The Alien in the Room

wait you're listening okay all right okay all right you're listening to radio lab radio from wnyc okay after all of that it is time to finally discuss let's if the question yeah the topic okay the theme of the moment perhaps climate change no that nobody cares about climate change man come on hey i'm latif nasr this is radio lab where despite what reporter producer simon adler just said we here at the show including simon do care about climate change but we're here today to talk about a different huge overwhelming thing that we're all in the middle of i mean i don't want to put words in your mouth but what i have been feeling yeah is a general sense of frustration yeah something that everybody's talking about but nobody seems to actually understand you and i have even done interviews together with people on this stuff that's right which is of course artificial intelligence so much of the coverage about this stuff right now is like this running debate right where you've got people on one side saying these ai you know they think they are intelligent and eventually they'll outsmart and destroy us all right and then on the other side you've got people being like no they they they aren't actually intelligent they're just mimicking us and it's not as big a deal as everyone says right and i i don't actually know who to believe yeah and i think it's because like i don't know what ai is like i don't know how it does what it does under the hood yeah because we don't know right this is one of the most extraordinary things about you know machine learning ai is that we don't really know what they are but after reading countless articles talking to tech people and scientists i finally felt like i was getting at that question when i talked to this guy stephen cave i'm the director of the lever home center for the future of intelligence he leads this sort of think tank at the university of cambridge and there's about 50 of us now trying to understand these systems using a really wide range of methods including tests taken from animal psychology tests designed to measure how well a mouse can problem solve and applying them to ai agents in order to understand well where are we in the kind of evolutionary cognitive tree of life of ai and they've actually turned these tests into a sort of competition that they call uh the animal ai olympics yes uh indeed okay well that just sounds fun right yeah exactly yeah so to do this they've created a slightly lower resolution toy story looking digital world okay or maybe even more accurately like if you know the game minecraft oh yeah yeah sure sure it looks like that it's this three-dimensional space filled with all these different bright primary colored objects okay and then they take these ai which are running on basically the same kind of engine that power chat gpt and they give these things a little avatar like a hedgehog or a pig or a panda and then they just sort of place them in in this 3d world and say there is food in here find it so it has to like navigate the digital world to find i mean like i assume it's not really food it's this green orb that they're looking for okay and i mean there are walls that they have to like figure out how to get around there are transparent walls but it's like it's like physical world problem solving absolutely and i mean well this is the sort of task that mice or pigeons can pull off pretty easily for these ai agents things like manipulating understanding gravity it's a real um challenge like they struggle to press a lever or perceive an edge which any animal can do or at least you know any mammal say and so effectively these systems don't have the common sense of a mouse whereas higher reasoning maths and so on they can do a hell of a lot better than humans can that's the that's the more of x paradox right like it's like easy things are hard and hard things are easy uh exactly you yeah and like we've known this for a long time and it's pretty obvious at this point but after running all of these ais through this thing dozens hundreds of times what steven has seen over and over is that they have a completely different profile of capabilities and skills than any animal they are not like us no i mean one of its capabilities might be convincing us it's human like but it isn't well okay so then what is it like i mean is the ai little tadpoles or what what is it well there is one metaphor that some people like to use and that's the octopus you know what's wonderful about the octopus is they are phenomenally smart that they can use tools for example without being taught um they develop sophisticated tactics of all kinds for lots of wonderful octopus escape stories well wait because that doesn't sound like ai at all um no then why this metaphor um well it's helpful not because ais are like them but because in a way it really shows how different intelligence can be okay i mean octopuses their intelligence is distributed through their tentacles he says you know we and all mammals have this one central brain but octopuses they have nine little brains one in the center and then one in each limb so their tentacles can function much more independently which is how they manage to have them all doing like clever things all at once um and you know this kind of intelligence is is fundamentally alien to us and that's a good way of looking at ai alien profoundly alien which on the one hand makes this thing feel sort of unknowable impossible to understand but then on the other well it is alien it did not evolve in some far off galaxy or even the depths of the ocean right like this is an alien we created year by year transistor by transistor and so this is what we're doing today we are going to trace the evolution of this alien in our midst this alien that we designed in the hopes at least of like coming to some deeper understanding of what it actually is today and then maybe if we're lucky that will give us some insight into this thing we are all almost certainly going to have to face off with at some point or another so this is great like i feel like we all need this we all need this explainer great uh fill your glass because here we go hey you guys can hear me yes i can hear you simon hello terry how are you very good thank you uh sorry for the slight delayed start here some classic technical difficulties you know so there there are a lot of different first contacts yeah we could point to with this alien species but the most fun place to start that i found is with this guy terry sanofsky professor at the salk institute for biological studies who uh yeah sort of like the midwife of ai is that a helpful way to think of you or no yes yes actually uh well it's obviously uh more complicated but but that's that's not a bad analogy terry trained as a neurobiologist he came up poking probes in monkey's heads to try to understand how the brain works but then in the mid 80s he teamed up with some computer scientists trying to make computers do and then he came up with a new technology that animal brain like things like hear and recognize sounds or visuals but but but it was going nowhere okay because every everything was based on rules at the time like all computer programming at this point it was this incredibly complicated set of like if this then that statement so if you see this and you see that but you don't see that then that means this right this sort of web of logic right which when it comes to recognizing sounds or pictures was a problem because for each rule there are you know tens of thousands a hundred thousand exceptions just too many nuances in the rules to hard code in and so it was uh clear that this approach this way of doing it through rules was really hopeless and so together with my friend and collaborator jeffrey hinton he started to wonder if there was a different way to tackle this learning and so with the small group with computers that were puny by today's standards they set out to build a machine that could learn and one of the first things they tried to teach it was how to pronounce english you know text to speech in computer science and amazingly demonstration of network learning by terry and charles rosenberg they have recordings from these early training sessions now if you want to learn from experience you have to have lots of data and so so you ready oh ah okay sometimes but can you They took a transcript of a kid talking, a transcript I had my friend and neighbor, Levon, reenact. And then what Terry did was give the computer this text and then also gave it the exact phonemes, like the symbols for the proper pronunciation for those words. No rules, just actual pronunciations. And then said to the computer, quiz yourself, like, go ahead and try, and then compare what you tried to the correct pronunciation. First recording, de novo learning. And here it is. Wow. Right, so it has no idea what it's doing. Yeah, not even close. No, no. It doesn't sound like a baby either. Like, that just sounds, like, glitched out. It's chaos, right? It's, like, noise effectively. Yeah. But then, as it continued quizzing itself, comparing its output to what it should have said. When I go to my cousins, I play badminton, all that. Slowly, we could actually hear the learning. You could hear it figuring out the difference between vowels and consonants. And then it would start pronouncing small words, you know, oh. And, you know, it only took a couple of days. When we walk home from school, I like to go to my grandma's house. Well, because she gives us candy. And it was acing it. And we eat them sometimes. Oh, we eat them sometimes. Sometimes we sleep over a night there. Sometimes, when I go to go to my cousins, I get to play badminton, all that. But, the really astonishing thing is that when they gave the program new words and new sentences that it had never seen before. He won't stop jumping around in the bathtub. It pronounced those two. He repeats jumping around in the bathtub. But, when he goes to bed, when he finally gets to sleep. It was phenomenal. Sometimes, I get to go to bed at 12.30. Sometimes, but most of the times, I don't. What we didn't appreciate back then was that NetTalk was a little bit of 21st century AI in the 20th century. That this process of learning was the future. And, we're still learning. Are we done? We're done. Thank you so much. Well, okay, but, but, but, like, what, like, what actually happened there? What is it doing? How do you get a machine to learn that? Well, take a baby human. Uh, you know, it's born with this clump of gray stuff in its head, which is really a bunch of neurons that are all connected in like a random messy way. Oh, they are connected? I just imagined a baby brain was like, nothing was connected. It was a blank slate. No. So, when the baby emerges, the neurons are all connected. Uh, they're just not connected in ways that make sense in terms of the world they've just popped into. Um, but then, when it gets some input, like it touches something hot, gets yelled at, gets cuddled, it, it starts to strengthen some of these connections and prune others back. Okay. Until you have this just unbelievably complicated network of connections that can recognize patterns in the world around it and, you know, know that this is a square, or if you poke a cat, you get scratched. That's right. In the brain, you can adapt to your world that you happen to be in by changing the strengths of connections between neurons. And so, basically, Terry and others wanted to create some version of that in a machine. Yeah. That, that, you, you, you hit it. The models we were developing, these neural network models, were based on very simplified versions of brain circuits. Okay. But how did you, how did you do that? Uh. Like, what is, what is going on under the hood here? Uh. Uh. Uh. Uh. What is going on under the hood here that allows it to, to do this? Well, we understand mathematically how they work. We're, and we're making progress now at trying to translate the mathematics into something that humans understand. And so, Latif, here is my best attempt to translate this for us humans. Okay. I mean, so just setting aside all of the technical setup on like, how is it even interpreting the data or, or what are you inputting? Uh, with the help of this guy, Grant Sanderson. Yeah. I, I run a YouTube channel that's named 3Blue1Brown. I often talk about math, but math-adjacent things as well. Great. We're just going to draw like a mental image of what one of these networks looks like. Okay. Let's go. Now, as we all know, these neural nets can do crazily complex things. But for now, we are going to give one a very simple problem. I'm going to draw a couple shapes. What shape is this? A circle. Nice. Can we get a computer to see a circle? How about this? A circle. A very childlike task. Yeah, sure. Now, first things first, to get an image into the computer, we're going to chop it up into a bunch of pixels, like a 10 by 10 grid of them. Okay. And we're going to imagine those pixels, uh, as 100 light bulbs. One light bulb for every pixel and light bulbs that will be on if their corresponding pixel is filled in with ink and off if their pixel is empty. Okay. So, you've got this circle of illuminated bulbs in this grid of bulbs that are off. Okay. I can see it. From there, for reasons that'll make sense in a minute, below that, we're going to add a smaller grid of 10 light bulbs, and then below that, just one bottom bulb. At the top, 100 light bulbs, and then another layer, 10 light bulbs, another layer, one bulb. Exactly. And that final bulb, that is just the end. That is just the answer. The output that, when it turns on, says, "Yes." Circles? Circles, that's right. There's a circle here. Okay. But this last bulb, it's a little bit special. It, uh, it's not like the other bulbs in that it's actually on a dimmer. So, it can also answer, like, maybe a circle, 'cause it could be a square if it's kind of bright, or I'm pretty sure if it's pretty bright, or if it's all the way on, that means this is definitely a circle. As a side note, I, uh, yeah, this feels like quite the challenge where we're torturing the poor audience members here. Probably, like, on their drive, and, and not able to allocate their, you know, like, visual cortex to, like, try to visualize all this. But, uh, setting aside all of the technical terminology. There's one last thing to do. We, we have to wire all of these bulbs together so that electricity can, like, flow from that top grid through that middle grid down to that last bulb, which will hopefully turn it on. So. We call up an electrician. We tell him go and connect every bulb in the top 100 to every bulb in the middle 10, and then go and connect every bulb in the middle 10 to that final bulb. So, literally every bulb is connected to every other bulb, basically. Exactly. So, that electricity can flow down from any bulb that's lit up and kind of cascade through all of them. Uh. And so, the electrician starts pulling the wires, soldering, and they say, "I'm done." But, the thing about this electrician is, they're shit. Like, they just do a terrible job. Uh, some of the wires that they put in are like a strong copper. Others are just twine, so they can't even carry electricity. And so, when this is all said and done, this network we get is kind of. It is kind of like a fresh baby brain with just random neurons clumped together. I see. Got it. And so, when we do send an image of a circle into it, into the machine. Hey, why are you using the microphone again? To record your voice. Lighting up some of the bulbs in that top grid. What shape is this? Uh. The electricity passes down through these random connections from the top to the middle, down to the bottom, and in all likelihood. Iron. Rectangle. It's completely wrong on this. That final bulb might be a little lit up, or half lit up, or just completely off. Okay. Now, when a child gets something wrong. No, what is that? And, like a parent scolds them, that is altering the connections between the neurons in the brain. Strengthening some, pruning others back, right? Right. And that, that is what we want to do with this machine. We want to mess with those wires. The strengths of those connections between the bulbs. Right, right, right. Now, we could just go in there and rewire this thing by hand. Yeah. We could pick out the important bulbs. Yeah. Yeah. because we know which ones are lit up for a circle and direct their current through the middle bulbs to that final bulb. But, you know, that would take just as long as hard coding it. Right. And so instead, we're going to give this thing the chance to learn all this, to learn what the connections should be. So when it gives us that first random wrong answer, there is a 12.2% likelihood of a circle in this image. We're going to say bad robot. There is absolutely a circle in this image. Try again. Okay, I will try again. But then after that first try, instead of us standing there saying yes or no, we are going to set it up to learn all on its own. We're going to step away and let math be its babysitter, be its teacher. And so this is the moment where we have to dive into the math a bit. Uh, okay. It's not that complicated. It's mostly multiplication. All right. Okay, let's go. First of all, these bulbs in the computer, they're really just numbers. One, two, three, four, five. And the wires, you can really just think of them as variables that multiply these numbers. X times two. As they pass through them. Y times 0.3. A good wire multiplies the electricity by five or whatever. A bad one divides it in half or even zeros it out. And that means we can just take this entire array of bulbs and wires and turn it into a giant equation. You know, A times B plus C times D plus E times F. There's some other math strewn in there very artfully and deliberately. But the key here is with a bit of mathematical trickery, this equation can represent the difference between the output it is giving There is a 12.2% likelihood of a circle. and the output we want it to give. There is a 100% likelihood of a circle. And if we do that, if we think, "Hey, I've got this function and I want to, like, find a minimum of that." Like, minimize the difference between your output and the output we want. There's a whole field of math that is just built ready to do exactly this kind of thing. This is what calculus is all about. Like, Newton, if he was rising from the grave, would just be, like, throwing fireworks right now saying, "Hey, I got this. I know how to do this one." So somehow, the calculus tells you, in math equation form, if you're getting close enough, closer to the right answer? Yeah, and now, don't worry, we're not going to go into the calculus other than to say, we walk away, and the calculus becomes the teacher. Okay. So. 12% likelihood. After the first wrong answer, the equation says no. The machine tries again. 25% likelihood. And the equation says closer, and the machine tries again. 77% likelihood. And each time it tries, it messes with the wiring, the weight of the connections between the bulbs. Getting it closer and closer. And closer and closer to right. Exactly. And what happens over time is that that middle grid of 10 bulbs, their connections back to that top grid are getting tweaked in such a way that it's like they're starting to pick up clues. Like, maybe it's getting stronger signals from bulbs that are part of a curve. Or maybe it figures out that the corner bulb can't be on for it to be a circle. And, like, the thing is, we actually don't know. I mean, when people talk about these things being a black box, this is what they mean. It's this middle grid, it's all automated by math, it's picking up something. And we might think. We don't know what the clues are. We just know that they're right. That the clues are, like, they work. Yeah, it's finding some signal that tells it there is a circle-ish thing here. And as it keeps giving answers, and the equation keeps telling it whether it's right or not, or closer, or further away. Eventually, each of those middle bulbs is receiving the right electricity from the right top bulbs to know if these characteristics of a circle are there. And if they are, they pass that along to the final light bulb, which will light up if enough of those characteristics are present. And at that point, yeah, our little network here has learned a lesson. here has learned to recognize this circle which that's actually kind of astonishing that's pretty amazing it is but it's only this one circle and so the the important thing is that if you do this process not with just this one circle circle circles but with tens hundreds thousands of examples circle circle you know big circles little circles messy circles circles drawn by you and me and you have the machine tweak all those different wires for all those different examples you can then take all of that and do one final actually very simple bit of math just average it all together all of the wire strengths you got from all the examples for wire one get averaged down to one value all the wire strengths that you got for wire two get averaged down to one value and if you've done this right you can then send in any of the drawings it's seen before or new drawings it's never seen circles drawn by a two-year-old or a picture of an orange and it will say yes there is a circle there holy cow now that process we just went through uh can recognize way more sophisticated things than just a a shape like cats or dogs and i mean the only real difference in the model is instead of a circle it's a circle and it's a circle and it's a circle instead of these three grids we just used these three layers you know an input a middle and an output you just add more layers of bulbs in the middle these multiple middle layers allow the computer to recognize progressively more complicated components of the picture so like the first layer might just find the edges the second might find textures the third forms the fourth maybe eyeballs because it's like because everything is made up of building blocks of the layer before it yes without crucially without anyone labeling any of those intermediate like it's it's figuring that out itself exactly and then using the same mathematical reinforcement it can tune and tweak to get shit right okay wow well i need a drink after all of this uh to sort of let all this settle in uh great okay like this i'm i wish my kids could learn like this like the way they learn is so physical so emotional it matters who's saying it it matters how they're saying it it matters the tone it matters all these different things like this is so clean and like crazy fast i mean what just took us 10-15 minutes to explain that all happens in seconds so it can learn the circle thing at at basically lightning speed but like a circle recognizing a circle is one thing sure and like now we're talking like actually making like making a you know a sonnet as if shakespeare wrote it that seems like a that seems like a very wide gulf it seems like there's still a lot of place to go for sure and our little alien is going to have to evolve here yeah but in in terms of its its architecture of how it does this it's basically the exact same huh the the only real difference is we're shifting its its focus from to a slightly different skill and we're going to get to that uh you want to predict what i'm going to say next right after a quick break exactly right after a quick break latif simon radio lab so you you asked this question to me before the break like how did this thing evolve from being able to recognize shapes to to uh generate stuff yeah and i posed that very question to grant sanderson okay yeah okay so i would say there's there's many different ideas at play here who again youtuber has thought about this question a lot and i'm a hell of a lot about this stuff and he says the important next step is to realize that yes you could think of what we did with those circles as having the machine recognize them or you could say we were asking the machine to predict the answer we wanted like with the circle example there's there's two things that it could predict circle or not okay so but that so that's so it's not anything meaningfully different it's just like let's just call everything a prediction right but it becomes important when we're talking about the answer to the question we're talking about generative stuff okay like in the case of language predict what word comes next so to explain um going all the way back to the 80s ibm began playing around with these chat bots that you could type to and it would respond hello there how are you today and the way it would do what it was doing was it would take every word that you typed in as your question turn those words into numbers we're not going to go into how because that would take an hour in it of itself but turn those words into numbers send it through this multi-layered set of bulbs But in this case, those bulbs, those layers it's passing through, they haven't been trained to categorize a sentence. Like, we don't want it to say, that was a question. Instead, it has been trained to spit out the word that is most likely to come next, to predict the most likely next word. Just one word. It's not even a word. Also, there's a nuance here between the notion of words and tokens, but excessive nuance. Yeah, but it's like, what is it even basing? Like, how is it predicting that? With a circle, you know it's a circle. We know the right answer. We're giving it the right answer. It's calculating back to that right answer. Right. But like, in a sentence that could go any million number of ways, how can it ever have a right answer to train back to? Well, so what IBM was doing was giving it a bunch of texts, books, transcripts, conversations, feeding that into this machine. And so then the right answer. The right answer was the most likely word to follow the preceding word of words. So it's like, it's just like, here's a giant stack of human talking. And in this giant stack, what's the most likely thing that would have been said next in this exact scenario? Exactly. That's right. And just one brief aside, because it's sort of fun. I think I have this right. That a word is a big. A big, long list of like 13,000 numbers. What? A computer has to turn a word, one word, just like a one word into 13,000 numbers? Yeah. And so like in the way that a pixel value in the circle example was like basically a zero or a one. It's like every word is this list of 13,000 numbers. It's so weird that it like, that that's, that's the simpler version for it. I know. Let me turn it into this like a phone book of numbers. Which is again, like, which, which points to how these things are so not us. Yeah, they're really not us. Not at all. Wow. But they're using us though, right? Like it's our talk that's getting turned into numbers. And it literally does it one word at a time. So after it's written the first word of its response, it just does the whole process over again. It takes all the words in your question, plus the first word it predicts. It sends all that through the network again. And then it just predicts the next word after that. Sends that through those bulbs again. And then the next word after that. Does the whole thing again. And plays the same game over and over and over. And one of the words in its vocabulary is the like, end conversation token. So it like, it has some notion of when to stop. But the act of stopping is itself just one more prediction. It's, it's one more probability in that big list of things that should happen next. And as I said, this is how they were doing it all the way back in the 80s. And I mean, if you interacted with a chat bot, even in like the 20 teens, this is the way they were doing it as well. Really? Do you have any recollection of when you first came in contact with one? Oh, God, um, I feel like it must have been one of those like customer service bots on a website kind of thing. And I'm sure, not just because it's a customer service experience, but because it was an early chat bot experience, it, it wasn't very good. No, no, no, terrible. No, terrible. And a big part of why they were bad. Was. They had difficulty dealing with longer stretches of, of, of, of text. This is Steven Levy. Editor at large at Wired. He's been covering this stuff for. Yeah, yeah, I mean. A long time. I published a book in 1992 called Artificial Life. I was two years old, by the way. Thanks for that. Sorry. Yeah, yeah, thanks. And he says, because it predicted words one at a time and one after the other, the longer the question or the longer the answer, the more likely it was that it would be. The more likely it was to miss or lose the larger meaning. And so eventually predict a word that just doesn't make sense or, or is out of place. Exactly. Huh. And so just to give one very concrete example to illustrate it, like the sentence, what sound does my dog make when I slam the door? It's like. That's so, I can see why that would be confusing. Right. Like you have to somehow know that in that sentence, dog is really the opposite. The operative term here, the important noun, it's not I or door. Right, right, right. And so in 2017. This guy, you know, Oscar Eap, who worked at Google. Set out to solve this dog door problem. He thought that the thing should be able to figure out, oh, this is the most important part of the sentence. This is what I should pay attention to. And now the question becomes like, how the heck does one go about doing that? And what they figured out was the, the problem here is we're giving it. One word at a time and we're having it predict one word at a time. And what we need to be able to do instead is have it somehow process the sentence as a whole so that, you know, something at the end of the sentence can sort of feed back on the weight or meaning it gives to something at the beginning of the sentence. And one way that you can just imagine it doing this is that instead of just making a prediction and giving an answer, you need to take in all the information. Make a prediction, but then just like set that aside because you're going to take in all that information again, and then we're going to send it through again and again and again, each time focusing on a different word in the sentence, generating a different possible prediction before landing on some final prediction, which God willing would be bark. It's like the computer simultaneously lives in the multiverse of that sentence where, where each word in that sentence is the most important. Yeah, and like I, I've, I've, I've looked at this stuff for months and I still don't totally understand exactly how a machine does this, but. Well, I mean, something like that. And also, you know. You can say no, you can tell me. Well, I mean, I, I, in the, in the raw sense. Yeah, that's, that's the, that's the idea. Like the complexity here, you can see it's going through the roof here, like where you're like, oh God, this is so much more computing you need to do. Totally. And, and this was a big barrier for a long time. I mean. And that's why these chatbots were almost as bad in the early 2000s as they were in the 1980s. And this is where we get to the next step in the evolution of our little alien friend here, which, as many evolutionary leaps are, was mostly a hardware upgrade. I mean, if you have been following the news about AI at all, you've probably heard this term GPU. GPUs, components that go into data centers. Or the company. Computer chip maker, NVIDIA. NVIDIA. The most valuable. The most valuable company in history. That makes these things. Its story, of course, wrapped up in the frenzy around the future of artificial intelligence. These things and this company have been at the center of the conflict between China and the U.S. when it comes to export controls. And the idea here is for the U.S. to kind of limit the ability for China to catch up when it comes to AI. And interestingly, what these GPUs, these graphical processing units, were originally designed for was computer games. Video games, things like that. And what they're really good at. What they're really good at is just doing a bunch of different math problems all at once. Exactly. It's just all about multiplying and adding numbers as fast as you can. There's some other things, but like, by and large, like, just do those two things and we're off to the races. And doing these math problems all at once, which is called parallel processing, that's exactly what these learning machines needed to do some version of that super complicated multiverse prediction thing we discussed. Sure, sure, sure. And so, with these GPUs and this new parallelized architecture. That Google named a transformer. All of a sudden, they could get a machine to parse those longer sentences and give at least reasonable answers to more complicated questions. All right. But what really sent these AI chatbots into the stratosphere was a kind of knock-on effect of this parallel processing. Because when you can process everything at the same time in parallel, you can actually train on a lot more material in the same amount of time. And so eventually, they just gave it basically the entire internet. Almost everything we humans have ever said on the internet as its training material and started sending that through this network of light bulbs and wires that was just unimaginably big. Like, to get a sense. In our smaller example with the circle. There's something like a thousand and some odd parameters. Right. A thousand or so of those wires. GPT-3, which was kind of dumb by today's standards, but it came out, had 175 billion parameters. 175 billion? 75 billion things that could be tweaked. Yeah. And many of the ones that we have now, they're trillions of parameters. And as they fed basically all the things we humans have ever said on the internet into this thing. Throwing way more training examples and way more compute than anyone would reasonably think to do. Slowly, they started to notice. That with a sufficiently large amount of data on a sufficiently large model run with sufficiently many cycles of training, these new computers do seemingly intelligent things. Now, a lot of what I just described was written up in a paper called Attention is All You Need. And these findings are really what unlocked these large language models like ChatGPT. And that's all it was really intended for. But. There was a passage in there saying we think this can work for images and video. And indeed. That turns out to work. same basic model of massive parallel processing with tons of input. That could predict the next part of an image or sound. The moment civilization was transformed. And that moment, that realization is really what triggered the explosive proliferation of artificial intelligences, different kinds, practically different species of AIs that we are living amongst today. New artificial intelligence systems. Machines that can teach themselves superhuman skills. Chat. GPT-3. GPT-4. Apple intelligence. Dolly. An app called Lenza. Bard. Bard. It's called Mid-Journey. Text-to-video art generated. It's crazy. Look at this. So I don't know what AI it is they're using. Yes, it feels like an episode of Black Mirror. So it's like, it's like all of these different apps doing all of these different things. And it's like, it's like all of these different apps doing all of these different things in all these different mediums. They're taking in a huge amount of examples. And then they're using fancy math to basically predict the next word, the next pixel, the next note. And from that, it's like generating this whole huge diversity of new stuff. Yeah, basically. And I just as we described doing something that I don't totally understand that that's more holistic than just looking at the thing that happens next. But it is drawing on the examples it's been given to decide what should happen next, which suddenly sounds not so simple. And it does send you into a spiral because it's like, it's like, is, is what I do any different from that? Just spewing out, you know, some iteration of everything else I've seen before this. Yeah. But first of all, you're not, you're not pulling from the whole internet, right? Like you have to depend on just the, the limited things you've experienced or can even maybe remember. And you're like math is also just way sloppier. It's not as accurate. And to that point, and maybe we shouldn't even go here, but there's this one other thing that you can control in these models, which is called the temperature, which is like this final knob you get to tweak on the thing. And so if you have, I think it's, if you have the temperature all the way down, it will give you the most likely thing to come next. If you turn the temperature up a little bit, though, it then is going to pick like the second or third most likely thing to come next. No. So you can control like how precise you want the math. Like you can, you can say, I want it a little stanky. Yeah. Like there's a little bit of randomness in it then that it's then acting upon in what it does next. So maybe you just want the temperature turned up on like every third word so that there's this, uh, almost spontaneous feeling, serenity, serendipitous creation, active creation that, that comes out of this rigid math. Like, it's like something startlingly creative might just be a less right answer. A less right answer. Wow. Yep. And just by doing that, it's going to keep doing stuff that we are going to get increasingly uncomfortable with. Yeah. Like right, right now there is an AI generated song on the billboard country charts. Really? I didn't hear about that. But like, if that's the case, I, I see no way that eventually a fully AI generated film won't hit the box office. Like that, that's just going to happen. But when it happens, it will be only because of all of this math. To me, I think the thing that makes me, it makes me realize is when you see under the hood, what you see is less like something spooky. And ethereal. Yeah. Like there are times when it gets spooky, when like, like there'll be a time, like I'll be listening to like an AI generated podcast and then one of the hosts breathes and I'm like, wait, that's so weird. Like it doesn't even need oxygen. Why is it breathing? And now it's like, oh, because you know that like, that's just the next statistical thing that would come in that sentence is a breath. That to me, that to me is like, it's, it's much less eerie because you can see where it got it from. Right. Um, well, okay. I do have one, uh, one bit more for you because, uh, I don't know, I, I still found myself wondering how it will feel as these things get better and better. And in particular, what it'll feel like in the moments we sit across from it. And it is better than us at something we have spent our lives working on that it is better than us at something we truly love. Yeah. Like many, many people or also all my friends tell me like, wow, you are the first professional gold player be famous because you lost the game. No. So yeah, it's me. And so, uh, I got in touch with this guy. I'm a professional gold player. So a three time European champion. So real quick go, it is an ancient Chinese game considered probably the most complicated board game in the world to teach a computer. This is a very simple game. It is a very simple idea. It is a very simple play because of just how open ended it is. All you really need to know is you are trying to control as much of the board as possible. You go back and forth with your opponent, placing one stone at a time and you control portions of the board or territory by either like cordoning off sections of it or encircling your opponent's stones. It's very simple idea, but it's difficult because with such simple rules, there are just this. Easy number of ways the game can play out. In fact, folks like to say that there are more possible ways for a go game to go than there are atoms in the known universe. Yes. Anyhow, uh, back to Fon. I remember I discovered a goal age like six in my school in Xi'an. And, uh, I feel something, oh, this game I can play and I progress very quick. One year after I learned my school, I'm number one. Three years after that. I'm in the best team in the province. And not long after that. I stopped my school. I only learn goal game. I mean for years. Every day, only, only thing you do is just to play goal game. 12, 12 hours, 12 hours of play. Yeah, it's no joke. Around age 15, he went pro and somewhere along the way, he says he noticed this almost magical quality of the game. Goal for me, it's like mirror. A mirror? Mirror. Yeah. Because when you play, you can see your mind on the board. He says all the choices you make, whether you're aggressive in attack or are patient and waiting, you know, in a sense, how you think stares back up at you. It's like print. Mind print. And your opponent's mind, he says, it's printed there too. So I play with someone. I don't know him. I never talk to him. Talk with him. I play one game. I know him. This is magical. But this mirror of his, well, it was about to get shattered. 2015, then Mr. Hassabis, researcher at Google, sent me an email like, we have some very exciting goal project. Can you go to our office visit? We will show you our project. I tell, yes, okay, why not? And what they showed him was this thing called Elphago. It was a computer that had learned how to play the game. And they asked him, like, will, will you play against this? So I tell him, okay, we can play together because I will win. It's just a program. Just a program. What can you do? You can win with me? Never. It's like zero percent chance to win this game. Zero percent. And why, why were you so confident? Because I know the best program this moment. I can give six stone handicap. Handicap. Handicap. Handicap game. Got it. Got it. Got it. Got it. So how you can possibly make this happen? The technique makes a huge difference. Just a month. It's impossible. And so a month later in this windowless office room, Fon faced off with this computer and it's human stone placing helper in a best of five game match. That first game, all the game, I feel good. I think I will win. But end of game. With just a few stones left to play. Oh, I still play. it. I make some mistake and I lost my first game. Okay. But, you know, he's thinking I was sort of arrogant going into this. I was overly confident. So next game, I will be careful. I will play more seriously. I will win the game. So the next day, next game, he sits down at the board, starts carefully placing his stones. And it's looking good on the board. But inside his head, I feel something really difficult, very difficult, very difficult. Because I like fight, but Afro could don't fight with me. And if I want to take something, Afro gave me very easily. Looking down at the board, he was not able to see his opponent's mind in the way he always had. No. There was no bravery. There was no subterfuge that he could sense. I see Afro. Afro won't do this. Afro won't do that. But why he won't do this? You cannot fund it. You can't. And so, he didn't know how to respond to it. And his mind started to race. Good mood, bad mood. Good mood, what mean? Bad mood, what mean? Good mood, what think my teacher? The good mood, what think my student? Everybody. All my friends. And he realized that with all these emotional pushes and pulls that eventually. I will make mistake. But AlphaGo, no. Never. When you think about this, the confidence is crash. It's crash. All crash. And I lost again. Very, very badly. And I lost again for third, fourth, and the last one. Yeah, damn. But, you know, this experiment is really good for me. Thank you. This is a moment I really see myself. Really? You think AlphaGo taught you to be more. Myself. Yeah. I think this AlphaGo teaching me about that. And why? Because I see myself. So, it's like AlphaGo teach me that our life, we will always lust. Lust. Lust. Lust. Sorry. It's real life. It's our life. I think this is human. This is important for us. I think what he saw in that game as he was losing was kind of what you were saying about seeing under the hood making AI less spooky. Like, he could see it wasn't magic. It was math with no mistakes. Right. And when he saw himself, you know, like not being the perfect Go player in any given moment or in every given moment, like that's what makes him a person. A person who could love something but still lose at it. Maybe feel bad about that and then use that feeling to figure out what to do next. Today, I'm teaching the Go in China with a student. Why are you teaching Go? The computer will always win. Yes, yes. But be careful. Because I think all you experiment to learn is still useful. So, don't worry. It will be coming. You can do nothing. Accept it and just learn. Yeah, I get that. Before we wrap this thing up, I wanted to put all of this in front of someone. And not an AI person, but somebody with a really wide scope on technology and history. And so I went to this guy. Tom Mullaney. I'm professor of modern Chinese history at Stanford University. I worked with him years back on a story about typing in Chinese. And he said, and he said, and he said, he's just one of the most thoughtful and informed people I know. Yeah, that means a lot. That means a lot to me. So, how would you respond? Well, I mean, everyday life is at its core a study of this awful, amazing, horrifying, never-ending surprise of what it means to be born and live and die as a human. And even if, at the end of the day, an AI is orders of magnitude smarter. AI, just by definition, cannot suffer and rejoice and live and die in quite the same way that humans can, in the same way that we cannot live and die and suffer and comprehend and feel the way an octopus can. I mean, the only thing an AGI will be able to do is contemplate, my goodness, what does it mean to be, an AI? And so I am not worried at all about what AI means with regard to meaning human identity, what it means to be human or any of that. Well, that was very beautiful. And while I love that I'm still like, but this is going to mess everything up so badly. I don't know. Oh, no. No, I agree. Okay, go. It's going to, no, it's, I mean, this is going to get weird down to the fabric. But fast forward this, you know, 20, 30 years, if we're still around at a sort of climate change level, when another future human is sitting in this fabric altered world, it will still be a group of humans rejoicing, suffering. Like it will still be that condition, and so it's kind of a it's kind of a liberal for me, it's like a little bit of a liberatory time it's it's it's a great maybe we'd get to free up a little bit more space to get back to work thinking about how to be human because we have not we've not even come close to solving that issue. © transcript Emily Beynon Special thanks to Stephanie Yin and the New York Institute of Go for teaching us the game. To Mark, Daria, and Levon. To Barbara Svenich. And, of course, thank you to Grant Sanderson for his unending patience explaining the math of neural nets to us. Grant is kind of like your favorite math nerd's favorite math nerd. His YouTube channel is 3Blue1Brown. Check it out. This story was reported and produced by Simon Adler with original scoring and sound design by Simon Adler. Which brings me to the last unsavory thing I have to say, which is goodbye to Simon Adler, who happens to be one of our best reporter producers here at the show and also a friend. He's going off to, among other things, pursue his music career. And this was his last episode on staff with us. Chances are, if you list out your favorite episodes from the last 11 years at the show, more than a couple will be his. Could be some of the tech stories he did. He did stories about drones in Ukraine, about content moderation on Facebook. Could be some of the international stories he did. He did a lot of stories he did. He reported about the hunt for an endangered rhino in Namibia. He did a story about a species of raccoon in the Caribbean islands of Guadalupe. He did a lot of stories about democracy as well. Covered a town, Seneca, Nebraska, that voted itself out of existence. He did a story back in 2017 about a New York City council race where the campaign manager was a little-known guy named Zoran Mamdani. Besides being a killer reporter, not to mention compounding, composer and interviewer, Simon has also spent so many hours coaching an entire generation and staffers and interns. He's so generous with his expertise and his time. Really someone who makes everyone around him better. Anyway, we have been so lucky to have him as part of our nerdy band for 11 years. Check out his band, Windstar Enterprises, on Instagram. That's Simon and another fellow former radio lover. Alex Overington. We just, we already miss you, Simon. And good luck out there. Oh, you want me to say this? Oh, that's fun. Hi, I'm Cordelia and I'm from New York City. And here are the staff credits. Radiolab is hosted by Lulu Miller and Latif Nasser. Soren Wheeler is our executive editor. Sarah Sandbach is our Executive Director. Our Managing Editor is Pat Walters. Dylan Keefe is our Director of Sound Design. Our staff includes Simon Adler, Jeremy Bloom, W. Harry Fortuna, David Gable, Maria Paz Gutierrez, Sindhu Nyanasambandan, Matt Kielty, Mona Madgavkar, Annie McEwen, Alex Neeson, Sara Khari, Anissa Vitsa, Arianne Wack, Molly Webster, and Jessica Young. With help from Rebecca Rand, our fact-checkers are Diane Kelly, Emily Krieger, Ana Pujol-Mazzini, and Natalie Middleton. Leadership support for Radiolab Science Programming is provided by the Simons Foundation and the John Templeton Foundation. Foundational support for Radiolab was provided by the Alfred P. Sloan Foundation. For more information, visit www.fema.gov.