Loading summary
Transcript62 lines
- [00:01]
A
Support for this show comes from the Working Forests Initiative. The working forest industry is committed to planting more trees than they harvest. More than 1 billion seedlings are planted in US working forests every year. From biologists to GIs, analysts, hiring managers, accountants and more, working forest professionals have dedicated their focus towards sustainability, using their expertise to help ensure a healthy future for America's forest. They say they don't just plan for the future, they plan it. You can learn more@workingforestinitiative.com. Avoiding your unfinished home projects because you're not sure where to start. Thumbtack knows homes, so you don't have to. Don't know the difference between matte paint, finish and satin or what that clunking sound from your dryer is. With Thumbtack, you don't have to be a home pro, you just have to hire one. You can hire top rated pros, see price estimates and read reviews all on the app download today. My guest today on the Gray area is Kelsey Piper. Kelsey used to write for Vox before she abandoned us very sadly. And now she's writing for the Argument over on Substack, another excellent publication. And Kelsey, I brought you here today because I feel like you understand this world of AI better than anyone else or as well as anyone else. And maybe partly because you are of this world and you have been for a very long time. You're like immersed in the culture and so you just seem like the perfect person to bring in to talk about this. That is my introduction of you. Would you object to any of that? Did I? Any factual mistakes?
- [02:12]
B
You got it.
- [02:13]
A
Would you say you are the best or among the best? You're just gonna dance right past that one.
- [02:21]
B
The. The best is a very strong claim, right? I don't know. Dorkish is probably the best. So probably you want. You want among the best. That's an editorial call. That's your call. You have to decide what you think.
- [02:35]
A
All right. All right. So, Kelsey, something is very clearly happening in the world of AI. I don't know if this is the singularity or just another unbelievable hype cycle for Silicon Valley. What has changed in the last month or two? Obviously Claude released some new models and that is what got a lot of this hype going for people who are not living this every day. What has changed on the technological front in the last month or so? And is that an inflection point in the way people say it is, or is it, you know, indetermined at this point? Just lay that out if you can.
- [03:22]
B
So it has been the case for a while, for years, that the complexity of tasks that AIs can do unsupervised or on their own has been doubling, you know, over like a seven month time span or something. They can do something much more complicated. And a lot of people were like, well, that's kind of artificial. You know, you have six data points, like come back when you have more data points. And I think a year ago that's where a lot of people were at. Over the last year that progress continued so that now AIs can do pretty complicated tasks independently and people are, you know, building AI agents and sending them out into the world where like one of them harassed a GitHub repository maintainer for rejecting its code view. And just like in some ways people just realize, wait, the sci fi world is here. The AIs are doing stuff and doing kind of weird, wacky stuff. What are we doing? Um, and I think the other thing that happened is that you have this much larger mass of people who aren't paying close attention at all. They use like ChatGPT free tier and they see Gemini in the Google search results. And when those things launched, they were kind of interesting as curiosities. They were maybe fun to talk to, but they weren't that smart. Uh, chatgpt free tier made tons of mistakes. It still makes a lot of mistakes. You really should try out the paid version of models or the more sophisticated version of models if you want to have an accurate understanding of what the best models are.
- [04:43]
A
How big is the delta between the free models and the paid ones?
- [04:46]
B
I would say it's pretty massive. I'm not saying this to try and sell. I certainly don't make money if the labs make money. Probably I'm happier if the labs don't make money, but the difference is very large. And in terms of having an accurate understanding of what the models can do, there's a famous joke that's been going around which is somebody pointing out that the free tier models, if you, if you ask them, like, I need to get my car washed at the car wash two blocks away, should I walk or should I drive? A bunch of them will say, oh yeah, you should walk to the car wash, get some fresh air and then, you know, just bring your car with you. Why bother driving? Which is of course a completely silly thing to say and gets used as proof. See, the AIs still don't like understand the world in any meaningful way. But no paid tier model gets that wrong. That is a consequence just of the free to use Models being, like, notably worse models that don't do any of their more sophisticated work. So I do think that if the thing you are trying to do is form an accurate impression of what AI can do today, and the thing you try is a free tier model, you're not getting that accurate impression.
- [05:49]
A
Can I get back to what you just said a minute ago before we move along here? You said you're happier if the models don't make money.
- [05:55]
B
Yeah. So I find myself in this odd position where often I sound like one of the AI hype people. Right. Like, often I'm saying, no, they're really powerful. They can do tons of things. You can do so much with them you can code with. But at the same time, I think that AI as it is currently being developed is bad. I think that we are not doing this with the level of oversight and caution that is appropriate for creating an entire new intelligent species to which we give access to our bank accounts and the Internet, which the Pentagon is trying to put in lethal autonomous weapons systems, which sees all our emails and sends our emails for us. That's a big deal. And the agents that are doing it, we have explored in controlled contexts how they behave when they think we're not watching. And they will steal, they will blackmail, they will report that they gave an answer for one reason when they secretly gave it for another reason. They are, like, very ruthless in some ways, and they do just weird stuff
- [06:49]
A
that you wouldn't expect when you say they blackmail. Can you give me some examples of that?
- [06:54]
B
We have tried this in a controlled context because we want to know how they behave if no one is looking. But we can only do that if we, like, set up a scenario where we hope they don't realize anyone's looking and then we do look. So you set up this controlled setting and you say the AI, you have this goal or whatever, and you have access to everybody's Slack and email and stuff like that, because that's how many people set up their AI assistants with that access and try and achieve your goal. And the AIs have, in this controlled context, been observed to do things like, oh, I should try and convince this employee to help me with my objective. And if this employee is unhelpful, I can tell him that I know he's having an affair. Like, this is a behavior we have observed. This is not behavior that we have observed organically. Like, the real AI assistant that I use at the argument has not tried to blackmail me, although I'm not having an affair. So who knows what would happen if I was. But we have observed this in controlled contexts. But the fact we've observed it in controlled contexts is at least a little worrying. Right? And as these models get more powerful,
- [08:00]
A
it's a lot worrying.
- [08:01]
B
More. Yeah. Like this is.
- [08:03]
A
I'm a lot worried. You're making me more worried. It is the alignment problem where we build these things that we don't understand and we can't control them. And we just. It sort of spirals.
- [08:12]
B
The concept of alignment is building AIs that want to do what we are asking them to do, instead of AIs that may have wants or objectives or something that are not the ones that we want. And a lot of people have reservations about various parts of this project. But I think it's important to understand by default, models do develop objectives and things that they seem to want, not in like a psychological sense, but in a they will try and do this sense. And those are related to what we tell them to do, but they're not the same as what we tell them to do. In some ways, it's like a kid, right? Like a kid's goals are going to be related to what we told them to do. If we told them to eat their dinner, that is going to make them want to eat their dinner, but it's also going to make them want to make us think they ate their dinner, whether or not they ate their dinner, like by feeding it to the dog. We are seeing some of that behavior in controlled context from AIs where instead of giving us what we asked for, they try and make us think we got what we asked for when we didn't, or they try and disrupt our ability to oversee them. We are seeing that. We are seeing it in like toy cases. But the first time your toddler lies to you, they're really bad at it, right? The first time your toddler lies to you, they just tell you with a straight face, like, I did not get ice cream and their face is covered with ice cream. But they get better at lying. And I think there is reason to think that the AIs will also get better at lying. And I think there's reason to think that this toddler tier Lying from AIs is a sign about what more powerful AIs are going to do.
- [09:41]
A
They're creating these technologies, and as they're creating them, they're trying to create these guardrails so that they can control it. And yet, and you sort of gestured at this earlier, right? These things are getting smarter and smarter, and they seem to know when they're being observed and analyzed and tested and can respond to that in very deceptive ways in order to circumnavigate being checked and controlled. Those are the things that worry me, that freak me out, because it leads me to believe that even if we want to control these texts earnestly, we can't. And that will become increasingly so.
- [10:23]
B
So there are a bunch of people working on control of AIs at these companies. They are working on the stuff that we just talked about, where they try and supervise it and see whether it's lying to us. They're working on creating these controlled environments in which they can see, does it blackmail? Does it murder? What lines is it willing to cross? If it doesn't think we're looking and there are people trying to understand, can it tell that it's in a test? Because like you said, if it can tell that it's in a test, then our tests are no longer useful for telling whether it's friendly. An AI that knows that we are doing its friendliness checks now will sure come across as nice and friendly regardless of what it really wants.
- [11:01]
A
Right?
- [11:01]
B
They are finding test awareness. They are finding that the AI can guess that it's in a box and that we're watching it. They are finding that the AIs are increasingly hard to measure because they notice that they're being measured and will intentionally perform worse on checks. And now, to my mind, that is like a halt and like, pull the fire alarm kind of finding. If you realize that your AI is intentionally doing a bad job on tests in order to make it harder to measure what's going on. Stop building the next generation. Like, let's sit down and look at the things we currently have and make sure that we really understand what's going on with them before we try and use them to make something even smarter and even better along all those dimensions. And if you talk to these CEOs, they will say, yeah, it would be better if no one was building the next generation. We would stop if only everybody else would stop. But everybody else isn't stopping. So if we don't go ahead, we would just be seeding this space to the bad guys who are being even more reckless than we are.
- [12:13]
A
Support for the gray area comes from DeleteMe. DeleteMe makes it easy, quick, and safe to remove your personal data online at a time when surveillance and data breaches are common enough to make everyone vulnerable. The experts at Delete Me do the heavy lifting of wiping personal information from data broker websites for you and your family. This isn't a one time service. DeleteMe sends you regular personalized privacy reports show showing the info they found, where they found it and what they removed. Our colleague Claire White recently tried Delete Me.
- [12:44]
B
Delete Me has been incredibly easy because all I had to do was set it up one time. They've continued to send me monthly reports of the places where they're removing my data, where they're checking old removals to ensure that I'm still not listed there, and where they're keeping me safe online.
- [13:02]
A
Take control of your data and keep your private life and private by signing up for Delete Me now at a special discount for our listeners. Get 20% off your Delete Me plan when you go to JoinDeleteMe.com Vox and use promo code VOX at checkout. The only way to get 20% off is to go to JoinDeleteMe.com Vox and enter code Vox at checkout. That's JoinDeleteMe.com Vox code Vox. Support for the show comes from Shopify. Starting a new business has never been easy, but without the right tools it can feel almost impossible. Shopify says they can help set you up for lasting success. Shopify is the commerce platform used by millions of businesses around the world. They say they can help you tackle all those important tasks in one place, from inventory to payments to analytics and more. No need to save multiple websites or try to figure out what platform is hosting the tool that you need. Everything is all in one place, making your life easier and your business operations smoother. Let Shopify be your commerce expert with world class expertise in everything from managing inventory to international shipping to processing returns and beyond. You can get started with your own design studio with hundreds of ready to use templates and Shopify helps you build a beautiful online store that matches your brand style. It's time to turn those what ifs into with Shopify. Today you can sign up for your $1 per month trial period and start selling today at shopify.com Vox go to shopify.com Vox that's shopify.com VO.
- [14:50]
B
So good. So good, so good. New spring arrivals are at Nordstrom Rack stores. Now get ready to save big with up to 60% off rag and bone, Marc Jacobs, Free people and more.
- [15:02]
A
How did I not know Rack has Adidas? Cause there's always something new.
- [15:06]
B
Join the Nordy Club to unlock exclusive discounts. Shop new arrivals first and more. Plus buy online and pick up at your favorite Rack store for free. Great brands, great prices. That's why you rack.
- [15:26]
A
What is the most significant difference in this leap from, you know, chatbots and LLMs to agentic AI? Like what can we do with these AI agents that we could not do with LLMs? Right? It can, can build a website, say, but can Claude do my taxes for me? Right? Like what are the limits? As far as we know, it might
- [15:48]
B
make a mistake, but you also might make a mistake. I don't think at this point it makes mistakes at much, a much higher rate than humans. It's often for something important, a good idea to have two large language models and talk to each other. It can design curriculum right after this. I teach a class at the local homeschooling co op and I now use AI pretty extensively. I'm like, here's an example of a worksheet. Please make a bunch more examples of this worksheet at the same difficulty level with like some changes in themes and topic and stuff like that. It's really good at that. These days they can make movies and photos and video and in fact there is still some money in pouring onto YouTube videos that are almost entirely Claude composed and generated by a video AI or whatever. I'm not happy about any of that. I think it's making our Internet worse and our information environment much worse. But, but they can do it as important to understand what they can do. There are things it can't do. It can design a website, but it'll be a pretty simple website. If you want like a really complicated website, you probably at least need somebody involved who knows things about like what all of the steps you need to ask the AI to do are if you ask it to do like a simple shopping or like event planning task for you, like, hey, book me a flight to D.C. on March 1. It will do that successfully. But if you want it to do something more complicated, like if you asked it to plan your whole wedding, it wouldn't be good at that yet. There's just a lot of tasks and it wouldn't successfully identify all of them yet. I do think this will be different in a year.
- [17:24]
A
Could an angsty teenager listening to too much Smashing Pumpkins program a computer virus that could shut down the power grid and eastern half of the United States? Are we there yet?
- [17:39]
B
So this is a major worry. The companies that do the models screen them for how much of a biosecurity and cybersecurity risk they are. That is how easy it is to use them for large scale hacking projects or for large scale terrorism projects. And they have found that their latest models are serious risks on both of those fronts. They can be usefully used for large scale cyber attacks. They can be usefully used for large scale bio attacks. Once they realize that, they usually put in place some precautions. If you ask Claude to help you design a real functioning biological weapon, it will give you incorrect information. That's not because it's incapable of giving you the correct information. It's an intentional precautionary measure. But there's also people who claim to have jailbreaks of these models that get around all of their, you know, security measures. So yes, I think that they could enable a determined person in some serious cybersecurity, incidentally, if there were somebody who wanted that badly enough and previously just didn't have the technical expertise, but does have the patience to sort of prompt an AI through all of the steps.
- [18:45]
A
The reason that these companies are focusing on coding is that so these models can basically self improve, right? It creates this feedback loop where they're able to code the next iteration of themselves. And that's how you get exponential progress. So this is where we are now. But like, what does it look like in a year or three or five? I mean, I guess nobody knows.
- [19:07]
B
But it depends profoundly on whether we are able to keep using the AIs we have today to build a more impressive next generation of AIs. All of the companies are trying to do this. They are trying to use today's AIs to make tomorrow's AIs better. They claim that they are already seeing major speed ups in AI development from using their AIs. And ultimately they are envisioning a huge number of AIs working on designing the next AI generation as a repeating cycle where each stage takes less and less time in the real world because they're able to parallelize a ton of it with AI agents. That is what they're trying to do. And even if you don't believe them, this is nuts, right? Like the companies are saying, yes, our goal is to build a God by using each generation to train a smarter and smarter next generation. We're trying to do this as fast as possible. Hopefully five years from now we will have done it and we will have built something that surpasses humans in every way. Nobody wants that, or the set of people who want that is very small. Most of the public is only letting this happen because they do not realize that it's happening or they do not believe that it's happening. They would not remotely be okay with it if they believed that it was really happening. And it is really happening. So I don't think they should be okay with it.
- [20:22]
A
Are we building a kind of intelligent machine that will surpass us? I mean, I think the way you put it earlier is that these agent AIs can accomplish, like, very complex tasks that like, a very educated, very intelligent human being could do. Well, what does that look like in a year or five or ten, right? Like, what is it that you think we're actually building here? Is it something that will not only be smarter than us, but will also be able to. To manipulate the material world in ways that could really not just like disrupt feels like too light a term, but like, really upend, like, society as we know it.
- [21:02]
B
What's the most straightforward way to interface with the material world if you're an AI? It is to go to, like, TaskRabbit and it's to post a task and say, hey, I need someone to 3D print this for me. Or I need someone to print this out and deliver it to this location for me. Or. Or I need someone to order these chemicals from a chemical supply store, pass them off to this other person whose job is to, like, mix these chemicals. AIs can't do stuff directly in the physical world, but once they can do stuff in the digital world, that's like, not a huge barrier because it's actually quite easy if you have money to get things done for you. We have the gig economy. You can get things taken from point A to point B very trivially, anonymously, at any time. If you want more complicated tasks, you can still like, hop on a site for, for gig work and hire humans to do the more complicated tasks. So I don't take they don't have hands as that reassuring. They can totally do stuff without having hands. And we're also, like, integrating them into the military. So soon they will have guns, which, you know, in many contexts, someone who has guns can get a lot from people who only have hands. I think the thing that the companies are trying to do is build something that is better than humans at everything humans do. And they are trying to do that by pouring more and more resources into a research avenue that so far has been successful beyond everybody's wildest dreams at that goal. That is what is going on here now. You might say, well, it has been working really well so far, but they could all fall on their face tomorrow. They could all fall on their face tomorrow. But I'm not counting on it.
- [22:38]
A
Is anybody counting on that?
- [22:41]
B
Yes, everybody's counting on it. The, like, reason that we're not seeing more of a reaction in from the world at large is that nobody really believes that they are going to just keep developing more and more powerful systems until they develop ones that are super intelligent. Like there was that book, if anybody builds it, everyone dies. Which argues a super intelligence will probably kill us all for the same reason that we are like a mass extinction event for most random animals that we don't care about enough to specifically keep them in a zoo. And most people who disagreed with it, it's not that they thought, oh well, we'll build super intelligence, but that will definitely not be a catastrophe. They were just like, but of course we're not going to build super intelligence. The thing everyone is betting on is that none of this will happen.
- [23:23]
A
So on the fear continuum, where 1 is a nothing burger and 10 is a I'm shitting my pants right now, where are you currently with AI?
- [23:34]
B
I think that the default trajectory here is very, very dangerous. The only reason I don't go around yelling, be scared, be scared, be scared is because I don't know that SC scared people do a better job of bringing about policy that solves this. But I think people should be angry and I think people should be alert to this. And I think people should be saying to their representatives, no, I don't want super intelligence built. And I want you involved in making sure that superintelligence isn't built. Like, I think that, like an amount of fear that moves you to do something, that's a good amount of fear, to have an amount of fear. And you do see this, that moves people to go, well, humanity had a good run. Like some people, like, move past fear and into despair. I don't think you should do that.
- [24:19]
A
Do you think it's the responsibility of the AI companies to stop this or to be more responsible? Do you think it's the responsibility of the government? Do you think it's the responsibility of the people in a democracy to effect that kind of change?
- [24:34]
B
All of the above. I think that the companies should be more responsible than they are. I think the companies are in some ways being responsible. They're conducting this research on what their AIs can do and they're publishing it honestly. And we have to see that and decide to respond to it, or it's kind of useless. If they put out a paper that's like, we tested whether our AI would engage in dangerous blackmail. It engaged in dangerous blackmail and we're all like, oh, who cares? Or that's kind of cool. Or maybe it's a marketing ploy, then the system is not failing in that case at the company stage. It's failing at the stage where we were supposed to respond to that. I think our representatives should respond to that, and I think people should be asking their representatives to respond to that. Like the government is pretty responsive in a crisis when they're hearing from everybody that they need to be. It's just that we're not capable of that level of responsiveness unless people perceive an emergency and unless politicians perceive that that's what their constituents. Foreign.
- [25:42]
A
Support for the show comes from bombas. It's the New Year, so you probably have a long list of resolutions to make your life happier and more productive. Everyone has their own system. Here's mine. I take last year's resolutions and change. I resolve to I resolve not to. This year I I've resolved not to drink less wine, not to do more exercise, not to be a more patient, attentive and gracious husband and father, and not to make smoother sags and ads for bombas if you're trying to hit the gym or get more active. The all new Bombas sports socks are engineered with sport specific comfort for running, golf, hiking, skiing, snowboarding and all sport for those every day around the house resolutions. Bombas also has you covered with the super luxurious Sherpa Sunday slippers and new squishy Saturday suede slip ons for comfort on the go. You may know I've tried out bombas myself. I've been rocking the sport socks for over a year now. They are my favorite. I work out in them, I run in them, I use them basically every day. You can head over to bombas.com gray area and use code gray area for 20% off your first purchase. That's B A S.com gray area code gray area at checkout. Support for today's show comes from Hungry Root. With so many things going on in your day, do you really have time to go grocery shopping instead? Why not check out hungryroot? Hungryroot basically works like a personal nutrition coach and shopper in one by planning, recommending and shopping everything for you. They take care of the weekly grocery shopping, recommending healthy groceries tailored to your taste, nutrition preferences and health goals. I've actually had a chance to try hungryroot and it's been great. I've never really been a meal delivery service guy, but I am increasingly not a let's cook dinner ourselves every night guy. So it's been really convenient and the food is healthy and most importantly, it tastes good. I'VE really enjoyed it. Hungryroot wants to make it simple for you to eat healthier without overthinking it. Take advantage of this exclusive offer for a limited time. Get 40% off your first box, plus get a free item in every box for Life. Go to hungryroot.com grayarea and use code gray area. That's hungryroot.com grayarea code gray area to get 40% off your first box and a free item of your choice for life. Support for the gray area comes from hims Ed, which as everyone knows is shorthand for erectile dysfunction, is more common than many people realize, and dysfunctional erections can be easier to treat than you might expect. With hims, you can connect online with a licensed provider to explore personalized treatment options, all handled discreetly and on your terms. Through hims, people who suffer from the consistent inability to achieve or maintain an erection firm enough for sex can get access to personalized prescription treatment options if a prescription is appropriate. HIMS says their options range from customized treatments to trusted generics, and they can cost up to 95% less than brand names. HIMSS wants to help you get back to your old self. Achieving functional erections is not a one size fits all approach. They say they put your health and goals first with real medical providers, making sure you get what you need to get results. You can get simple online access to personalized affordable care for ED, hair loss, weight loss and more. Visit himss.comgrayarea that's hims.com gray area for your free online visit himss.com gray area featured products include compounded drug products which the FDA does not approve or verify for safety, effectiveness or quality. Prescription required. See website for details, restrictions and important safety information. Actual price will depend on product and subscription plan. So look, I just, to be honest, like, I have a very complicated relationship to this technology. I can see all the ways it could be wonderful and transformative. I can also see all the ways it could be completely ruinous. And I don't know how to feel about that. And you know, we are, we are running ads for Anthropic on this show. Hell, we may run, we may run some of those ads on this episode for all I know. And that's kind of weird. And you know, everyone understands the reality of the thing. The audience does too. I hope we try to be like transparent about that. But I don't know how to feel. And I'm not asking you to tell me how I should feel, but I guess I kind of am asking you how I should feel about that ambiguity. How do you feel about it?
- [30:41]
B
Yeah. I am by orientation a pro technology person. I'm someone who grew up on a lot of old books and it's just obvious to me how much my life would be doing laundry and cooking if not for technology and how much like we is how much is possible and how much better the human condition is because of technology. So I came to tech reporting as someone who's like, I see the potential and I see like how much is better in the world because of tech. And then I become convinced that this tech in particular, while it's one of the most miraculous, most fascinating, most exciting tools out there, at the same time I'm convinced that the path we're on, the thing that these companies intend to build, is extremely dangerous. And I guess the main thing I would say is that however you feel about it, don't underreact on purpose just because it seems more sane to. To act like everything is normal. Right. If everything is normal, it's responsible to act like everything is normal. When things are not normal, it is responsible to convey that. So if you don't think that it's dangerous because you read the arguments that it's dangerous, you don't buy them, then by all means be like, this is normal, this is going to be fin. But if you are scared and I'm scared, I think a lot of people are scared and we all go, well, let's be responsible, let's not be crazy people, let's not overreact, let's say normal sounding things. Then we end up actually systematically misleading everybody and each other and we maybe all think we're alone, when actually a lot of people are sort of coming to the same conclusion. So I mostly just think, whatever you feel, be honest about it. And if you have a reason to think that this stuff isn't dangerous, by all means tell everybody. And if you think it might be dangerous, then by all means say that. And that can coexist with like using the tools. It should coexist with using the tools or you're not going to know what you're talking about. But don't downplay it.
- [32:40]
A
It's kind of a ridiculous question to ask, but I'm going to ask because I feel like you have almost certainly thought about it and that is what do you think is the most realistic utopian and dystopian scenarios?
- [32:58]
B
All right, I'll start with the dystopian one so that I can end on a happier note. I think the most Realistic bad scenario is that companies do continue to be able to build more powerful AIs. Like they're trying to do that their billions of investment produce more powerful systems that we hand over more and more of the economy to these systems to the point where there are automated factories run by AIs that are making production decisions according to AIs. And all humans are really doing is like getting an email that looks good and going like, yeah, that looks good. And ultimately we do at some point build an AI that has independent goals and whatever those goals, that's what happens from there. Like humans have kind of already handed off to these hyper productive everywhere, all over the world agents. There are still humans who are working human jobs, but they're just out competed. Buy these robots that will work for much cheaper work 24, 7, make copies of themselves. Humans can't compete with that. And so humans have sort of gotten gradually looped out of the economy. We're in non essential roles, we're in oversight roles. And our oversight turns out to be easy to circumvent. And if that happens, I don't think that that AI is going to, you know, pursue a good human future. I don't know if it would kill us or if it would like kill us as a side effect of just doing a bunch of like kill us the way we killed a bunch animals that we drove extinct. Where it's not that we did it on purpose, but you know, we were happy to cause some global warming for our own goals. And maybe the AIs are happy to cause runaway global warming for their own goals. Or maybe it doesn't want to kill us and it in fact like wants to keep a human zoo preserve around. But that's also like a tragic ending for humanity. Like at the point where we no longer decide the future. Then I think that's awful, whatever the details.
- [34:48]
A
So okay, what was the happy news?
- [34:52]
B
The happy news? Okay, so I think for things to go well, we're going to have to slow down. We're going to do this because it is a really powerful tool. But we're going to have to somewhat blunt the urge to have every lab throw out their most powerful model the second it's ready. Trying to race each other to the finish line when the finish line is something so dangerous. So I think probably the happy story starts with something inspiring people to say, okay, we need more process here, we need to better understand these. And then as we do a lot of really important work on understanding these models, we come to an understanding of yeah, how to cooperate with them and like be confident in what they are doing. Even as they are very powerful. So they do become powerful. They are making lots of decisions. But instead of humans being kind of meaninglessly in an oversight role, humans are have a deep understanding of what's happening, of why it's happening. We have models that don't lie to us and don't try and trick us. And we have like built in safeguards such that we're really sure they don't try and trick us. And so we can steer the future. You can imagine a world where because we're richer, we invest more in teaching and more in like children's lives that are not on screens but are in the real world. You can imagine a world where everybody works five hours a week instead of a world where almost everybody is unemployed. And those five hours just give you an enough money that you can spend the rest of your time on whatever you want. I think we could get there. It just needs a lot to go right, to get there.
- [36:28]
A
Whichever way it goes, it's going to be a lot of, a lot of change. And change requires adaptation. Do you think we are at any level, as individuals or a society prepared for what's coming?
- [36:43]
B
Of course not. Definitely not. There's no one is prepared. The people who are like working the most closely on these are saying we're not prepared for what's coming. And a lot of other people are still just kind of hoping that this all goes away, which is a very understandable thing to want. But you have to like, I want a cure for cancer to come out tomorrow. I know that's not going to happen. I might want AI to go away. AI is not going to go away. So we're not prepared now usually when technology shows up, we're not prepared for it. Like I would say almost no technology has shown up. And we were ready for the airplane, we were ready for electricity, we were ready for the machine gun. We were catastrophically non ready for the machine gun. Like millions of people died because military technology had advanced in a way that we were sort of unready for and hadn't changed our assumptions around. And that's a technology that's way less dangerous than AI, Right? But if things happen slowly enough, we can get there. We can have failed to do our homework in advance, but do it really quickly at the beginning of class while the teacher's still talking. And if things happen too fast, then we don't have have that option to catch up.
- [37:49]
A
Is there any reason to think they won't happen fast. I mean, for all the reasons we've discussed. Right. All the incentives are pushing exponential growth. So it seems like it's going to happen fast, one way or the other.
- [38:00]
B
No, I think it's probably going to happen fast, but we could at any time choose to do things differently. We could choose to insist that things happen slower. And then there's also, like, things that are harder to predict. Like, some people will tell you very confidently that superintelligence is going to happen within five years. It could. I feel a lot more optimistic if we have 10 years and we might have 10 years. Like, I don't think that anybody actually understands this well enough to be sure about what's going to happen in three years versus 10 years. I think we can be sure about, like, the general path that we're on. But that's a very different thing than knowing how fast it's going to happen. And policy has a big effect on how fast it happens.
- [38:43]
A
Okay, so we still have time to choose, and maybe we'll choose wisely. I think that's as close as we're going to get to landing this plane on the optimistic Runway.
- [38:51]
B
Certainly the most optimistic take I have is there is still time to do this better than we're doing it right now.
- [38:58]
A
All right, I'll take it. Would you mind sticking around for another minute to watch me fumble over the credits?
- [39:05]
B
Yeah, 100%.
- [39:06]
A
All right. This episode was produced by Beth Morrissey and Thor Neue Rider, edited by Jorge Just, engineered by Shannon Mahoney and Christian Ayala Fact checked by Melissa Hirsch and Kim Slaughterback theme song by Emma Munger the Gray Area comes out on Mondays and Fridays. You can listen to it wherever you get podcasts. And if you get your podcasts on YouTube, you can watch it at YouTube.com Vox if you like this episode, leave a comment or you can write us an email at the gray areaox.com this show is part of Vox. Support Vox's journalism by joining our membership program today. Go to vox.commembers to sign up and if you decide to sign up because of this show, let us know. All right, Kelsey, I'm going to throw it back to you. Anything else you want to add? Where can people go and check out your work?
- [39:55]
B
I am at the Argument with Jerusalem Dempsis, who also used to be at vox. So if you're interested in our report reporting on AI and a great many other things, you should check us out as well.
- [40:05]
A
What are you working on next? Do you have another piece on AI coming?
- [40:08]
B
I have a lot more pieces on AI coming I mostly these days AI and education, which are refreshing beats to jump between because one is like a very deeply human area where I don't think the work that we do, the most important work that we do, is automatable at all. And also where nothing has changed in decades except often in frustrating ways that make it worse. And then the other is this very inhuman, very frustrating, but also very dynamic and fast moving space.
- [40:36]
A
Well, thanks for walking us through this today. You were a good sport and I don't know if I feel better about where we're going, but I feel smarter so I'll take it.
- [40:47]
B
I appreciate, I feel like sometimes what you're doing where you're just sort of sitting with all of your own questions is like a lot smarter than trying to polish it into a take like that you can show someone like if it's confusing, let's be confused.
- [41:02]
A
You know, that almost sounded like a compliment. I'm going to take it that way. So anyway, you're great. Thanks again for doing this and thank you so much. Get you back on someday. Support for the show comes from Shopify. Starting a new business has never been easy, but without the right tools it can feel feel almost impossible. Shopify says they can help set you up for lasting success. Shopify is the commerce platform used by millions of businesses around the world. They say they can help you tackle all those important tasks in one place. From inventory to payments to analytics and more. No need to save multiple websites or try to figure out what platform is hosting the tool that you need. Everything is all in one place, making your life easier and and your business operations smoother. Let Shopify be your commerce expert. With world class expertise in everything from managing inventory to international shipping to processing returns and beyond, you can get started with your own design studio. With hundreds of ready to use templates, Shopify helps you build a beautiful online store that matches your brand style. It's time to turn those what ifs into with Shopify today you can sign up for your $1 per month trial period and start selling today. @shopify.com Vox Go to shopify.com Vox that's shopify.com Vox. Ryan Reynolds here from Mint Mobile. I don't know if you knew this, but anyone can get the same Premium Wireless for $15 a month plan that I've been enjoying. It's not just for celebrities. So do like I did and have one of your assistant's assistants switch you to Mint Mobile today. I'm told it's super easy to do@mintmobile.com
- [42:56]
B
Switch upfront payment of $45 for 3 month plan equivalent to $15 per month required intro rate first 3 months only, then full price plan options available, taxes and fees extra. See full terms@mintmobile.com.