Mike Israetel: evaluation

Transcript context

…This is the other issue that that I have with your perspective, that you have, I would call it, a cortical centric view of of cognition. Absolutely. Yeah. Exactly. So this is another 1 of those leaky abstractions I was talking about. By the way, there's a great book called The Brain Abstracted by Marvita Chirimuta. I interviewed her recently. And she said that 1 of the most pervasive myths in neuroscience is that we use these leaky abstractions and idealizations to talk about cognition. And usually, it's using the most recent technology at the time. So a few hundred years ago, we were describing the brain in terms of pulleys and Pulleys and levers, yes. That's right. And then it was as a prediction machine, as a computer, and all this kind of stuff. At the end the day, this is an example of these are grounded things that we understand. They're really good models because we can both talk about computers. We both know what computers are. But the brain doesn't work like that in any sense. And a great example of this is knowledge. You as a personal trainer know this, right? So I'm sure in the past you've trained folks in the gym and what you've probably realized is that in your mind, you've got a distilled abstract beautiful like you understand things, you've thought about it, you've distilled it, you could write it down in a book and you probably think all I need to do is just tell people and and I I've done all of the thinking. I've figured all of this out. I just tell them and what you learn to your chagrin is that there is no substitute for experience because part of knowing is actually doing. It's just the pure experience of lifting the weights, just feeling the sensations in the body going through that process. That's why knowledge is non fungible. I've been fairly impressed incrementally with modern AI's ability to render pretty good recommendations on exercise and sports science and I'll push you back on 1 of those things. Know a lot of people, they're great people that have a lot of what you would call embodied knowledge about training. Because they haven't been interested in or capable of doing the abstraction to distill principles, they're not really that great at training other people. They They know what works for them and they know scarcely of that and they're just not good at generalizing because they've never abstracted out the concepts. I've also had many discussions with some of my very very smart friends like 1 of my friends graduated from Harvard, masters data science, worked for Apple and a bunch of other AI companies and stuff. And his ability to pick up how to do an exercise incredibly well just off principle discussions and then trying it like twice. He's instantly like 1 of the best trainers I know in the area. 1 shot. Why? Because Jared and I can distill to you like 5 principles of how to do an exercise well, which are geometrically not complex, heuristically very cogent, and you will instantly be good at teaching people how to exercise without ever doing the exercise. Is some nuance missing? Absolutely. There's some stuff and we can abstract how it works that you have to do to actually like, oh I see moving the hips back like this. But if someone was able to describe that to a computer in vectors of how hips should move back in a squat versus just down, especially if so so I'll tell this. If I was able to train a model, and I'm not currently doing this and had no plans to do so, of giving it 1000 samples of a squat done right, 1000 samples of a squat done wrong, and about 10 rules of how to squat, it would instantly be 99 percentile of teaching people how to squat with 0 grounded experience. And I would also say, what we consider grounded experience is mostly visual data stream. Like, yeah, touching stuff is cool that it gives you something like I'm not letting the computer have sex with me until it's had sex with test animals in the lab, that came out all wrong. But you know, lot of our data is visual. Huge huge fraction of them. I suspect that if you let a neural network train on all of YouTube, which I understand isn't happening because there's like copyright bullshit, which drives me insane. You just like have like an algo just eat all of YouTube, which which is preposterous amount of data. It would know more about how the world looks in real life than any extant human. Period. Because there's more visual data on YouTube than I have collected with my own eyes by, I don't know, 12, 50 orders of magnitude something like that. And so 0 embodiment. We did the embodying for it by putting up camcorders into people's faces while they throw up from drinking challenges, and then downhill skiing, and then the pictures of microbes. And so 0 embodiment. We did the embodying for it by putting up camcorders into people's faces while they throw up from drinking challenges, and then downhill skiing, and then the pictures of microbes. If you really train a visual model, and even better a 3 d relational model on YouTube alone, you will get a more grounded understanding than your eyeballs, which by the way are also just cameras, can ever reference to your brain, which is also a computer. Period. The brain is a computer. Is it a different type of computer than a CPU? Yes. Is it different type than a GPU? Yes, but closer. Are we going to replicate the exactly how the brain works computationally? Yeah. Sure at some point. I suspect we'll need to do that because there's no good reason to believe that brains are like super super good at thinking. They're just evolutions best crack at it. You know, like a tank isn't a walker. Are you a Star Wars fan? No. Do you know Star Wars? Do you know what an AT AT or ATST is? Okay. You know those dog looking big walker things that shoot lasers out of their faces? Like what's better that or a modern battle tank? Well modern battle tank by a long shot why? A battle tank sits like 8 feet off the ground. It can hide in trees and it can shoot like 2 miles out. An AT AT walker stands above the tree line and you just shoot 1 leg and it falls. Why the fuck would you do that? Well, it's a pretty decent attempt to replicate animal locomotion, but the assumption that animal locomotion is like real the best way to do it is not. So just the same way as like a rocket or a supersonic aircraft is like categorically better at flying than a bird is in almost every respect. AI and machine intelligence is going to surpass human intelligence and real deep understanding of the real world just by bypassing our architecture, and then later being so smart that it can come around, scan the human brain, be like, oh, that's how they were doing it. Oh, that's kind of cool. I don't think there's a huge ROI of what it's gonna get out of that because it's gonna be way ahead. But basically suffice it to say, I think that what we assume we're getting from sensory experience is like just a few, I don't know gigs or some number of bytes of data. Like I've seen a lot of shit in my time, but if if it's all also by the way wildly abstracted. Like if you ever think about memories you've had from childhood, and you try to really parse like how did this dresser look? What was my mom wearing? You realize 2 things. 1, you had precious little actual visual data. It's like a 10 second long video, maybe at any kind of fidelity and the fidelity sucks. And 2, is you're actually hallucinating a lot of that, like you're just pretending in details. If you let an AI train and eat all of YouTube at 4 k and really like develop the model from that, it would see and understand the visual three-dimensional space around it and the world at a level of embodiment that shits on all of us instantly. That is my conjecture.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.