00:07:21.620 --> 00:07:23.959 the weights is not enough. It's a black box. 00:07:24.199 --> 00:07:26.319 You can use it, but you can't truly understand 00:07:26.319 --> 00:07:29.459 it. Which is why this new release from the AI2 00:07:29.459 --> 00:07:33.910 team, OMO3, is such a big deal. They've... They've 00:07:33.910 --> 00:07:35.930 radically redefined the term. They're providing 00:07:35.930 --> 00:07:38.230 everything, the full training data, every line 00:07:38.230 --> 00:07:40.449 of code, every training checkpoint. The checkpoints 00:07:40.449 --> 00:07:43.430 are like saved games, right? The model at different 00:07:43.430 --> 00:07:45.949 stages of learning. Exactly. And every decision 00:07:45.949 --> 00:07:48.269 they made along the way. It's a full -on anti 00:07:48.269 --> 00:07:51.170 -black box effort. It's all about traceability. 00:07:51.589 --> 00:07:53.670 That is a level of transparency we just haven't 00:07:53.670 --> 00:07:55.910 seen. It's like, imagine if Tesla didn't just 00:07:55.910 --> 00:07:58.189 give you the car blueprints, but handed over 00:07:58.189 --> 00:08:00.910 the entire assembly line. Okay. The notes from 00:08:00.910 --> 00:08:02.949 the engineers, the faulty parts they threw out. 00:08:03.019 --> 00:08:05.959 Every software tweak, it completely opens up 00:08:05.959 --> 00:08:08.519 the science behind it. Let's talk about the actual 00:08:08.519 --> 00:08:11.980 models. They released Olmo 3Think, a 32 billion 00:08:11.980 --> 00:08:14.740 parameter model. And that one is focused specifically 00:08:14.740 --> 00:08:16.459 on chain of thought reasoning, so it's perfect 00:08:16.459 --> 00:08:18.319 for researchers. And you have the base models, 00:08:18.500 --> 00:08:22.680 a 7B and a 32B. And these have a huge 65 ,000 00:08:22.680 --> 00:08:25.769 context window. That's 16 times bigger than the 00:08:25.769 --> 00:08:28.529 last version, so they can handle enormous documents 00:08:28.529 --> 00:08:31.250 for coding, math, comprehension. And there's 00:08:31.250 --> 00:08:34.669 also a smaller 7B instruct model, right? Tuned 00:08:34.669 --> 00:08:37.190 for chat and using tools, which is great for 00:08:37.190 --> 00:08:39.690 running locally. Yeah, but what's really revolutionary 00:08:39.690 --> 00:08:42.330 is the cost efficiency here. Right. The sources 00:08:42.330 --> 00:08:45.629 say the 32B model rivals the performance of something 00:08:45.629 --> 00:08:49.610 like Quinn. But it used six times fewer training 00:08:49.610 --> 00:08:52.409 tokens to get there. Six times. That's an incredible 00:08:52.409 --> 00:08:54.730 reduction in resources. Which means the training 00:08:54.730 --> 00:08:57.730 cost was only about two point two million dollars. 00:08:57.789 --> 00:09:00.929 That's that's. astonishingly cheap for a model 00:09:00.929 --> 00:09:03.110 this capable. It's like building a high -performance 00:09:03.110 --> 00:09:05.549 rocket on a motorcycle budget. It just shows 00:09:05.549 --> 00:09:08.669 what smart data curation can do. Whoa. I mean, 00:09:08.669 --> 00:09:11.509 imagine scaling that kind of full traceability 00:09:11.509 --> 00:09:14.809 to a billion queries, knowing exactly why you 00:09:14.809 --> 00:09:16.710 got a certain output, not just guessing. And 00:09:16.710 --> 00:09:19.210 it's all out there now. AI2, Playground, Hugging 00:09:19.210 --> 00:09:21.429 Face, you can run it locally. And it's all under 00:09:21.429 --> 00:09:24.169 the Apache 2 .0 license. That permissive license 00:09:24.169 --> 00:09:27.899 is key. It is. But that ability to trace a response 00:09:27.899 --> 00:09:30.320 all the way back to the original training recipe? 00:09:30.970 --> 00:09:33.950 That's the unique part. Researchers can now dig 00:09:33.950 --> 00:09:36.710 in and see exactly where a bias comes from or 00:09:36.710 --> 00:09:39.409 why the model learned a certain thing. So if 00:09:39.409 --> 00:09:42.909 you look past just the low cost, the single biggest 00:09:42.909 --> 00:09:46.470 benefit here is that complete transparency allows 00:09:46.470 --> 00:09:49.149 for critical investigation and much faster model 00:09:49.149 --> 00:09:51.549 improvement. It just eliminates all the guesswork. 00:09:51.769 --> 00:09:54.710 So we've really covered three huge shifts today. 00:09:54.870 --> 00:09:57.850 We have, first, robotics found a way to achieve 00:09:57.850 --> 00:10:00.500 complex dexterity, not with expensive... gear, 00:10:00.700 --> 00:10:03.539 but with a clever, cheap, crowdsourced data trick, 00:10:03.679 --> 00:10:07.740 the $200 gloves. Right. Then second, the AI performance 00:10:07.740 --> 00:10:09.919 race is shifting the whole landscape and putting 00:10:09.919 --> 00:10:13.120 this just unbelievable strain on our energy grid. 00:10:13.299 --> 00:10:15.419 Meta trying to become an energy trader is the 00:10:15.419 --> 00:10:17.419 perfect example of that. And finally, the very 00:10:17.419 --> 00:10:20.000 definition of open source has been expanded in 00:10:20.000 --> 00:10:21.779 a radical way. Yeah, showing that efficiency 00:10:21.779 --> 00:10:24.360 and transparency can actually rival sheer scale. 00:10:24.440 --> 00:10:27.159 These three things, dexterity, strain, and transparency, 00:10:27.320 --> 00:10:29.440 they're all colliding right now. And that collision 00:10:29.440 --> 00:10:31.779 is going to define the next few years in tech. 00:10:31.919 --> 00:10:33.659 That's the big takeaway. So here's a thought 00:10:33.659 --> 00:10:36.159 to leave you with. If these highly efficient, 00:10:36.320 --> 00:10:40.179 fully traceable models like OMO3 become the standard 00:10:40.179 --> 00:10:44.460 models that cost just a couple million dollars 00:10:44.460 --> 00:10:48.220 to train and give you total insight, how quickly 00:10:48.220 --> 00:10:51.659 do the old proprietary black box AIs become obsolete? 00:10:51.919 --> 00:10:54.559 The ones that demand meta -sized power grids 00:10:54.559 --> 00:10:56.860 start to look really expensive and you can't 00:10:56.860 --> 00:10:59.200 even see inside them. something for you to consider. 00:10:59.360 --> 00:11:01.159 That's a really important question for where 00:11:01.159 --> 00:11:02.779 this whole industry is headed. Thank you for 00:11:02.779 --> 00:11:04.740 joining us for this deep dive. We look forward 00:11:04.740 --> 00:11:06.779 to diving into the next stack of sources with 00:11:06.779 --> 00:11:06.899 you.