Speaker 0:  robots have officially stopped being clumsy lab experiments and finally started getting to work. generals ai just dropped gen one, a model hitting a ninety nine percent success rate on tasks that used to break a robot's brain, and it's doing it three times faster than anything we've seen. so these robots finally get human hands. the gen one is an embodied foundation model that masters complex manual labor with near perfect reliability. it bypasses the rigid programming by learning from five hundred thousand hours of real human movement data. so is this the gpt three moment for physical labor where we've officially moved from robots that follow a script to robots that actually understand the physical world. what do you think about this, john?

Speaker 1:  you know what? it it it kind of is. so let's let's let's dive into what gen one actually is. so it's called the robotics foundation model. so it's a similar concept to llms like gpt, but it's designed for the physical world. so instead of generating text, it generates real time actions for robots interacting with objects. so it's trained on hundreds of thousands of hours of real world robotic data. try saying that five times real fast. allowing it to generalize across tasks instead of being programmed for one specific job. so the key breakthrough here is reliability, repeatability. so previous general purpose robotic systems struggle to exceed, i didn't know this, 64% success rates on real world tasks, where gen one jumps to 99%, which is a massive leap. so why this matters? well, robotics isn't limited by capability. it's limited by consistency. at 60 to 80%, systems are demos. at 99%, systems become deployable in production environments. and so i would say this is the same threshold shift, and feel free to to correct me, but that we saw in speech recognition, machine translation, and even self driving perception. what do you think about that? do you think that's kind of the breakthrough here or the threshold?

Speaker 0:  well, i i think we're getting there. i like to think of this as an update like a line cook in a busy kitchen. old robots were like a microwave. they could only do one thing, and if you put the wrong dish in, they'd fail. gen one is like a professional chef who could handle the chaos. if a plate slips or an ingredient now is out of place, it doesn't freeze. it just adjusts the grip and keeps cooking. we've got a little video of it putting, taking money in and out of a wallet on the screen for those who are watching. however, if you noticed, i have no bills, that are that flat and straight ever, in the wallet. so it's getting there, but i wouldn't consider this a real world task when you look at it putting the money in the wallet. that's a completely empty wallet with an easy access, with completely flattened bills. but it does a good job. it force make sure it's in there. it sees what's not in there, grabs the edges, puts them in. however, if you've got a foldable wallet and the money isn't completely flat like that, what's gonna happen? i think we're taking things to the next level. i think this is a huge leap forward. i am just not buying all the hype that says, oh, this is the next thing. i mean, it, i guess it's in the name. it's gen one. so at what version at what point does this, become really usable? i don't believe the 99% rate just because looking at this, i'm saying the 99% rate in an environment that is friendly to them versus a normal environment.

Speaker 1:  so would you say that this is the gpt one of llm, for robots?

Speaker 0:  yeah. i definitely think we're getting there. i don't know about gpt one because i have to say i didn't use gpt one, so i don't know how crappy it was.

Speaker 1:  but it's not it's not that i yeah. anyways yeah.

Speaker 0:  yeah. so, the 99% is verified across 200 repetitions, a massive jump from the 64% in gen zero. twelve second cycles of box, folding speed that, the test, i don't have here for you to show, improved from thirty four seconds to just twelve seconds when it was folding boxes. 500 k data hours. it was trained on petabytes of human to object interaction. so this is, like you're saying, more like a world model than, say, a text model where we're talking about interacting with three d objects in space and time, not just giving it text. so, i'm really interested in this. this is a very significant leap in physical ai. it's the biggest since the transition from hard coded robotics, rpa, to machine learning based vision. so it effectively renders twenty twenty five era scripted robots obsolete. so we can now have, robots that can respond and interact with their environment if it's not, a 100% way where it should be. if, you know, the thing is not in the fridge or it's not in the right spot in the cupboard or on the counter, you can react and figure it out. so i think we're definitely getting there. i i just i've heard this before. like, we're gonna have full self driving car five years ago, weren't we? but, you know, laws and registration, restrictions as well as capabilities aren't quite there.

Speaker 1:  well, we do have it. it's called, what's it called? the mode on, tesla it's not fsd. it's a special mode on the tesla.

Speaker 0:  yeah. and it's a self driving five. but it's

Speaker 1:  not dog mode, but, like,

Speaker 0:  it's a good one. with the waymo taxis and whatnot, i'm saying that when they encounter situations they don't understand, they break down. yeah. this kind of technology is going to change that. ideally, getting it to react in real time. like, we've heard the waymo and the other ones, where, if they're they go in a situation where there's police or fire or whatever, they just stop and start blowing their horn or shut down, and they don't know because they don't know what to do, and they're in the way of everything. so that's what i mean by, we're not a 100% there yet because it doesn't know what to do with the edge case scenarios. but, i mean, if you're still in the driver's seat with your hand on the wheel, then you're fine. however, it can handle 90% of it. but, like, you know, we do construction on the road. you drive to toronto, you know, where the lines, switch because of construction, you know, the car can't tell which is the right line and which isn't. no. so i

Speaker 1:  can't even tell which is the right line and which isn't. so jeez.

Speaker 0:  yeah. so i that's what i mean by real world scenarios. that's not, i think it can if the situation is clean and predictable and follows normal guidelines, great. yeah. however, as soon as it goes off the rails, that's when gen one starts to kick in to give us more functionality, but and it's so much more reliable than gen zero. the question is, how rapidly are we cycling? when will we see gen two? when will we see gen three? is this three months, six months, years as ai gets faster and faster, we get new models, two or three new models, flagship models from the main companies almost every year now versus it was one yeah. a couple years ago. so things are really accelerating. the data hands used for the screening are actually low cost wearable pincers, not high end haptic gloves. this allows the general ai to crowdsource data from actual warehouse workers rather than engineers in a lab. so meaning, you know, people can wear them and use them, and it gets real data, and that's how part of the training process. while tesla's optimists are still struggling to perform useful work this year, general ai compressed a five year development cycle into just five months by applying the scaling laws typically used for llms. and this is what i mean by it took five years to five months. when's the next generation coming down? is it gonna take that five months to one month? is it gonna take you know, how quickly does this keep going? and the robot's ability to shake a plastic bag to settle an item inside was an emergent behavior. no engineer programmed the shake, and the ai figured it out, was the most efficient way to use gravity just like a human would. so i don't have a video of that, but the idea of it finding a plastic bag and need to shake it to open it up a little better to put something in it that it figured out on its own is a remarkable breakthrough. yeah. so the last 1% problem in the high stakes of manufacturing, a one percent failure rate is still a disaster. so we need to see if gen one can maintain the position in high heat or extreme dust industrial comp, environments with dust and whatnot. general generalist ai didn't just build a better robot. they built a brain that makes every robot on the planet smarter. so now that we've solved this 99% success rate, what is the one chore you'd pay a thousand dollars to never do again? let us know in the chat. i'm see i'm guessing we're gonna see a lot of laundry and stuff. so i'm curious to see what, you would offload to a robot.