--:-- [Theme music fades out]
--:-- David Pierce: Hey there, what's up? How can I brighten your day today? Does that voice remind you of somebody? Hi, how you doing? I'm well.
--:-- [Guest 1: Will Knight, Senior Writer at Wired, joins the conversation]
--:-- Will Knight: OpenAI is actually pulling Sky, that new voice, the one for its new chatbot, just days after rolling it out. Because, yes, it sounds a bit too much like a certain Spike Jonze character.
--:-- David Pierce: Ladies and gentlemen, her. In the past week or so, a bunch of companies have actually announced a bunch of new AI tools. And this is becoming something of a pattern. Roll out a flashy new AI product, something glitches horribly, roll it back in. But wait, weren't the CEOs of these companies just telling us about how they were about to change humanity forever? We're going to do this. It's going to be net great, but it's going to be like a technological revolution. And those always come with change.
--:-- [Announcer voiceover]
--:-- Announcer: Ahead on Today Explained, AI's hype versus AI's reality.
--:-- David Pierce: those email sequences, those real-time reports, those powerful automations, which they claim can help you build what matters most, your company. You can go to adio.com slash todayexplained to get 15% off your first year. That's A-T-T-I-O dot com slash todayexplained.
--:-- [Announcer voiceover]
--:-- Announcer: Having a night out with your friends is awesome. Dealing with hangovers sucks. Meet Senorita, the THC-infused margarita that lets you skip the booze but keep the vibe. Made with real fruit, organic agave, and Himalayan pink salt, every sip is like a mini getaway. No fillers, no hangovers, just full flavor and a better buzz. Find your flavor at senoritadrinks.com. Use code VOX20 for 20% off. Must be 21 and over. This product may cause intoxication. This product has not been evaluated or approved by the FDA. This product may contain up to 0.3% THC.
--:-- David Pierce: It's Today Explained. I'm David Pierce, filling in as host today. Will Knight is here with me. He's a senior writer at Wired, where he covers artificial intelligence, which means that last week, he too was watching Google's annual developers conference. And this year, as you can imagine, was just 110% AI, generative AI in everything. turned up to 11. It was the most AI event they've put on and probably the most AI one I've ever seen.
--:-- Will Knight: Everyone get out your computers. It's time to have ourselves a nice Google. Yeah, so I think the most noticeable thing is going to be the change to search they showed. I mean, this is their cash cow huge business and it's so important for many other businesses, right? and they are rolling out these AI overviews, which are generative AI summaries of the search results. Instead of having to go through all these links, ideally it will summarize it for you as if the AI has gone off and read all the web pages and it's got all the answers.
--:-- David Pierce: So whatever's on your mind, whatever you need to get done, just ask. And Google will do the Googling for you. It doesn't always mean the answers are going to be right. Someone was pointing a camera at a broken film advance lever on a SLR, like a film SLR camera. The question was, why can't I advance this thing? Why is this lever not moving all the way to advance the next thing? The answer that Google delivered in its own video and highlighted is the most wrong answer.
--:-- Will Knight: But it does also really threaten to kind of upend this symbiotic relationship that companies have had with Google because it means no longer people have to necessarily have to go and click on those links. And this summer, you can have an in-depth conversation with Gemini using your voice. They showed off new versions of Gemini, which is their answer to chat GBT, which can see things and talk to you as it sees the world. So it's almost like a little robot in your pocket where you can show things. We're calling this new experience Live.
--:-- David Pierce: When you go live, you'll be able to open your camera so Gemini can see what you see and respond to your surroundings in real time. One of the other things that caught my eye was they start to show off what they call, and other people call, agents. So these AI programs you don't just talk to, but you'll give them a task and they'll go off on the web and try to complete it. The idea is that this could be a totally new way to sort of use computers.
--:-- Will Knight: It's pretty fun to shop for shoes, and a lot less fun to return them when they don't fit. Imagine if Gemini could do all the steps for you, searching your inbox for the receipt, locating the order number from your email, filling out a return form, and even scheduling a pickup. That's much easier, right?
--:-- David Pierce: Having played with some open source ones, I've played with some of these people who have cobbled together these tools. They are equal parts really, really amazing, and you can see the potential. And equal parts are just bonkers, and they go off the rails and do something. And the stakes are that much higher if you're sending emails or changing your calendar or you've given it your credit card or something. So if they can figure out how to make them reliable, which I don't think they can quite do now, so they have to really limit what they can do, then I think it could be really transformative.
--:-- Will Knight: And the usefulness, I think, is going to be less clear. It depends if they can get them to actually work. It does seem like that's kind of an amazing baseline that we're still at, that if only they actually work, how cool it would be.
--:-- David Pierce: Well, the announcements keep coming. One way or another, they're going to keep launching stuff. And the other one was, what, like 24 hours before Google OpenAI did a big launch. Compare and contrast the vibes for me. Google has a big, giant developer conference. What did OpenAI do?
--:-- Will Knight: Yeah, this was a much more small-scale, intimate thing just at their headquarters. Today, we are releasing our newest flagship model. This is GPT-4.0. The CTO on stage with a couple of engineers showing off the new chat GPT. I will say that the vibe was also somewhat bonkers because they showed off this cool, impressive new model, but they also revealed that the latest thing they've done is make it remarkably like the AI from her.
--:-- David Pierce: Hey, ChatGPT. Hey there, what's up? How can I brighten your day today? In that it's kind of inclined to flirt with people. Wow, that's quite the outfit you've got on. Which is a twist, a plot twist I wasn't expecting. OpenAI new voice chatbot is getting attention not for its ingenious tech but because it sounds suspiciously like US actress Scarlett Johansson Johansson says OpenAI CEO Sam Altman wanted to hire her to voice Sky but that she declined the offer due to personal reasons.
--:-- Will Knight: She points out Altman insinuated the intentional similarity when he tweeted a single word, her, the day the ChatGPT product was announced. So they just said that they were going to change the voice because of this backlash over it being a little bit too like Scarlett Johansson and a bit too sexy. So what else did OpenAI announce?
--:-- David Pierce: They had the new voice, which was definitely kind of the star of the show, the not Scarlett Johansson, but kind of Scarlett Johansson voice. What else did we get from OpenAI? Right. So they showed that underneath the hood, they have this new model, which was ChatGPT 4.0, which is a completely new reimagined model, which takes video and audio so it can do the voice, but it can handle video as well. So they showed all these examples where in real time you can talk about what you're seeing.
--:-- Will Knight: Show it to me whenever you're ready. Okay, so this is what I wrote down. What do you see? Oh, I see. I love ChatGPT. That's so sweet of you. The idea here is that you may have this kind of new paradigm in personal computing where you've got something that's always seeing what you're seeing. It can remember where you left the remote control. It can tell you about everything you're looking at.
--:-- David Pierce: We were expecting maybe a brand new, super-powered new model, which was going to be able to do way more things and be way more intelligent. So it does feel like a little bit of a cop-out or a bit of a swindle that that's the new kind of AI. It just happens to sort of have more of a personality, as it were.
--:-- Will Knight: Do you think we're getting inklings of what the kind of first huge killer app of AI is going to be? I mean, you mentioned search, which seems like it has the potential to be one. Maybe it is just chatting with a stranger. Maybe it's some of the agent stuff that we're seeing. What is your sense of, are we getting glimmers of where this is going?
--:-- David Pierce: I don't think we've seen what would be the killer apps. I think there could be, you know, huge companies built on top of kind of quite mundane or seeming, much less sort of sexy uses of AI and these models that just to, you know, automate all sorts of tasks. And that could be really big.
--:-- Will Knight: Financially, I haven't seen something that's like, oh, this is really going to change everything. I mean, that was one of the fascinating things with ChadGBT. It was clearly a really big research advance and it was wild to play with it, but it was never the case that you could say, oh, this, I can really see how this was going to change work. Even in cases where people would say, oh, you can write essays with it, you can do... But there were all the problems with it. And I mean, you can already copy essays off the internet, right? I mean, people do use these tools.
--:-- David Pierce: I think they've sort of crept into their workflow somewhat in sort of smaller ways. But it's not like if we've seen like this completely killer app akin to something like the smartphone or the internet yet.
--:-- [Announcer voiceover]
--:-- Announcer: In a minute, we're going to ask whether all of these problems with all of these AI tools are actually fixable.
--:-- David Pierce: Support for Today Explained comes from Built Rewards. These days, it seems like everybody has a rewards program. You swipe your card and you get one cent on the hundred dollars. But with your rent, frequently, you just pay month after month. You never get anything from it other than that roof over your head. Not anymore, says Built Rewards.
--:-- [Announcer voiceover]
--:-- Announcer: Built Rewards says it lets you earn points on your rent. By paying rent through Built, they say you can earn flexible points that can be redeemed toward hundreds of hotels and airlines, perhaps a future rent payment, your next Lyft ride, and so much more.
--:-- David Pierce: Your home isn't just where you sleep, it's your community. And Built wants you to go out and do things in the community. You can eat at your favorite local restaurants, then earn more points. You can get VIP treatment at certain fitness studios. You can enjoy exclusive experiences just for Built members. every month.
--:-- [Announcer voiceover]
--:-- Announcer: When you're ready to travel, they say Built Points can be converted to miles or hotel points all around the world. You can earn points on rent and around your neighborhood wherever you call home by joining built.com slash explained. That's J-O-I-N-B-I-L-T dot com slash explained.
--:-- [Announcer voiceover]
--:-- Announcer: Every day millions of customers engage with AI agents like me We resolve queries fast We work 24 and we helpful knowledgeable and empathetic We built to be the voice of the brands we serve Sierra is the platform for building better more human customer experiences with AI No hold music, no generic answers, no frustration. Visit sierra.ai to learn more.
--:-- David Pierce: Support for the show today comes from LifeKit, the podcast from NPR. What's in your life kit? I'm asking. They didn't ask that. Every day you have to make choices that might shape the direction of your life, but you can't flip to the back of the book to find an answer key. Instead, you can listen to the Life Kit podcast from NPR.
--:-- [Announcer voiceover]
--:-- Announcer: Life Kit offers real stories about common issues like relationships, finances, parenting, and your career. If you're looking to move with more intention or just need thoughtful guidance on living better, Life Kit delivers strategies to help you make meaningful, sustainable change. From fitness routines to mental resilience, navigating personal goals to tackling burnout, each episode is full of relevant insights and clear takeaways about things we all could use a little help with.
--:-- David Pierce: Life Kit wants to help you approach your own life with confidence and clarity and walk away with a game plan you can implement right away. Enough from me. Go listen to the show, Life Kit. It's a podcast from NPR.
--:-- [Theme music fades in]
--:-- David Pierce: Google.com. What is it? What does it mean? Why are we here? No one knows. And you're not going to find out. Not today. Explained. We're back.
--:-- [Guest 2: Julia Engwin, New York Times contributing opinion writer and the founder of Proof News, joins the conversation]
--:-- Julia Engwin: Tech CEOs have been hyping this AI future for a while now. You know, I've always thought of AI as the most profound technology humanity is working on. More profound than fire or electricity or anything that we have done in the past.
--:-- David Pierce: Julia Engwin is a New York Times contributing opinion writer and the founder of Proof News. She argues that the product demos we saw last week are, let's just say, not quite on the level of fire just yet.
--:-- Julia Engwin: Well, I felt really vindicated by those announcements because I think they were really underwhelming. You know, Sam Altman promised us that he was going to show us something magic. And it was kind of a routine update. Like you would see, you know, the new iPad actually was probably slightly more magical than this update of OpenAI.
--:-- David Pierce: Same thing with Google, right? They pulled out all the stops for this announcement. Google! No! But like, I'm hard-pressed to tell you anything that really was like a compelling, like, oh my God, I'm so excited to try this.
--:-- Julia Engwin: And so I feel like the problem is that they started the AI conversation with, AI is so smart that it's going to kill you. Like, where do you go from there? Right? Yeah. Like, it's really hard when like the gap between that and the reality is that it can't really answer even the basic question.
--:-- David Pierce: and then the gap just gets wider and wider. Yeah. So with that as kind of the hype machine, where do you feel like we are in real world, like ground truth AI stuff? What is your sense of where any of this stuff actually is right now?
--:-- Julia Engwin: Well, I mean, if you just look at the studies, it kind of consistently comes back to AI is like a 50-50 coin flip. So you look at medical diagnoses, right? There's a bunch of papers. The most recent one that I looked at from Stanford's human-centered AI lab shows that when they were looking at how AI performs on citing medical studies, it was basically 50 percent the evidence didn't support what AI was saying, right?
--:-- David Pierce: Even some of the spectacular things that were touted, like remember when ChatGPT supposedly aced the uniform bar exam and they said that they had scored in the 90th percentile, like a new study from an MIT researcher actually found that it was in the 48th percentile.
--:-- Julia Engwin: So we're seeing that it does more than maybe would have expected a computer to do a couple years back. But it's not reliable. Coders who work with AI as like a coding assistant will tell you, you have to know enough to debug the code that it generates for you. So it requires a bit of expertise basically to fact check the AI.
--:-- David Pierce: So what do we do with that, though? Because on the one hand, 48th percentile on the LSAT is something. It's better than my computer would have. It's better than I would do on the LSAT, I suspect.
--:-- Julia Engwin: Yeah, definitely. Me too, for sure. On the one hand, cool when it tells you the truth. On the other hand, I now have to go fact check it every single time. And so maybe we've accomplished nothing because I have to just go do the work that it did to make sure that it's telling me the truth this time.
--:-- David Pierce: So I don't know. I'm so torn in this moment whether to be impressed that we've made any progress at all or totally annoyed by the fact that the progress we've made just makes me do even more work.
--:-- Julia Engwin: Yeah. I mean, this is why in my piece in the New York Times, I referred to it as like a bad intern. His work you have to check, right? Because it's basically like, you know, sometimes you get a great intern and they are like producing great stuff and sometimes you don't. And sometimes it's more work than actually just not having help, right?
--:-- David Pierce: And so I feel like we're kind of in that space with AI. I think it wouldn't be as frustrating, honestly, if they hadn't promised us that we were on a steady march to what they call artificial general intelligence, which is this sort of mythical concept that there's going to be one AI machine that can do everything, right? Like it's going to be an expert at chess and at like drone strikes and law and medicine.
--:-- Julia Engwin: And, you know, none of us knows a human being, right, who has all of those expertise. And so So they have told us, like, this is moments away. We're almost there. In fact, we are so close to being there that we're worried.
--:-- David Pierce: And I think that that is the problem, is that we were told that, and then we have the reality on the ground is so different than that. Yeah, so on that front, let's run with the intern analogy a little bit here. Because if you have a bad intern, the theory, supposedly, is that you can teach that intern how to be better, right?
--:-- Julia Engwin: And that the work you're putting in to do twice as much work because the intern's a doofus comes back in the end because you eventually teach the intern how to do the job. And I would think there's that there's some of that happening in AI.
--:-- David Pierce: But you've also described some pretty big barriers between where AI is and where all of these folks are suggesting that it might go and is worth waiting for. Can you walk me through some of the big barriers between here and there?
--:-- Julia Engwin: Well first of all I do want to say like it definitely possible We going to get better interns out of this eventually but I think there are some major barriers One major barrier is energy use. So AI is incredibly expensive in terms of its energy use. It's why there are really only the big tech companies are actually only the really main ones able to compete in this space because you need incredible cloud computing resources.
--:-- David Pierce: And then they're running up against energy uses. Basically, they need these energy intensive data centers to power this AI models. One study finding that training a single large language model program takes the same amount of electricity needed to power 120 homes for an entire year.
--:-- Julia Engwin: Microsoft just broke ground on a new one. But in Virginia, like somebody who runs the data centers there said like we're done. Like we can't expand anymore. Sam Altman has said we have run into an energy wall. This seems to me where things are going. Like we're going to want an amount of compute that's just hard to reason about right now. How do you solve the energy puzzle? Nuclear. That's what I believe. Fusion. That's what I believe. Nuclear fusion.
--:-- David Pierce: Yeah. So they have an energy barrier. They also have a data barrier. The current hypothesis for AI is that you need more and more data to make your models better. I think I just want to put a pin in that, that that may or may not be true.
--:-- Julia Engwin: But if it is true, and that's what they're operating under, like they have already scraped all the data that is available on the public internet. And, you know, as we've seen from the lawsuits by New York Times and Authors Guild, it was not a consensual scraping.
--:-- David Pierce: And just to be clear, I'm not looking to shut down AI or turn the clock back. I just want guardrails so that AI fairly compensates the people whose work comprises its entire brain. So there's two problems. One, all those people are suing them and may want their data back out of those models.
--:-- Julia Engwin: And two, they're talking about basically creating synthetic data to train their models. So basically having the AI invent data to then make their models bigger. And then people have talked about that leading to like model collapse.
--:-- David Pierce: And so, you know, you have two huge inputs into AI, both of which are looking pretty shaky. And so it does, I think, raise the question of like how much exponential progress they're going to continue to make.
--:-- Julia Engwin: And I would say this. I think there is an interesting question about whether figure is really better. There are a lot of really interesting experiments being done where scientists, doctors are putting more qualified, smaller data sets together and then trying to build AI applications on top of those.
--:-- David Pierce: And I haven't seen enough to know yet how successful those are going to be. But I think that's a really interesting question. Like, I would feel more comfortable myself querying an AI model that was only populated by, like, peer-reviewed medical studies, for instance, right?
--:-- Julia Engwin: And so I think there could well be an interesting AI future that involves a lot more specialized models. And that would make sense, by the way, with the history of technology, right? Specialized machines have generally been what we rely on. And so the idea that AI was going to be that was always kind of weird.
--:-- David Pierce: Totally. So let's just go glass half full here for a minute and put on our rose-colored augmented reality glasses. The problems you're describing, especially these kind of big infrastructural ones, are they solvable? Is there a world in which we get through some of these big barriers in the next few years?
--:-- Julia Engwin: I think maybe. I always would say you have to bet on American ingenuity, right? What's one thing about this country that is wonderful is that when we want to do something kind of no holds barred, we try to figure out how to do it.
--:-- David Pierce: But I think the question is at what cost, especially on the energy side, is it worth pouring all this money into bigger and bigger AI models and data centers? or should we have a nationwide EV supercharging network so everyone can have an electric vehicle? Like, you know, we're in a world where everything has a tradeoff.
--:-- Julia Engwin: And so I guess the question I have is, yeah, maybe we could if we put all things aside and decided, like, this is the thing that we want. I'm just not sure the return on investment is worth it.
--:-- David Pierce: Well, and there are some folks arguing for precisely that, right? I mean, Sam Altman is out here saying pretty straightforwardly without a hint of irony that it might cost trillions of dollars to get AI to the point where we need to and making the case that that is the correct use because the benefits at the end will be so huge and so society serving that it'll be worth it.
--:-- Julia Engwin: Not to put words in your mouth, but I'm guessing you're fairly skeptical of that argument. If you were going to make that claim and have it be believed, you have to show some evidence, right?
--:-- David Pierce: And like the reality is, he's a little bit the boy who called Wolf. He said so many times how much magic we're going to see just on the next round and the next model. So I would believe it if like Jan LeCun said it, because he's been much more measured.
--:-- Julia Engwin: He's a meta, he's an AI pioneer. And he says, look, it's going to take forever. We are nowhere near close, and it's going to take a lot of hard engineering and iteration. And like, that feels like right to me.
--:-- David Pierce: Today's show was produced by Amanda Llewellyn, edited by Amina Alsadi, fact-checked by Laura Bullard, and engineered by David Herman, with help from Andrea Kristinsdottir. I'm David Pierce, and this is Today Explained. Up! Google! Google's gonna wake you up! Thank you.
--:-- [Theme music fades in]
--:-- David Pierce: it's landing your fundraise. The truth is, banking can do more. Mercury brings all the ways you use money into a single product that feels extraordinary to use. Visit mercury.com to join over 200,000 entrepreneurs who use Mercury to do more for their business. Mercury, banking that does more.
