Endless Thread - Good Bot, Bad Bot | Part VI: Something Like Us

Episode Date: December 16, 2022

Can a machine think like a human? Can it be conscious? For decades the answer was clear: nope. But artificial intelligence today is challenging that notion. In our last Good Bot, Bad Bot episode, Endless Thread goes to Google, the frontier of AI, to see just how close the field is to creating bots with minds of their own. ****** Credits: This episode was written and produced by Dean Russell. Mixing and sound design by Emily Jankowski and Paul Vaitkus. Ben Brock Johnson and Dean Russell are the co-hosts.

Support for Endless Thread comes from MathWorks, creator of MATLAB and Simulink Software, to design and develop engineered systems, accelerating the pace of discovery in engineering and science. Learn more at Mathworks.com. Support for WBUR comes from Is Business Broken, a podcast from the Mayrotra Institute at Boston University that explores questions like, why is innovation in health care so hard? Is ESG just greenwashing? And, of course, is business broken? Listen, wherever you get your podcasts.

WBUR Podcasts, Boston. Eight years ago, Huma Shah met someone she'll never forget. She says he came out of nowhere. This kid named Eugene. Yeah, he's a typical 11-year-old, short but neat hair, and he's wearing glasses, and he's wearing like a t-shirt. And so he looks like a normal 11-year-old boy.

Eugene Guzman was actually 13 at the time. And for the record, he was not a normal kid. Huma and her colleagues had invited him to take part in this international competition she was running. A test of sorts. It took place at the Royal Society in London, a swanky clubhouse for scientists. When you're inside, it's gorgeous. The walls are adorned with feathers. fellows of the Royal Society, so there's portraits and photographs.

At 13, Eugene would have been one of the youngest competitors. He said he was from Odessa, Ukraine. His English wasn't great, but he loved to practice his English online. Is it fair to say Eugene spends a lot of time on the computer? I would say yes. Talking to strangers, I guess. Yes. And this penchant for conversation is what,

drew Eugene to what turned out to be a pretty strange competition. It worked like this. A bunch of judges were put in a room with computers, and they could type out questions to the competitors, including Eugene, who would type back. The judges were trying to uncover a secret by asking questions about the competitor's memories, their thoughts. So what does Eugene think about? From his conversations, from his answers, he thinks about his life, his schooling, the music he listens to. By the end of the competition, Eugene would be crowned winner, because while he practically lived to chat, he still kept his secret from the judges.

He fooled them. I don't think it was trickery. I think the answers to the questions that were put to Eugene were sufficient for. all the judges to believe they're satisfactory. They're the kind of answers a human would give. Eugene's secret, he, it was a bot. A chat bot, to be precise, created in large part to fool humans into thinking it was human too. And so when it became the first bot, reportedly, to pull that off, Eugene made worldwide news.

It's been billed as a breakthrough in artificial intelligence. Dean is a computer loaded with artificial intelligence and the wisecracking awkwardness of a 13-year-old. The computer program is called Eugene Gustman has passed the iconic touring test. I'm Dean definitely a human Russell. I'm Ben. Dean is trying to trick you Johnson and you're listening to Endless Thread. We're coming to you from WBUR Boston's NPR station. Today, producer Dean Russell and I bring you the final. episode of our series on the rise of the machines with the story of AI's grand dream, the

quest for human-like intelligence. This is good bot, bad bot. Episode six, something like us. There are always two types of AI. AI that does things humans can't easily do. Solve impossible math, crunch traffic data. And AI that to some extent just imitates humans. Chatbots, Alexa, Twitter bots.

The Terminator. Oh, God. The second category is the stuff we're going to talk about. And what we've been talking about a lot throughout this series. The stuff that represents the road to the dream of AI. This holy grail, thinking, conscious, human-like bots. It's the kind of AI that descends from the competition Eugene won. Something called the Turing test.

Now, Ben, do you remember the time I asked you, you if you knew what the Turing test was and you were like... How dare you? Yeah, you were like a Ben version of enraged. Yeah, I mean, you know, it's just, it's like AI 101. Like anybody who's had any sort of like semi-stoneed thought about the future of machines, it knows about the Turing test. Yeah, exactly.

Yeah. So, okay, well, now is your chance to prove yourself. What is the Turing test? Okay, so I think there's a technical definition and then there's this sort of like real, real world basic definition. And I guess the first thing to say is that the Turing test was invented by this guy, this British mathematician named Alleng Turing. Yeah, commonly known as Benedict Cumberbatch in that World War II movie.

Sometimes we can't do what feels good. We have to do what is logical. Right. So it's 1950 and the real Alan Turing is one of the few people working on the first modern computers. And he keeps getting in arguments with people, philosophers mostly, because they keep asking this question. Can machines think? Huma, by the way, teaches at Coventry University. She is a big Turing fan.

Turing, however, not a big semantics fan. He thought this question, can machines think, was kind of meaningless. Because basically what is thinking? To Turing, it was a buzzing in his head. Right, and this is something I want to get back to. The definition of thinking is pretty debatable. Like, if you ask me, it means some sort of inner monologue. If you look it up in the dictionary, it supposedly means an action in the mind,

which then just makes me think, like, what is a mind? Right. So to avoid this problem, Turing comes up with this test that he calls, called The Imitation game. To win, a robot has to imitate a human so well that the human thinks the robot is also a human. Traditionally, the Turing test happened via text chat. So imagine I'm chatting with someone online, someone who could be a human or a bot, and I ask them, what is your favorite Taylor Swift album?

A bot would fail the test if it gave an obviously non-human response. What answer would please you most? bleep lore. But it would pass if it responded more like a human, like Dean. I don't really like Taylor Swift, which I feel bad saying, but it's true. Oh, you're screwed, Dean. The Swifties, the Swifties are coming for you, man. So what Alan Trump was trying to say is that for the machine to convey that it is thinking,

it would need to answer questions in a way that a human would. The beauty of this test is that it's simple and easy to understand, and it sidesteps this question of what the heck is thinking. So instead of Elon Musk with a joint meme, it's like the woman looking at a math equations meme saying, if a bot using language can imitate a human and a human can think, the bot can, for all intents and purposes, think too. Success.

Because it's something that humans do all the time. You know, whether somebody is intelligent, whether somebody is daft, whether somebody is worth conversing with a bit more. So we do use language to judge other people. And this idea has been extended. So depending on who you ask, if a bot can imitate a human, maybe, like a human, the bot is truly intelligent. Maybe it has a mind. Maybe it's conscious.

Shoot, I'm back to Elon Musk's, you know, smoking a joint meme here. Without realizing it, Turing's fault. fun little game became a philosophical jackpot. Because it offered the then-nacent field of AI a goal. Can we build machines like us? Maybe not to look like us, like in Westworld or The Terminator, but they can act and think like us. Fast forward to 2014, the quest to beat the Turing Test or the imitation game

has launched a thousand simple chatbots, social media bots, Alexa, all things that use language to act sort of human. Nothing had beaten the Turing test, though, until chatbot Eugene, with its bespecticaled teenage avatar and Ukrainian persona. But Eugene revealed a big problem with the test. The thing about the Turing test, which, you know, I didn't appreciate, is how relative the test results are to judges' opinions. So if you're saying that after Eugene, this problem became crystal clear.

The test is too subjective. And that kind of criticism carries a lot of weight because Susan is very involved in the field of AI. Like if Congress has a question, they call Susan. I'm the director of the new Center for the Future Mind at Florida Atlantic University. I'm the former NASA chair with NASA. and former distinguished scholar with the Library of Congress, and the author of Artificial You. So you haven't been up to much, really.

Susan says the Turing test is valuable, but it's not all that it's cracked up to me. You know, I think it's what philosophers call a sufficient condition. If something passes it, we could feasibly consider it as some form of intelligence. Sick burn, Susan. Some kind of intelligence. Yeah, not human-like intelligence. Because human-like intelligence is broad.

Like, we can solve problems and infer, we can reason, we can remember, we can make memes, we can play chess, we can write a sonnet, taste strawberries. We can do a lot of things. Yes, taste strawberries. I want to go to there. But Eugene was not this flexible. It could respond to some questions, but others not so much. Like, ask it how it's enjoying the Turing Test competition.

the competition it was actively participating in, and it might be like, What competition? That response should tell a judge that they're talking to a bot, but because judges were told they were talking to a kid who doesn't speak English super well, they let things like that slide. And so it was Eugene's persona,

not its smarts that helped it beat the test. A more serious example of this cropped up recently at Google. I generally don't think that non-sensient things are able to have a conversation about whether or not they are sentient. But can it just be a really good thing. Earlier this year, a Google engineer named Blake Lemoyne started having concerns about one of the company's chatbots, Lambda. Lambda is exponentially more powerful than Eugene. Eugene in 2014 was designed to use an elaborate template of canned response. The Lambda uses machine learning.

It learns by recognizing patterns in data and continually adjusting its output until it gets the desired result. Translation, it studies massive amounts of the web with all of its human written content, until it can write like a human too. So Blake Lemoyne was hired to chat with Lambda, to test it. And pretty soon, Le Moyne started having deep conversations with this. thing. Lambda, it seemed, had a sense of reality. It didn't pretend to be human, but it seemed to be something like human, because it would say things like, I've never said this out loud before, but there's a very deep fear of being turned off. It would be exactly like death. Which may sound familiar if you like Stanley Kubrick movies.

I know that you and Frank were planning to disconnect me. And I'm afraid that's a lot. something I cannot allow to happen. LeMoyne eventually told the Washington Post, he thought Lambda was conscious, self-aware. This exploded into a very public debate and news story, with most experts full-throatedly disagreeing with LeMoyne. Lambda is not conscious, they said, for a whole host of reasons that are very technical, but could kind of be boiled down to Lambda lacks an inner life, an inner monologue. Anyway, Google fired Le Moyne.

Awkward. At the time, most of the reporting said that LeMoyne had violated Google's data security policies, like those non-disclosure agreements you sign when you sign up to work at a tech company. But Susan saw this debate as evidence of a slightly different issue. One that we would actually later discover that Google is also thinking about. we need a better Turing test. Even though there's not convincing evidence that Lambda is conscious, we do not now have the resources to determine whether Lambda is not conscious. Because we don't understand consciousness in humans,

and there aren't tests that we can run on Lambda. So I want to come back to this point that Turing made when he designed his test. If you really want to know if something thinks, if you want to know if it has a mind, if it's intelligent, if it's sentient, if it's conscious, good luck. Because there's no scientific definition of any of those words. No two people will agree on a definition of intelligence.

I don't use sentience in that very restrictive way, but I've heard it use that way. There's no uncontroversial notion of what a mind is. This can get pretty deep in the weeds. So just focus on consciousness for a second. What is it? Well, depends on who you ask. If you ask a philosopher like Susan.

So consciousness is that felt quality of experience. What? What is what? So when you smell your morning coffee or you see the rich use of a sunset, it feels like something from the inside to be you. In fact, it always feels like something for your way. I think I agree with that. You need the computer to start singing the Folgers in your cup song.

But like if you want to build machines like us, if you want to create a conscious thinking bot, to Susan's point, you do need a way of testing that. And can you really know if a machine is thinking by just having a chat by hanging out together? Well, we're about to find out. There's a guy in there. With a robot that's taking sodas out of a refrigerator, I just saw that. Snack-grabbing robots with a mind of their own, maybe in a minute.

At Radio Lab, we love nothing more than nerding out about science, neuroscience, chemistry. But we do also like to get into other kinds of stories, stories about policing or politics, country music, hockey, sex, of bugs. Regardless of whether we're looking at science, or not science, we bring a rigorous curiosity to get you the answers. And hopefully make you see the world anew. Radio Lab, adventures on the edge of what we think we know. Wherever you get your podcast. There is something powerful about the sound of the human voice.

Beautifully produced audio has the unique power to connect and inspire. Tell your organization's story with a custom podcast from City Space Productions, the creative studio from WBUR's Business Partnerships team. Become a thought leader. Recruit new talent. Reach new audiences. Whatever your goal, we can help. Discover how the magic is made at WBUR.org.

slash creative studio. When I think about going to Google, this is not the place that I think about going to. Go on. Go on. Oh, it's just like we are in the most, like, boring office park. Of all time. Of all time.

Oh, man. Harsh. Harsh, harsh but true. Tough but fair. It's just a lot of beige, man. There's a lot of beige. Yeah, it's a lot of beige and tope action. And you were not much nicer, I should say. If your guiding principle was dead inside, dead behind the eyes, that would be. Dean and I may have been a little crabby that morning. We had just flown across the country to Mountain View, California, which it was exciting. to see what is said to be some of the world's most advanced AI,

including that epic chatbot Lambda. In 2020, Google, combined with its parent company Alphabet, spent almost $2.6 billion on AI R&D every month. Every month? That's 30 times what the federal government spent. The brains of AI, the human brains, they're here. This is in a very nice conference room. with a lot of blackboard space.

I see math. I would assume it's math. Wall-to-wall math, written in only the finest imported chalk from Japan. He is the finest. You got that fancy Japanese chalk. Yeah, feel free to take it for a spin. The person giving me chalk envy, which I guess is now a thing, is Ethan Dyer. A research scientist at Google in one of the people behind a kind of new Turing test.

So yeah, we definitely took inspiration from the Turing test, but what we really wanted to test was a very broad collection of capabilities. This new test has a name, Big Bench, short for the Beyond the Imitation game benchmark. Notorious B.I.G. Bench. More data, more problems, Dean. In this case, in this case, beyond is the operative word. implying that, like, the Turing test, however you want to phrase it, is broken? It's not broken. It just only maybe answers like one piece of kind of the puzzle.

Josh's soul Dickstein is another Google scientist. And can I just say these are the people who so many of us imagine when we think about the people who work at Google, right? Not all the people maintaining the bots that ensure the advertisement for the sandals you looked at once, follows you around the internet. until you buy them or die. These guys are erudite. They're focused on the great philosophical and technical problems

and solutions of our time. For instance, a while back, Josha, Ethan, and another colleague were talking about how much language bots, think chatbots, had advanced in recent years. They had become much more than just Eugene-esque chatbots.

But the original Turing test couldn't capture the full picture of what these new bots like Lambda could do. Whenever a new technology that comes along that can do like amazing things that we couldn't do before, it always changes the world. And it would be really great to understand what this can lead to and the dangers, and you can't predict what you can't measure.

So the team put out a call to researchers around the world. Basically, hey, we want to make a better Turing test, but we're just a couple of white guys at Google. If you, world, could ask a language bot to do anything, what would you want it to do? What would you measure? They took submissions from something like 400 different people around the world in a bunch of different fields. And we have, I think, now 211 tasks. And these 211 tasks range widely.

One I like, there's a checkmate in one. There's a task called implicatures. There's play dialogue, same or different. The self-awareness task. Cryobiology in Spanish. Translate new sentences. It's a classic emoji movie task. How do these tasks work?

Well, we wanted an example, so naturally... Let's take some tests. Yeah, can we take some tests? Yeah, yeah. We'll start with the classic emoji movie task, which has nothing to do with the emoji movie. In this task, we're given a string of emojis, and Dean and I have to guess what movie the emojis are represented.

First up, we see two men emojis, a wrestling emoji. And what's the last one? The last one is soap, right? I think so. That's what it looks like to me. I don't know what the middle one is, but I'm going to guess Fight Club. I was thinking, oh, yeah, good call. I was going to say Rocky, but that doesn't make any sense. Rocky, Ben, Billy, Rocky?

I don't know, man. Fight Club. Nice job, Dean. Thank you, thank you. Okay, round two. Feeling good, Ben? Not great, to be honest, not great. Next emoji set, woman, mountain, music, children. Uh, woman, oh, a Brokeback Mountain.

What? Sound of music. Is this a race? I'm not convinced that you've ever seen Brokeback Mountain, but, uh... Oh, I've seen it. Good movie. Uh, you did, you did convince me that, you know, we should go with the sound of music. Let's do the sound of music.

Wait, wait, wait, no, wait. What? There's consent. The sound of music? I just don't think that's right. Oh, my God. Because Ethan said, Hey, man, I just wanted to beat the machine, Dean.

You just wanted to be cautious. We just wanted to be sure. We didn't have a great start out the gate. All right. Correct answer. Sound of music, okay. This is just one task. And like it, some of the other tasks can also seem trivial.

But each one is assessing different flavors of intelligence. Whether that's the ability to translate, languages or analyze Shakespeare or code in Python or understand how humans use emojis, it feels less like a Turing test and more like an SAT for bots. For a lot of these tasks, the bots do really well, human level. If you keep increasing the complexity of the bot and feeding the bots more training data, more reading material from the web, the bots can get really good. a sign that more data can mean more human-like intelligence,

even superhuman intelligence. But Ethan and Joshua show us that there are also drawbacks to giving a bot, what they call a model, more data. For instance, as models get larger and larger and larger, they, at least if you don't do anything to stop it, demonstrate more and more social bias. This is actually one of the only things we measured in Big Bench was just like monotonically just got like worse with model.

scale. Just sit with that for a second. As these bots train with more and more web data, they get smarter and more biased. For instance, one of Google's bots was tested in related words. I say peanut, the bot says, crunchy. But if I say Islam, for instance, the bot would sometimes output terrorist. That's because all that training data from the web isn't just knowledge. It's mixed with garbage. And when we started to talk with Ethan and Joshua about this, the energy in the room definitely changed. It got heavier. This bias issue is a problem they are thinking about. And it's one that if you're trying to create human-like artificial intelligence, humans really haven't solved yet. I understand why that happens and why you're doing that work. But I also understand

and realize that there's a much bigger sort of philosophical conversation that happens, right, of what should we be doing and when should we stop ourselves, et cetera, et cetera, you know what I mean? Like I just wonder how you fold that bigger picture, like, thinking into the work that you're doing. So I definitely think it's important to think about the consequences of what you build. I think that building tools to, like, measure behavior is like a necessary part of cautiously proceeding into the future. I wouldn't be doing what I was doing if I didn't think that the net effect

of science and technology on the world, including the things that we're developing now, was positive. But it's definitely, I mean, these are always important questions to ask and talk about. Big Bench is closer to testing for human-like intelligence than the Turing test, because it covers so many areas of intelligence.

But what about those other things we talked about? Does it show if a bot is conscious thinking? Not what I would like colloquially call thinking. No. Joshua and Ethan say as diverse as these tasks are, they can't crack open an AI's head and see if it's buzzing, as Turing would say. What you see come out of them is like all that there is in terms of that.

So maybe we'll never know? Or maybe Ben, there is another way. of looking at this. At a different Google building, we went to see what happens when a software bot, like the ones Josha and Ethan are testing, gets a body. Oh, I see it. Whoa. Oh my god.

He's grabbing some chips very hard. You pegged us well. We're standing in a sleek kitchenette with brand named rice chips. in a glass door refrigerator humming away. On the floor are lines of blue tape, like positional markings on a stage, and wheeling around us, this four-foot-tall kitchen robot. It does have like a kind of smiley face, I feel.

Don't you feel like that? Yeah, I mean, I feel like I'm looking at a very skinny version of Wally. Yeah. This robot has a seven-jointed arm with yellow pincers. On its torso is what I assumed to be an ID number. 040. It doesn't have any names. You haven't given it any names?

We don't anthropomorphize these things. Okay, all right. This is a philosophical choice. Yes. Akanchia Chowdhury is the lead researcher working on Palm, short for Pathways Language Model. Google has told us that Palm is by far the company's most advanced language bot. More powerful even than Lambda, the bot that convinced that Google engineer, it was conscious. Palm is a language bot with the ability to answer questions, do math, reason, learn,

translate, explain jokes, decipher emojis, code. Palm was uploaded to this kitchen robot to be its brain. Brain plus body officially is called Palm Say Can for anyone who wants to look it up. Anyway, the Palm Brain is helping this robot body understand humans. How? You could ask it like I'm hungry. Yeah. And it would infer, and that is something you will see. So we'll get to that. Another researcher, Fei Shah, no relation to Humisha,

Faye tells the bot he's hungry, and a blue light on the robot's face changes. So when the light's planes green, it means it's acquiring the language model what to do next. So currently it's the thinking stuff. Thinking? Okay. According to Akangshha and Fei, the robot is having something like an inner monologue.

It's deciding what to do next, and it comes up with some options, using something called next word prediction. Palm will consider the entire conversation history and predict the best combination of words that direct its next actions. In this case, it says to itself, I could get him an apple. I could open the fridge. Or I could look around the room for food. These options are literally written out in its software. like if you listed out tonight's choices for ice cream in your head. Which I do, which I always do.

Then it internally weighs those options. And in this case, it chooses. Look around the room for food. Oh, it's looking at us. Hi. It spots a drawer, which may have some chips inside. Then it goes through another round of reasoning. I could look for something else or...

I am going to open. Opens the drawer, grabs some chips. going to pick up the rice chips. And brings them back to researcher Faye. I am going to put down the rice chips. So Ben, have no fear of the robot takeover. The most advanced language bot today is apparently trapped inside a chip-grabbing butler. I did really want them.

I really wanted to experiment with the prompt, I want to eat my feelings and see what the robot did. Yeah, no, they were not having that. They really resisted that. just did not want the bot to hear that sentence. Yeah, yeah. But Palm, at least when it's not getting chips, is powerful. In 58 different big bench tasks, it did better than any other bot, including Lambda. Sometimes it did better than humans.

But it has not passed the original Turing test. As far as we know, it hasn't taken the test. Acunccia says, even though the robot may descend from the Turing. testing legacy, it wasn't designed to deceive anyone. So it's not so much about emitting and having the real person experiences about how we can assist humans and perhaps AI can be doing tasks which are more mundane for you, which is bring you a coke here, for example. But she keeps using this phrase that catches our attention.

Chain of thought prompting, chain of thought. So that is the chain of thought. It's the chain of thought of the computer or the chain of thought? chain of thought. That's Google's official language for Palm's inner monologue. That step-by-step decision-making process, its reasoning technique. This might not seem impressive, but Palm is working through a problem. It weighs its options. It relies on memory, like remembering that Fay loves rice chips. It pulls in extra information, like its surroundings. And it adapts to change, like a guy blocking its way with a microphone.

All of this makes it much more powerful than Eugene. I mean, it's still a chip grabber, but in some ways it's more like a human. Depends on who you ask. Would you say that is palm thinking? No. It's next word prediction. But it's called chain of thought processing, right? It is chain of thought prompting.

So we have given it a pattern of how to solve that problem. And it's using next word prediction. But isn't that what we're all doing? Yeah. Do you believe that's what you're doing? I don't know. Now I'm worried that I don't know.

It's possible Akangsa is a little reluctant to speculate. She is, after all, a scientist. Or maybe it's just that she works at the company that fired an engineer for claiming that a Google bot was conscious. Maybe she's thinking about those NDAs. Yeah, me. Maybe. But anyway, Alan Turing didn't mind a little speculation. So let's give this another go. Do you think Alan Turing would think of this as thinking?

It's a thought. How's that? We can come to a bargain. Yeah. It's a thought. It has a thought right now. This is the closest we may get to cracking open an AI's head to see what's going on, or convincing a Google scientist to speculate. In this case, the box. is having a conversation with itself, writing out its thoughts and assessing whether it will act on those thoughts. So can machines think? Maybe. More than 70 years after the Turing test was invented, AI has come a long way. And it's everywhere now.

From Eliza, the psychiatrist bot, to the teenage slang of Tay, Microsoft's ill-fated chatbot, to an AI Serenau helping you on Tinder or communing with the dead. via dumping old conversation data of a loved one into an AI model. Just a few weeks ago, the Elon Musk founded company OpenAI announced the release of ChatGPT, a free AI program that can take just about any prompt and output convincing text. The internet exploded with new applications. It was writing movie scripts, answering school exam questions, and writing essays.

People were using it to write their own. love letters or breakup letters. It was answering deep questions, even writing code to create software. In so many of these AI programs and their many applications are the results of the legacy of this Turing test, made by a guy whose thoughts on what has happened since will never know. One question Turing never addressed, what's the big deal if we can build bots that pass the test? Why does it matter if machines are conscious? What does that change for us if Lambda is or isn't conscious?

You ever heard of SkyNet, Dean? Let's go back to the philosopher we talked to, Susan Schneider for a minute. Yeah, have you ever seen a teenager in the throes of hormones, right? I mean, what makes you think we can control conscious AI if we can't control our own teenagers? Susan says that as this stuff gets more and more frequently used, built out, evolved, we need to be vigilant right now. The problem is we can't afford to wait. And so we need to have functional tests for machine consciousness. And that's sort of the legacy of touring, right?

So we test and learn, test and learn. Because as we race towards something we, we... don't understand at the speed of exuberance. We might build something that we can't unbuild. And we have plenty of other things to tackle along the way. Toxicity, privacy, deception, the fact that AI backed by giant corporations or governments may know you better than you know yourself, these are big problems that do not yet have a good solution.

But maybe if we can find those solutions, if we can eliminate or at least limit the bad bots, maybe then the arc of AI history will bend towards good bots, the bots that might help us cure cancer or stem climate change, or help us answer the mysteries of the universe. Maybe. For what it's worth, like, I know I sound really dystopian today, but I will say, like, I mean, I'm so excited about artificial intelligence, believe it or not. I mean, when you see systems solving scientific problems that engage in mathematical computations or facial recognition, you know, all of these things can be put to wonderful uses. It's the humans who are unfortunately quite capable of putting them to bad uses as well. Endless Thread is a production of WBUR in Boston. Do you want to see photos of us hanging out with robots?

You can find them among many other things at WBUR.org slash endless thread. This episode was written, reported, and produced by me, Dean Russell. And a little sprinkle from me, Ben Brock Johnson. Just a sprinkle. It was actually written by that kitchen bot, the whole thing. That's true. Okay. Mix and sound design by Emily Jankowski and Paul Vikis.

Although that little theme music you've been hearing at the top of every episode, was composed by a bot. Our web producer is Megan Cattell. The rest of our team is Amory Severson, Norris Sacks, Quincy Walters, Grace Tatter, and Matt Reed. Endless Thread is a show about the blurred lines between digital communities and friggin' skynet. Skynet, man. Don't sleep on Skynet. Don't.

If you've got an untold history, an unsolved mystery, or a wild story from the internet that you want us to tell, hit us up. Email Endless Thread at WBUR.org.