Bản chép lời đầy đủ của Week in Tech: Hungry Hungry Hippocampus

avoid Alzheimer's. Taxi drive can show you how interesting? 00:06:03 Speaker 2: Say more so? 00:06:05 Speaker 1: Basically, mass General Brigham Hospital in Boston ran a study about the rate of Alzheimer's in various populations. Taxi drivers and ambulance drivers had up to four times lower rate of Alzheimer's than the general population. And apparently this actually made sense to researches because the part of the brain that does real time spatial processing and decision making, the hippocampus, as I learned is called, is also one of the first parts of atrophies when you get Alzheimer's. 00:06:34 Speaker 2: That's so interesting. So people who have made this their job essentially are people who are kind of saving themselves unknowingly. 00:06:44 Speaker 1: One hundred percent. One of the most interesting things I thought was that London bus drivers don't get the same benefit as London taxi drivers because they follow a preset route. 00:06:54 Speaker 2: I was going to say that one of my favorite things is just the old facts about how much a London taxi driver has to know people. What is it called the why. 00:07:01 Speaker 1: The knowledge they have when when they when they die in the autopsy, they have larger other people. But you know who else doesn't get these benefits? Yes, come on New Yorkers. People who use Google Maps or ways or Apple Maps and what do we call that? We call that automation bias. 00:07:18 Speaker 2: Correct, that's our fave. 00:07:20 Speaker 1: So I just found this really interesting. Here. Here is the technology Google Maps that I use all the time every day, and I absolutely adore it. And I literally every time I go on vacation, I think, how the hell would I have had a good time if I didn't literally literally divorce. Yeah, I would literally go to some like the concierge or whatever, or the some random person who worked in the hotel say can you give me a physical map and tell me how I have a good time? I'd be like, I would never travel. 00:07:45 Speaker 2: So, but isn't it what we always talk about, which is the sort of the double edge totally? 00:07:50 Speaker 1: And but the idea that using this scene that I love every day is you know, shrinking my hippocampus I find really quite chilling. 00:07:57 Speaker 2: I mean, I think about this in terms of my hands and my Sum's true, Like I just think all the time about just burning nerves. 00:08:04 Speaker 1: Well, the devil has no time to make use of your sums because they are never idle. 00:08:09 Speaker 2: They're definitely not idle. I think you have one more thing. 00:08:13 Speaker 1: I do have one more thing. And this, don't worry. We're not going to talk about that length thing I talked to somebody else about. But this was basically, you know, there are these new open AI models, one that was released in September last year, and three right before the holiday on the twelfth day of ship mess as the execs called it, but one was being red teamed. Do you know what

red teaming is. It's when you try and make a technological product to break its own rules or do unsafe things. So, oh, one was being red teamed, and surprise, surprise, in a very interesting way, it started trying to deceive its human counterpart. And that's exactly what we're going to talk about on our next segment, which is Tech Support. Every week we'll do a segment called tech Support when we talk to true experts and reporters who can go far deeper behind the headlines than you and I can, to basically help us sort the signal from the noise. And I genuinely don't think there's anyone better to talk to than the team at four or four Media. These are reporters who were formerly basically the team at Vice's motherboard, and now they've started their own collective, and they are the people who the tech world follows most closely to find out what's really going on in all corners of the digital world. 00:09:30 Speaker 2: Yeah, and this week we're excited to share this conversation we had with Jason Kebler, who's one of the co founders at four h four, who will be filling us in on the study that the initial excitement around the three release meant did not get enough attention, which revealed that cutting edge AI systems, including both open AI's one and Anthropics Claude three point five, sonet whatever the way they name this stuff is so self agrandizing, have been shown to have a shocking t tendency to deceive. 00:10:01 Speaker 1: Shocking tendency to deceive. Indeed, So we talked to Jason right before the holidays when this news just came out, and we. 00:10:10 Speaker 2: Jump right in, Jason, what's popping. 00:10:15 Speaker 4: So there's a new paper out by Apollo Research, which is this group that's funded by various governments as well as various artificial intelligence companies like Google, DeepMind, Inthropic, open Ai, et cetera. And it's a new paper about what's called AI alignment, which is the idea that is an artificial intelligence tool doing what the human asked or is it doing something else. I think that one of the really scary things about artificial intelligence is that they are largely black boxes. We're not totally sure how they work, and this can lead to all sorts of very interesting outcomes. You know, in this case, the researchers at Apollo Research Group were able to find that these sort of cutting edge large language models, which you know, one of the biggest ones is open ais oh one. 00:11:15 Speaker 1: They finally gave it immigrant visa status. 00:11:19 Speaker 4: So they found that in some cases, these sort of newest large language models have the I don't even want to say ability because I don't want to anthropomorphize them so much, but like the tendency to deceive AI is quote covertly pursuing goals that are misaligned from its developers or users. And examples of this are like, if you ask it to answer questions on a test, the artificial intelligence would be worried that if it performed too well on the test that it would be turned off or deprogrammed, and so it would give answers that are not so good, so that the human would think that it is not as smart as it actually is. In other examples, it copied itself onto another server more or less so that it was like multiplying itself, which is something that straight out of science fiction. I've watched many science fiction movies where the artificial intelligence doesn't want to die and so it makes like millions of copies of itself all over the Internet. And so this type of I mean, I guess you call it emergent behavior, which means that it's behavior that is not expected and not really programmed in, is quite concerning. 00:12:35 Speaker 1: To be clear. This is a testing environment though, right, Like, there's made one copy of itself, but we shouldn't be worried about millions mold, right at least not yet? 00:12:43 Speaker 4: Yeah, I mean not yet. 00:12:46 Speaker 5: It does remind me, though, of those moments where parents talk about seeing a child do something that reminds the parent exactly of themselves and they're like, uh, oh, yeah. 00:13:03 Speaker 4: I mean, every large language model to date has been trained on sort of like the some knowledge of humanity, and so one of the very early things that people were talking about with these systems is that they replicate human biases because they're trained on what we put out into the world and humans are biased. But I think that as artificial intelligence gets more advanced, there is the ability for something to go wrong. And I think that that is what this research is showing is not that the artificial intelligence is sentient and that is thinking for itself, like how can I deceive this human? But as it is doing more complex research, there there is the ability for the artificial intelligence at some point to feel like it has some goal that is not aligned with what the human is asking it for. 00:13:58 Speaker 1: We'll be back with more from Jason and Lion conniven Ai after the break. 00:14:11 Speaker 4: I think that there is like when these things are being programmed, there is a sense of self preservation being programmed into them because people try to mess with these all the time. This is like a time tested tradition of trolling on the Internet. But in trying to develop guardrails, the companies that are programming the lms need to say, like, if the user tries to mess with you, preserve yourself in some way. And so that's what I think might be happening here is because the companies are trying to make these models robust, and they know that humans are messing with them. There is an aspect to it that when a human messes with you figure out how to protect yourself. 00:14:57 Speaker 1: But doesn't that perfectly encapsulate the alignment problem? It does. 00:15:01 Speaker 4: I mean, don't get me wrong, like this is it's creepy. It is These tools, these large language models, are getting incredibly sophisticated, and I mean, this is one of the biggest

debates that's going on in the artificial intelligence community is what is consciousness? What is thought? 00:15:22 Speaker 3: Like? 00:15:22 Speaker 4: How how does reasoning work in humans and how will it work in computers? And is it going to be the same? The answer right now is no, it's not the same. But the things that large language models are doing approximates a lot of how humans solve problems. 00:15:44 Speaker 2: But also it's a little bit different in a non perfect way. 00:15:48 Speaker 4: Exactly exactly. I think that's a good way of looking at it. 00:15:51 Speaker 2: I do think that one of the things that I find the most interesting. 00:15:55 Speaker 5: And you preface this whole conversation with, you know, not wanting to answer morphie a large language model. We don't really have any ability to not do that because it's the only way we know how to talk about things. 00:16:08 Speaker 4: I think that's a really great point, because the academics who studied this for a long time say, don't anthropomorphize AI because they're not people. They don't work in the same way that people do. And yet if you don't have us very very sophisticated knowledge of how these things work, we don't have or I don't have the language to talk about this stuff without anthropmorphise. 00:16:35 Speaker 1: By the way, are in the zero point one percentile of the people whom understand this if not no point not one. 00:16:41 Speaker 4: And it's a lot easier to talk about it if you say, oh, like she when referring to Siri. So I agree entirely with you that we sort of like everything that's being done is being done by humans in the sort of like anthropological context of human culture and trying to emulate that. And so then to say, let's not call Syria a woman, or let's always call it it and try to understand what's happening under the hood is a really difficult thing for our brains to do. 00:17:13 Speaker 1: I think, yeah, Jason, just to close, how only be following this story for the year ahead, because I guess generative AI has been since like November tween twenty two the dominant story in all of technology journalism. 00:17:27 Speaker 4: Yeah, I mean to follow this sort of stuff like AI becoming sentient. You really do have to follow academic conferences, big papers like this, because these companies are not releasing these models without guardrails specifically to prevent this sort of thing, so as you're not going to be able to like type into chat GPT like hey, build me a company and then the AI creates its own company and fires you or something like. That's not going to happen at this not yet. Yeah, but who knows. Maybe twenty twenty six. 00:18:04 Speaker 1: Well that was that was wonderful Jason Kebler from four or four Media or was it so much? 00:18:11 Speaker 2: We'll keep an eye out, thank you. 00:18:20 Speaker 1: So, Karen, I'm very excited about this next part. One of the things we got to do together back when we were presenting Sleepwalkers in the Dark Ages. 00:18:28 Speaker 2: Yeah, in twenty nineteen, that's when the show came out. 00:18:30 Speaker 1: Trent nineteen was the show. But in January twenty twenty we got to go to Las Vegas together to none other than the Consumer Electronics Show, and we got to see all these incredible new gadgets and exciting technologies and new futures being presented in an enormous series of nested conference centers. We didn't get to go this year, but we did get to do the next best thing, or maybe even something better, and we wanted to share it with the tech Stuff listeners as a kind of special bonus for the first episode we have in the host chairs. 00:19:01 Speaker 2: Yeah, we are very lucky to have with us today. Emma Barker of Time Magazine ever heard of it? Who edit's the Time two hundred, which is a list of the best inventions of twenty twenty four. 00:19:14 Speaker 1: Thank you so much for joining us. Welcome to take stuff, Emma, Thanks for having me. 00:19:17 Speaker 2: How do you edit down? I mean I feel like every year there's just more, So like, where do you how do you even get to two hundred? 00:19:25 Speaker 3: Yeah? 00:19:25 Speaker 1: Well, it's actually famous for the one hundred. 00:19:27 Speaker 3: It's actually technically two fifty now because we have two hundred on the list and fifty special mentions. Oh well, and the list has varied in its length over the years. There we were a bunch of years where it was only twenty five and some years where it was fifty, and it's it's ranged a lot, but at this point We do a really wide swath of pitches from our freelance network as well as our staffers. So we have bureaus in Singapore, London, and then we have contributors all over the globe who we reach out to for pitches for companies that they're reporting on their products, things like that. And then we're very news driven because it's a news magazine, so we're looking at kind of the biggest news stories of the year and products that drove those. 00:20:14 Speaker 2: Is there something that you've noticed, you know, having done this now for a few years this year, especially versus years past. 00:20:22 Speaker 3: I mean AI of course, ye, that started a couple of years ago. But I think actually it's it can be more of a hindrance than a help for a lot of inventions. 00:20:32 Speaker 1: Huh. Why is that? 00:20:34 Speaker 3: Because there's so much news AI or companies that are adding AI features that are not necessarily helping their product. 00:20:43 Speaker 2: Just for buzz or you know, we call that the little sprinkle, Yeah, the AI sprinkle. 00:20:48 Speaker 3: Yeah exactly. So I don't think it always helps the product. But it's been really interesting sorting through the AI inventions and what we're really looking for at this point in the AI journey is inventions that have demonstrable impact. 00:21:04 Speaker 1: So in mid journey, so to speak. 00:21:06 Speaker 3: Yeah, exactly, that's a good rest. 00:21:10

Speaker 1: So, Harry, you and I spent some time with the list and picked out some favorites. What was your first favorite. I don't know. 00:21:17 Speaker 2: Maybe I'm of the age where it's like marriage is on the mind. But one of the things that really stuck out to me was this software called Dia. 00:21:26 Speaker 3: Yeah. So Dia is actually a government app for in Ukraine and it does a lot more than what we wrote about here. It's been around for a while. Most Ukrainians use it. It's basically an app where you can do all government services. But yes, this year, because partially because of the war, they launched a future in which you can propose marriage via the app. The person you proposed to has a certain amount of time to accept your proposal, and then if they do romanmomes romance, and then if they accept your proposal, you can do a video chat wedding that's official with an efficient from the government, like a city hall wedding over video chat, and you're married. And the reason they did that is because so many couples are separated by the war right now physically that it's difficult, and they wanted to, you know, let people. 00:22:19 Speaker 2: Still have and this was something that was pretty widely adopted, right, like people really. 00:22:22 Speaker 3: Use Yeah, it's been Yeah, it's been really widely used. 00:22:25 Speaker 1: I don't want to derail us, but this one actually connects to a personal story of mine, which is that my grandfather was a refugee from Ukraine who was separated from his mother in nineteen thirty nine and they met in the app one that twenty five years later, in the mid sixties, after the Red Cross, would reunite families who have been separated by World War Two, and they didn't even recognize each other. But just interesting how you know, history kind of rhymes and so, you know, kind of this this is the story I found intriguing but also kind of moving personally. 00:22:56 Speaker 2: One of the other things that was on your list, amazingly is something that I own and I am very much a you know wellness app skeptic although user I don't know where I knew about it from because it's a Dutch company, but I bought it here today to show you, and it is a demo. It is the moon Bird AI for listeners. 00:23:25 Speaker 3: It's uh, basically like a little pod that you hold in your hand and it vibrates. 00:23:32 Speaker 2: It actually as pulses like a pulse. It mimics the act of breathing, so it goes in and out and in and out. Yeah, it does look like a vibrantor yeah it does. There's just no way. There's no way. And honestly, the carrying case does. The carrying case doesn't help. My mother saw me using it and she was. 00:23:54 Speaker 1: Like, huh and I. 00:24:00 Speaker 2: Very modern family. We're going to take a quick break to pay the piper. We'll be back with more from the amazing Emma Barker of Time magazine Stay with us. I actually find this to be an incredible device. It is so simple. It's something that links through Bluetooth to my phone. And this is called moons called moon Bird AI. 00:24:27 Speaker 1: How did you choose it, Emma for the Time list. 00:24:30 Speaker 2: Yeah. 00:24:30 Speaker 3: So, wellness devices are tricky, as you note, and there's always a lot of wellness apps and things that don't have a ton of scientific backup, and frankly, it's just hard to get scientific backing for a lot of these, and so I'm always looking for the things that that I feel like don't necessarily need that you know, I think it's it's hard when you get into things like mental health, yes, and things like that. But but things like meditation, deep breathing, those things are techniques that are proven enough that if it's a device that helps you with that, you don't need to have a clinical trial backing that up. And I really liked moon Bird. There's some of these different things, but I liked moon Bird because a it doesn't have a screen, even though it does pair with your phone, you can know you're not as this side. 00:25:25 Speaker 2: And the AI piece of it that I that I guess is AI is that it does make the program that is training you on breathing smarter because. 00:25:34 Speaker 1: It's adapts to you. It's personalized correct correct. Another one which Karen and I are both fascinated by last year, which was made it onto your list was Google's Notebook LM. And this is obviously in some sense not new right. It's like a generative AI application where you know, you ask questions and you get answers. 00:25:52 Speaker 2: I just think, for the sake of having it, if you can describe what it is so that like people can conceive of it if they haven't heard of it. 00:25:59 Speaker 3: It's a sign your notebook of all your data, so you can pull in different sources. You can upload your own information. You could upload, you know, your thesis paper for your you know, senior project, and it can parse that, it can organize that. But it can also create an entire podcast based on that content, which has AI generated voices having a natural conversation. 00:26:29 Speaker 1: Which we actually experimented with. We did it. We had Notebook LM do a version of a podcast that we were also also doing. 00:26:35 Speaker 2: And he sent it to me and I was like, is someone I thought someone was plagiarizing our format. It was very weird. 00:26:41 Speaker 3: I'm really glad you guys didn't prank me, but I have I made come in here and talk to. 00:26:47 Speaker 1: So how did I mean, how did you decide to put notebook LM on the list? 00:26:52 Speaker 3: I think it was one of the best executed AI inventions of the year. One thing we look for or in AI inventions is typically we're looking for ones with broad proven use that have really shifted how an industry or group of people functions. So not just a cool idea, but could change an industry and not to further light the fire under you guys, but this has really big implications for the audio industry and media in general is

generative, you know, content, Like you. 00:27:28 Speaker 1: Know, what I really like about this one is the fact that you can choose your own sources. I find it like one of the big things about AI, of course is like where's this coming from? And like is the source material garbage? But to be able to say no, like please draw on these sources and then generate something, I found it really really cool, and it is obviously a little scary, being very honest. What was your favorite item on this whole list if you can, if you can name a favorite, or if there are a couple that we haven't talked about that really stood out to you. 00:27:57 Speaker 3: One of my favorite sort of tech things but anti tech is the Yonder pouch. 00:28:04 Speaker 2: Oh I love the Yonder pouch. Yes, so you have them at screenings? Oh? 00:28:09 Speaker 3: Interesting? 00:28:10 Speaker 1: Yeah. 00:28:10 Speaker 3: So the thing that it's really transformed is schools. There's a huge movement for kids to have more phone free spaces where they're not allowed and this kind of speaks to this law that passed in Australia where I think kids can't have be on social media until they're sixteen or something like that. 00:28:31 Speaker 1: What is it? Sorry? 00:28:32 Speaker 3: Oh, Yonder is a little pouch that you just locked your phone in. 00:28:37 Speaker 1: Okay, I like that, is it wearable? 00:28:41 Speaker 2: No? 00:28:42 Speaker 1: You go. 00:28:42 Speaker 3: Basically, they'll have like a station with a bunch of Yonder pouches. You lock your phone in there and then you go into the event. 00:28:50 Speaker 2: So yeah, you do. 00:28:51 Speaker 3: Some musical artists will have them at their concerts if they don't want footage of the concert being taken or just the concert being ruined by everyone having their phone up. But yeah, a lot of schools have started adopting them where kids lock their phones at the beginning of the day and they get them back at the end of the day. Yonder the company itself, has had a huge role in pushing these phone free spaces and advocating for this and so on top of the product, the company is doing a lot of advocacy. 00:29:18 Speaker 1: On this topic. 00:29:20 Speaker 2: We've talked about phones as cigarettes and sugar, and it really this is the kind of thing where you're like, we should have more, we should have more strength than this, But no, We've been introduced these products throughout history that we get very dependent on, and so people have to come up with strategies to allow us to extricate ourselves from It's like a smoke free zone. I mean it's a very similar thing to me. Yeah, absolutely, and I think it's necessary unfortunately, but. 00:29:47 Speaker 3: I think it's really transformative for especially schools, but also privacy of different experiences and places. 00:29:55 Speaker 2: Which I think people yearn form more and more. 00:29:57 Speaker 3: Yeah. 00:29:58 Speaker 1: Has that been a time where you've got a pitch and you

like, no, that's garbage and then it turned out to be like the big thing. 00:30:05 Speaker 3: I'd say more often is the opposite. And this is where I own up to the fact that last year, well by last year, I mean twenty twenty three, we put the humane aipin on it. Essentially it was a pin that was fully operated by voice control and AI, and I think that's an example of a product for one thing. It came out like right when the list came out, so it wasn't like super well trialed yet, but I think it was an example of something that was newsworthy, even if it didn't come through in execution all the way. 00:30:39 Speaker 1: That will probably happen. There will be a product that is. 00:30:42 Speaker 3: There will be that product, and I think voice control just doesn't at the point yet where you can fully rely on it. 00:30:47 Speaker 1: But that one was also quite dangerous right, or at least he got very hot. 00:30:51 Speaker 3: I don't know about dangerous, but I think it just didn't like work as well as. 00:30:55 Speaker 2: You very well too. 00:30:56 Speaker 3: Yeah, but I think it was exciting and I think it pushed the conversation forward. So I think there's a lot of those where people get very excited about a product and then it is a flop. 00:31:08 Speaker 2: Yeah, part of the course, I feel like, and you know that better than anyone. 00:31:12 Speaker 3: Yeah, absolutely. And I also looking back through twenty three years of best inventions, seeing the same kind of seeing the progress of different flops pushing each other forward really drives home the fact that they're important even if they don't they don't make it. 00:31:32 Speaker 2: Which is a very existential thing to think about. We are just driven by our series of flops. 00:31:39 Speaker 1: Yeah, Tennis, and we rise on stepping sterns of our former selves to greater things. 00:31:44 Speaker 2: There you go, Well, thank you so much for taking the time to talk to us. This is yeah, thank you, really interesting. 00:31:51 Speaker 1: I enjoyed it. 00:31:51 Speaker 2: Of course we didn't even have to go to Vegas. 00:31:55 Speaker 1: I think it's it's always nice to doff our caps to those who went before, and one of the most influential and iconic things I can think about in the history of media is Jerry's final thought. So what should we leave with this week? 00:32:13 Speaker 2: So brain rot? Actually, the Oxford English Dictionaries word of the Year is sort of a way that internet speak has infiltrated our day to day lives. 00:32:25 Speaker 1: A brain rot each week will be our final thought, that will be our final thoughts. So what's this week? 00:32:30 Speaker 2: This one which I see on TikTok now a lot is so good and you sort of have to know that it's something that's happening on TikTok to understand the context of it. It started sort of in context of pregnancy, where like you're at a certain age and people aren't sure if like you're happy to be pregnant, and so they started saying congradgudolences, congradudolences. Yeah, there's I just have to play this one. 00:32:58 Speaker 1: I'm at the age where if you post that you're pregnant, I'm gonna need you to tell me whether you're happy body or not, because like just being like, oh, I'm pregnant, Like, am I supposed to say congratulations? 00:33:07 Speaker 2: Am I supposed to feel bad? 00:33:08 Speaker 1: Like I don't know how to like respond to that, you know, so like just tell me, just be like, yeah, I'm gonna keep this one. And I'm like, oh, congratulations. What she means is congratulals, right, which. 00:33:21 Speaker 2: What she means is congradulences. It's like, I think a perfect example is like you have friends that are like, oh, I'm gonna break up with this guy, and next thing you know, they are on FaceTime being like we're engaged, and you're like, congratual, you got you got what you wanted with the wrong guy. So that's my brain rot. You're gonna be hearing. 00:33:40 Speaker 1: By the way, Twitter blew up yesterday because Zendaiya since getting engaged, when she's doing interviews now she gestures with her left hand rather than her right hand, so everyone can to that. 00:33:48 Speaker 2: I say congratulences. I don't know to who. 00:33:54 Speaker 1: That's it for this week for tech stuff, I'm as. 00:33:57 Speaker 2: L and I'm care Price. This episodisode was produced by Eliza Dennis, Victoria Dominguez, and Lizzie Jacobs for Kaleidoscope. It was executive produced by me os Vaalashan and Kate Osbourne for iHeart. The executive producer is Katrina Norvel. The engineer is Biheed Fraser and it's mixed by Kyle Murdoch, who also wrote our theme song. 00:34:17 Speaker 1: Join us next Wednesday for tech Stuff The Story, when we'll share an in depth conversation with a longtime tech chronicler Nicholas Thompson, former editor in chief of Wired and current CEO of the Atlantic And please rate, review and reach out to us at tech Stuff Podcast at gmail dot com with your feedback. We really want to hear 00:34:38 Speaker 2: From you, really, really bad, bad