Transcription
So, it's not like we're just x amount of times smarter than an ant. It's like a different brain architecture that allows for completely new abilities. So then you can imagine there might be a step after humans, and a step after that. They're not just smarter, but they're like—it's a step function. This is the end of the sort of biological range of intelligence. And then up here, that's artificial superintelligence. We can't even perceive what's on the next level. Like, as ants are to us, if that same distance we travel up that ladder, we can't even imagine that.
And the biggest illustration for me is like AlphaFold. And then it started spitting out predictions for other proteins that we we haven't sort of discovered or knew the shape of. There's some sort of a pattern that we can't even understand or predict. We feed into this neural net. And it's like, "Oh yeah, obviously there's this pattern to how biological life is formed with the building blocks. Obviously, like here's here's what I can make one up. Here you go." We had a good run in terms of being the greatest, the apex intelligence in the planet. We've been the smartest things on the planet for a long, long time.
Hi, I'm Dylan Curious from the YouTube channel Dylan Curious, and I am really excited to talk with you, Wes. I'm very glad to have you here. Let's get started. I would love to just start with a question of like where, when did you have a defining moment where you thought that AI might kind of be worth all this effort? So, for most people, I feel like the ChatGPT moment was the start of something, right? For me, I saw Elon Musk tweet, "ChatGPT is getting scary good," or something like that. I was like, "What is this ChatGPT thing?" So I I went, I checked it out, and it was it was okay. It definitely had something to it. It wasn't magical. It wasn't mind-blowing, but there was some something there. I was like, "All right, this is this is interesting. All right, whatever."
Um, next I started learning more about OpenAI, and they had this really interesting hide-and-seek reinforcement learning, kind of like a video game where they're training these AI agents to play hide-and-seek. And you know, it wasn't anything new. It wasn't anything mind-blowing. I think they just did a really good job of presenting it, um, in a very engaging way. They did a great job of showing the data and illustrating the concepts. And it had these little blue and red people running around with smiles on their face. So it was like very like visually also kind of uh fun to to to see. And the premise was very simple. So there were two teams, two players on each team. Ones were one team was the hiders, one were the seekers. And basically the seekers—you remember that? Yeah. Um, it's from a while ago. So this is before ChatGPT. This was before LLMs really even took off. This was a number of years before that, I believe, and it was basic reinforcement learning, which has been around for a while. And so the red team would get plus points if they could keep the blue, you know, the the hiders in their visual field, and the hiders would get, you know, virtual high-five reinforcement learning type of thing if they were able to stay out of the visual field of the seekers. Very, very basic.
And so the first thousands of simulations or hundreds of thousands that they were playing, the AI agents couldn't do anything. They were just basically mashing buttons randomly. So they were like spinning in circles, kind of like having these like jerky motions, like no sign of intelligence or int or just just anything. And that kind of like—I kind of I I was aware of this but kind of seeing it live, um, was very interesting. And in fact, understanding that kind of ties into a lot of the stuff that's happening now because we didn't start them with any sort of human data, if that makes sense. Do you know what I mean? We didn't tell them, "Well, this is how I would hide." We're just like, "No, you figure it out. We're not going to tell you anything. We're just going to punish you if you do it wrong, and we're going to give you a high five if you do it right." And over time, over millions and millions of iterations, you saw intelligence emerge. And going through it, I was like, "Oh, okay. This is—I mean, this is learning." Like, to me, I was like, "This is what we would call learning. There's there's no way around it. This is machine learning." Um, and that was kind of the second kind of piece of the puzzle that kind of like opened it up for me. I was like, "All right, no, this is—it was something resembling the human brain gathering information and this and that." And by a billion iterations, they found a way to—they found some glitch in the system that the developers were not aware of. So, they were able to like launch themselves in the air to land where the the se the hiders were, and and that was a surprise to the developers. It's it's creativity, and it's like almost like, you know, the whole idea of like—we've we created their little environment. We thought we knew everything about that environment, but they figured out something where the developers like, "Wait, you did what?" Uh, I I just thought that was incredibly incredibly interesting.
And then so, so it's like ChatGPT, it was that, and then shortly after, a few weeks after, I think after ChatGPT came out, I think 3 or 4 months later, you had GPT-4. And that's when uh Microsoft published the paper "Sparks of AGI." So they were saying this is kind of a prototype AGI. Reading that paper, that was like the final—whatever you want to call it—nail in the coffin where I was like, "Okay, okay, this is a thing that will attempt to do whatever you throw at it, and maybe it'll do well and maybe it won't, just like a human being—like might fail, might—but it will attempt it. It will it would understand what you're trying to say." And so between the reinforcement learning side of it and kind of the more general aspects of the LLMs, GPT-4, somewhere between those two moments, it just like dawned on me, okay, I'm like, "We're on to something big." And I I I didn't know where it was going, but I was like, "There's no chance that this is not going to be a big, big, big deal." And I immediately went all in on it.
Yeah. You know, like that's so interesting because yeah, I can pinpoint kind of two kind of like a camel. There were two humps also where I first thought, "Okay, like there's something different between programming something and learning something." For me, that first moment was when I was watching Watson. This was a long time ago. And it was learning the different ways you could write the characters like A, B, or C. And it was just learning not the averages, but something general about what the letter A is and the letter B is. A simulated environment started making me realize like, "Oh, this could be like walking or reading MRIs or all these other things that didn't seem like it was a narrow domain." And then you have that second shock moment where you're like, "Oh, this is like really close to what it feels like might be happening up here."
Definitely. And yeah, I like the fact that you said, "So it's not programming, it's it's something else. It's learning." That was a big deal when I was like, "Wait, this isn't code. This isn't somebody writing out scripts. This is something again." Yeah, like the developer doesn't know how it works either. They're like, "I fed in the data and then I tuned the model, and now it's solved, but it's not programmed." Like there's no way to go trace that all the way back.
Absolutely. That was like a big mind shift for me. It was just like, "Okay, this is different." Yeah. For how many people actually still use ChatGPT? I suspect most of them think of it as a program. And we're kind of at a disadvantage trying to explain that too because you pull them both up on your iPhone, but one's an app and one's ChatGPT, and they just feel so similar. It's hard to get your head around the fact that behind the scenes one learned everything it's doing and one was programmed to do everything it's doing.
Absolutely. And that's definitely one of the things that I think more people need to realize because also they're comparing it to um things they're scripts, and they're like, "Oh, it's not as good as this program," and it's like, "Yeah, but one is static and one is like learning—it's completely different." Yeah. Someone was like complaining that it's like it can barely do the square root of some big number, and I was like, "Yeah, but it also can like pass the medical exam and like write short stories, which your calculator can't do." So like it's a little bit impressive, you know, if you think about it.
Absolutely. So, are you kind of surprised that it felt like humans were so on this planet as the center of the universe, and then we're like, "Okay, maybe we're not." And then you think intelligence is like the thing that makes us special because you see technology take away all these other things, and now it's like we're grappling with the idea that maybe even intelligence isn't that special, right? Do you feel smaller and smaller and smaller over time as you dig into this stuff?
Yes, but it's good. Um, I think when Ilya Sutskever, like a years years ago, he said, "If the only thing you value about humans is intelligence, their intelligence level, you're going to have a bad time." And that's literally what what I think he meant. And I think that's what we're seeing because anytime AI intersects within somebody's realm where they they feel very proud of themselves, they're like, "Oh, I'm so good at this thing." And then AI comes out and just like levels them, people get angry. And I see this like in in comments on my my videos, other people's videos, on Reddit, whatever. People are very scared of that. There's a lot of anger because they're they feel threatened that this thing is better than them. If they if they're very proud of their intelligence and their abilities and this thing comes out, it's smarter. If that's the only thing you value, you feel like you're under attack. Um, and so that's what I mean by like if you let go of that a little bit, if you let go of that, I don't know, ego. I don't know if that's the right word, but like you just can't take yourself too seriously. Like yeah, we've been we had a good run in terms of being the great the apex intelligence in the planet. We've been the smartest things on the planet for a long, long time. Um, and now it's this emergence. It's so weird. We're going to give that up to something we created, which is so different. Like usually evolution came up with something that became the new apex predator. We might be the first ones to really just build something that we then succumb to.
With "succumb," you mean it's going to kind of supersede us, not necessarily kill all?
Well, the weirdest thing is I guess "succumb" like in the sense that like even—Okay, like let's imagine it's like super benevolent, and we live in like the perfect zoo. It's still kind of succumb to like—it is now in charge of the direction of humanity, I guess, you know, and um, it's just strange. I mean, maybe bad or good, but yeah, I mean I I don't know if that's going to be as big of a deal as people think it is, and I might be wrong on this, but I mean like I'd love to hear—yeah, I mean like calculators are better than us at calculating stuff. Like nobody has existential dread over that. The chess AIs are far, far superior to humans. People still play chess. People still enjoy playing chess.
There's this guy Scott Aaronson who is just really smart, top-tier physicist. He worked for Google on their quantum chip.
Scott Aaronson.
Yeah, I'm pretty sure that's that's it. And Ilya Sutskever invited him to Moonlight for the OpenAI safety team, um, a number of years ago. So he—which was not his field, right? So um, Scott Aaronson's field—he's a physicist, again extremely smart, but AI was a brand new thing to him. So he came into it not as an expert but just as a brilliant person from whom they wanted him to contribute in whatever ways possible. And so it was interesting to see this really smart person start fresh in a brand new industry, brand new field like AI and grapple with some of these concepts. And some of the stuff that he came up with was brilliant. There's a talk that he gave; it was a TED Talk eventually, but before that, there's another talk on YouTube where he's practicing in front of a room, and he's a lot more like relaxed and natural, and it's it's a great it's a great watch. But towards the end, people are asking him like, "So what happens when these AIs are better than us at everything?" And he's like, "It's not going to matter because like, you know, we have cars that are faster than human runners. We still watch the Olympics. We have chess AIs that are better than this. We still watch the Olympics. It's like—or we we still play chess." So he was saying that we're going to compete with each other, and we're going to get enjoyment from that. We're going to build—we're going to write books because we're going to enjoy writing books, and other people are going to enjoy reading other people's books. Right? If we're going to go for a hike, we're going to enjoy the hike because of the experience that it gives us. So, but what he's basically saying that instead of like everything being measured in economic units or like us trying to climb some social ladder, it's just—we're going to be doing it for the experience, the—you know what I mean?
Yeah.
Yeah. That's okay. So, cuz in my head I'm just playing it through like one by one. I'm like, "Okay, so we invent chess AIs that are better than us, but we still play each other for fun, and some people learn from it. And then maybe it drives cars better, and then it starts getting into more like decision-making things." And this is where it gets a little fuzzy for me. Like I can imagine a government, a president AI that does a better job than any human that we vote for. So we kind of put it in charge. And then I think about like medicine or mechanics kind of being outsourced until there's really nothing in my life of utility or of kind of use. But then I I guess there would still be like—I'd get with my friends to like learn about medicine even though like I don't need to. Or like would we just kind of hike and and do just pure physical things? And then what if a robot's like better at hugging you and like loving you and like—then people are like, "Well, it's really hard to get another person to act the way I want, and I can just date this thing"—then it doesn't—it kind of feel like we've—I don't know—the chess analogy breaks down a little bit.
Yeah, I totally get that. And this is going to be very interesting to see kind of what happens because on one hand, yeah, like right now a lot of the things that we do, they're kind of economically driven, right? Uh, we make decisions, and people that make good decisions get rewarded for it like in in in society, like if you're running a company or whatever. Um, okay, so let's take—let's say you take that away. Um, and now you just do whatever you want. Does that make for a good life? I mean, we see people that are born into wealth that are miserable or people that achieve a lot of wealth that are miserable. I mean, like like Jim Carrey talked about it; he used to talk about a lot how he like strived so hard to make money, and then he made all this money, and then he's now saying like it's all meaningless and like don't even bother making all that stuff cuz it's meaningless. Like okay, let's imagine for just hypothetical—like let's say there's a digital Wes Roth twin, right? Like you have this iPhone app, and it's just pretty much almost every time it makes a decision for you, sets up an email, like makes a restaurant order, like it's pretty much always what you think you would have done. Is that the kind of thing you would also outsource your vote to? Like would you say like, "Why don't you go look at all the different candidates and like vote for me?" Is that the kind of thing that you would say—what—like, "I want to impress this girl; like you come up with a fun joke, and then I'll say it." Like would you go that far, or would you just truly say like, "I'm not going to open the app. I'm not going to talk to my digital twin. I'd like to just do this on my own."
I mean, that's very thought-provoking. Um, and I talked about this a little bit on a live stream. Somebody hit me with a similar question. I was like, "Oh, man." Cuz the thing is I I—it does feel so—so the the idea is like what if like all the decisions were made for you? Everything was nice and comfortable. You didn't have to stress out; like you had everything. Whatever joke you want to tell a girl, like you had the best joke right there and there. Um, and we see a little bit of that in modern society where we have more and more stuff, and people are not necessarily happy. In fact, a lot of people are miserable. So, I think you know what—you know what I think—okay, I think they nailed it in mat in the original Matrix movies. Remember when um Agent Anderson is is is that his name or Agent Smith?
Agent Smith.
Yeah, the um—he he goes, "You know, at first we created the perfect utopia for you humans, and you guys were miserable, so we just scrapped that, and we created this crappy world, and you guys were just happy. You needed to be miserable." So, I think part of the—we do need to be a little bit miserable; that there needs to be some sort of a negative negativity or some sort of um something that we're striving against, running. You know, we we all got to work. It's like certain dog breeds. If they don't work, if they don't feel like they're contributing, they're not happy. I think humans have a little bit of that. So, I think that at some point we're going to eventually—AI or no AI—at some point we're going to reach through just automation or whatever; we're going to reach a point where no one's going to need to work as hard, and we're going to need to—and we're going to be a little bit more miserable. We're going to need to figure out meaning that's outside of all of that. So, I I—if I understand your question correctly, it's like would you—I don't think we're going to be happy if everything is handled for us. I I don't—
What's the solution to that?
I I don't know.
Yeah. I mean, because it's s it's such a human problem because—Yeah. like you know—Yeah. You know you don't want to fail right now. So you want to lean on the cheat code, and then you lean on the cheat code month after month after month, and your life has no satisfaction. You know it always feels like the wrong move in the moment, and it always ends up being the wrong move in the long run. In a video game analogy, you're playing a game; you're interested, but you're struggling. You type in a cheat code, and you go from struggling to—it's effortless, and you lose all interest. It's like life is probably the same. Like if you have everything, you're going to lose all interest. So it seems to me like building in some sort of artificial scarcity or artificial strife, hardship. That seems to be a good thing for humans. So at some point we're going to be thinking about how how to do that. I don't know. Like for me right now, if I'm not feeling good, what helps sometimes is I go and I have like a hard workout or I go running. I do something that's not necessarily pleasant, you know what I mean? It's something that's hard, and that surprisingly I feel a lot better afterwards because it's sort of like that contrast. So if you're—if there's no contrast in life, you lose the ability to really appreciate the good stuff. So I think that's going to be a big question. It's like how do we create contrast?
Well, yeah. And if we have—Yeah. And if we have 8 billion people on the planet, and most of them need to now learn that that skill, which is like the self-awareness that like, "I don't want the easy answer even though it's clearly all around me." It's like we have to get used to the Star Trek replicator world and not just leaning on it. You can't just get addicted to the holiday, you know? You have to somehow know your own weaknesses and like put some artificial fences around yourself to make it harder.
So yeah, like do you feel like we're going to regret inventing AI, or do you think this is the same kind of progress of technology lifting up the the world and a better world?
To me, you know, each big technology from the past kind of—what it gave us is leverage to do more. And that of course usually means we can do a lot more harm, we're a lot more good. AI is the biggest leveraged technology we've ever had and potentially will will ever have. So yeah, I think it's going to be the the, you know, the day we invented AI is going to be like the day we rule forever and just like—why do we do it? Or it's going to be the greatest day for humanity ever, and it's going to be one or the other. You know what I mean? Uh, and I don't see how it could be neutral. And do you know what I mean? Like it's going to have a big impact, positive or negative.
Like what would you put your P(doom) at?
I don't like the idea of just talking about P(doom) because I think there's a number of other really bad dystopian outcomes that has to be considered. I feel like if we're just talking about P(doom) and that existential risk, some sort of catastrophic X-risk, we're ignoring like, for example, somebody jokingly said, "Well, okay, we have our P(doom). What's what's what's P(1984)?" So like what's the chance of—you know, that Orwell book?
Yeah.
Yeah. Like what if like some tyrannical regime takes over? Everybody's surveilled all the time. Everybody's brainwashed, and there's no way to go against it or or overthrow it. Because I mean if you're talking about P(doom), a lot of the solutions to not pursuing AI development is kind of more centralization, more cooperation. There were some talks about maybe even some sort of a like a world government organization that's like top-down making sure that nobody's developing AI. Um, I mean, if you're really concerned—if you have a high P(doom)—what is your only solution? It's really like get the governments involved and like shut down development, right, if you if your P(doom) is 99%. Like what else what else do you do? But then what happens if, you know, the government decides to abuse that or develop it but for their own needs. So it's like—I think we need to talk about all the risks holistically. That's why like Dario Amodei how he kind of approaches cuz he's like, "Hey, yeah, we got to talk about X-risk. We got to talk about P(doom). We also got to talk about what happens if an authoritarian regime gets access to this technology before everybody else." So he's like, "There are multiple bad things that can happen. We have to look at them holistically," if that makes sense. And of course, if you centralize everything, you increase the the risk of an authoritarian regime or whatever, potentially, right? Like people with high P(doom)s like Bostrom—they—yeah, like they kind of—you start asking like, "What does this mean?" Well, it means governments get more kind of like tight-knit about what's going on. And if people are not—some countries are not compliant at a certain point. You know, if they're really not compliant, like the same way with a nuclear, you know, checking program, like you might have to, you know, bomb a server or center or something and like stop and like show that you're very serious about them participating or cooperating, you know. So, we're talking world government with complete control over countries, something in that direction. And there was a speech where
Um, somebody was actually kind of suggesting that talking about that. So it's not me just making stuff up. I mean, that's how you would deal with that. That also scares me. You know what I mean? Right. Yeah. No, that comes with a much bigger sense.
I'm kind of curious, just out of like the the people that you've sort of seen, like the Elon Musks, the Sam Altmans, the Demis Hassabis of the of the universe, like who are you kind of leaning towards like like agreeing with? Like who do you sort of trust the most and sort of not trust the most? Or like, how are you feeling about the leaders in the AI space or the CEOs of the major tech companies?
I feel like, first of all, I try not to fall into the thing where anybody's like an angel or a demon. They're all good or evil. I don't think anyone's like that. Everybody's like this flawed mix of whatever. So, I don't think everybody's somewhere in in in the middle. Um, in terms of trust, the only person like I feel like I fully trust is is Dennis. I don't know why, you know, he just I actually feel the same way. Yeah. He was just there before everyone else. And he tried to like he wasn't in favor of pushing this into product so quickly, and he seems like he's genuinely smart and he's actually there on the ground level building stuff. He understands the technology, and it's weird because when you do interviews with him, like no matter what they throw at him, like you want to talk biology or this, he he knows everything, and he's got like a just a really great grasp on everything, and he genuinely seems like he cares not about himself or whatever, but humanity. Everybody else, I don't think they're necessarily bad, but they're more more human, sort of, you know what I mean? Like just like the rest of us, they have their own ambitious ambitions and goals and whatever.
I mean, do you think Sam's pushing product too fast? Or do you feel like his take is incrementally pushing stuff out allows humanity to kind of like absorb it and figure out what to do with it? Um, that kind of after like watching what's been happening, that kind of is making sense to me. Uh, the idea that let's just iterate and push it out, iterate and push it out, and then we're able to sort of like on the fly build certain guardrails and figure out where it's going to start breaking stuff. I know I'm sure a lot of people think that's crazy. They think that's dangerous. I'm not saying it's safe either. I'm sure at some point they're going to push a model that's going to freaking break something. Uh, meaning that it's going to cause some damage, cause some issues. But maybe that's how that might be the safest way. It does seem like it has a little bit of an evolutionary or marketplace dynamic to its kind of safety, like in the same way the cybersecurity is kind of constantly evolving with cyber attack.
Exactly. And so I think I think there's some truth to that. Now, of course, that might play into, you know, his hands in terms of like the company valuation and stuff like that. So it might be he he's might be playing his own game, but just on the surface it makes sense, this idea of rapidly pushing stuff out there and just letting societ forcing society to figure out how to deal with it. I know that sounds bizarre. That sounds risky, but at this point it might not be such a bad idea. It's like a very evolutionary process, like like you said. I think that's a good way of looking at it.
And then Elon Musk went from just assuming that everything's, you know, AI is going to kill everything to now I don't know if he's more positive about it or at least he wants to at least control it to be part of the development. Um, I'm very much looking forward to cuz they're building that giant data center, the Colossus or whatever it is.
Yeah, the Memphis one. I think it is.
That's where it is. Yeah, it's massive. Bigger than I think anything else as a single data center. So the next iteration of Grok, if they build it with the full power of that thing, that thing's going to be massive.
You feel like scaling hasn't hit a a roof? I never quite know. Like I hear people say like, "Oh yeah, we've kind of now it's an architectural problem." Sometimes it's just a scale problem. But the thing is like scaling is an exponential cost to like a linear progression, right? So you 10x the thing, you get, you know, you double the the sort of the the ability to the model. And certainly the jump from, you know, 4.0 0 to 4.5 wasn't super awesome and impressive. But here's the thing, we also have found a lot of other ways to scale, not just more computes, you know, with with test time computes and a lot of the different sort of architecture changes and stuff like that. So, I feel like we're finding other avenues of improving it that are also scaling it, but it's probably not scaling in the same like not just more more computer or whatever. So, I think that's not going to be like the only path to improving it. There's going to be a number of different paths, and I think that still just we're going to need more Nvidia chips and more, you know, stuff cranking out data and stuff like that.
Like when you look back, is there a favorite uh video that you put out there that you're like most proud of? The videos that tend to do poorly are some of my favorites. The ones where I just go deep into some rabbit hole and it's just kind of there for me. Uh, but it's not like what appeals to most people cuz new model releases and stuff like that, that's what people want to see. The stuff that I mess around with is like, for example, there's one that showed that models that AI models have their own mental models that they understand how the world is, even parts of it that we don't kind of tell them about. They kind of build an abstract representation; they guess at the existence of the outside world. That was one of my favorite ones actually.
Is that new? Is that the anthropic paper? This was actually one of the older ones, and that's the one where Andrew Ng was saying these models understand because they so uh Oll GPT I would look at that because that one is mind-blowing. So in Oll GPT they fed them games from um a game called Oll and just the notation of the moves made, and after a while the model started to be able to predict what move comes next, which is interesting, and that's kind of like what they're supposed to do, but later on once they started digging deeper they realized that the model had kind of a representation of what the board would look like. Nobody told it about a board. Nobody told it about players. Nobody told it about squares, but it it it created a mental model of what the board looked like. See, like what if that's how consciousness builds up, you know? Like you think about yourself when you were like 3 or 4 years old. You just didn't have so much buildup. So it almost felt less.
Yeah, dude. That's crazy. It's weird to think about. So I've only like kind of skimmed somebody who summarized it, but I know there's an anthropic paper from this week where they were using those autoencoders to try to look into how some questions were answered. And when they asked it to do a a math edition, when they asked it through theory of mind, how did you come to the solution? It said, I carried the one and did all this stuff. But then when they watched it through the autoencoder, it actually had like a whole network that activated about like what are ridiculous numbers, and it kind of honed in on what the average answer like zero wouldn't be a good answer. A billion wouldn't be a good answer. And then another part of the network kind of did a different thing than it thought it did to calculate it. Like it it did add the numbers but only in the first column and then just generalized on the other half, and it was like it's so fascinating. It's it's so bizarre. It's so fascinating. I can't wait to discover more about this stuff cuz that's like a whole rabbit hole of like what is happening those neural nets cuz it's not like in the beginning people like oh yeah just stochastic parrot predicts blah blah. No way. There's some insane insanity happening there, and yeah, just fascinated with that. I mean, that's part of why it's kind of fun to make all these YouTube videos. I'm, you know, I'm still not like a profitable channel yet. But I'm like, the reason I love this is cuz I just love thinking about it, and like we're going to be on such an adventure for the next 5 10 years. Like the whole world is going to change, and like we're going to be trying to digest it every single day, and it's going to be I it's going to be wild.
All right, I know this has been an exciting discussion. If you want to hear more, we're going to continue this on the YouTube channel of Dylan. I'll leave all the links in the show notes and in the pinned comment. Definitely check it out. Give them a follow. And we're going to be talking about life extension, free will, and all sorts of other AI-related stuff. So check it out, and I'll see you there.