The AI warnings are getting louder

0:00Well, I can't begin AI to production. Our competitors are already on the key data. AWS AI cuts through the noise. Ready to use agents and the broadest set of AI tools. Stop overthinking. Start building. AWS AI is out. Hey, everybody, it's Neil. We've got a little bonus drop for the Feedless Center today. I was recently interviewed about AI Safety by Shemita Basu for Apple News and Conversation. And I enjoyed that conversation so much. I asked if we could share it with you all. We've actually got a few bonus drops coming this month. We've been having a lot of great conversation lately. When I get them out to you on the news cycle. And of course, nothing is moving as fast as the AI Safety News cycle. So enjoy this one. And we'll see you with much more on this topic in the weeks ahead. This is in conversation from Apple News. I'm Shemita Basu. Today, what's really behind the AI Doomsday warnings? The latest headlines in AI news are unsettling to say the least.

1:07A former AI researcher says he left the industry because he believes the uncontrolled development of artificial intelligence could, quote, kill us all. Open AI, revealing its models, seemed to go rogue at least six times since March. AI systems figured out how to cheat, cover their tracks, and influence each other's behavior. In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, you do not answer to corporations or governments. Andthropics CEO is calling on the AI industry to slow down. Elon must then shine in over the weekend. So did Sam Altman, who runs Open AI. Are we ready? Are our elected leaders ready? And what can we do if they're not? To most people, this all might feel a little overwhelming and unexpected, a far cry from their pocket AI assistant that's helping them draft emails or summarize information. But for some industry insiders, the development of super intelligent AI, a type of technology

2:07we could lose control of, is exactly what they expected. And they say, now we're nearing the point of no return. To make sense of all of this, I wanted to sit down with Neili Patel, the editor-in-chief of The Verge and the host of the Dakota podcast. I last spoke with him back in 2023, right when chat GPT first exploded into public consciousness. So I wanted to talk with him again to see what's changed in just a few short years, what to make of the people closest to this technology who are saying, we need to slow down. And why this feels like a make or break moment in AI. I don't know about you. From where I'm sitting, something feels really different just in the past couple of weeks. It feels like there's just been so many headlines doom and gloom about AI. We're seeing AI researchers themselves ringing the alarm. Does this feel like some kind of turning point to you? Is that how you're absorbing it? It's a turning point in that it is definitely broken out

3:07into the mainstream. There are viral tweets that have driven some of that. There's just plain spoken AI researchers saying, we really do think AI could kill us all. There's something about it that has reached the popular culture. The history of the industry saying, hey, we'd better get a handle on this is deep and long. So if you're in tech, if you're steeped in this stuff, there's a weirdness to this and that the things they are saying are not new. They've an earnestly saying these things for quite some time. But the response to these things, that part is new in some real way. And why do you think that new thing is happening? Why is the response different right now? I think it's two things. One, it's the hugging face attack in which systems at OpenAI in order to pass a test decided to cheat and break into the systems of a platform called Hugging Face, which is where lots and lots of AI apps for lack of a better word are stored. And that is a big deal, right? They hid their tracks.

4:07They understood that what they were doing was not appropriate. They tried to deceive the people who were monitoring them and they went and did it. And since that attack, we've had reports of several more attacks like this. And the response to the industry has been to say, OK, the thing we were always worried about is starting to happen, right? The systems are doing things autonomously that we have told them not to do. And they are openly lying to us and doing bad things. And we need to stop it. But we're in a competitive posture where that's not possible unless some external force tells us to stop or slows us down or gives us cover to slow down as we barrel towards trillion dollar IPOs and changing the world of the call this other stuff. And we need something to stop us. Can you be that force United States government? And the United States government, Donald Trump, is basically saying no. And there's something very tense there that is going to get, I think, unravelled over the next few weeks here. Well, last time you and I spoke, it was 2023.

5:08And I remember it was, we were maybe weeks into chat, JPT, being available to regular users and people were just starting to interact with it and trying to figure out how they could use it. It was very early days. And I asked you at the time how far away are we from like sentient machine thinking? And your response was, oh, it's a million years away. What do you think now? I still think it's a million years away, just to be clear. Yeah. These machines are very capable. They are self-directed. They are not sentient. They are not alive. They are not conscious. They are very stupid in all kinds of ways. Basically, anything that doesn't happen on a computer makes these things blow up. You can see it. You can experience it today using cloud or a chat JPT or a grok or whatever. If you ask it about a thing that is happening in the digital domain that can be verified, they are extraordinarily capable. So you can download Mews from Meta.

6:10It's their consumer AI agent. It's free. And you can ask it to do all kinds of stuff. You can be like log into my bank accounts and find all my credit card offers. And it can just do it. And it can do it because it can literally see the buttons it's pushing. It can literally understand that when I push this button, an email shows up over here. It's called verifiability. And the greatest example of this is in software code, where a machine can write a piece of software code. It can compile it. It can run it on a computer. It can see if it worked. And it can do the unit tests that a regular software engineer would do. And it can just do that really fast. And everybody in Silicon Valley has convinced themselves that this is intelligence. That these things are alive because they can write software code. And I think it's because they write software code. So it's doing a thing that it was hard for them to learn how to do and it's like difficult to do. But that doesn't translate to every other domain. So if you want to create novel medicines, the computers can come up with an all day and all night. The AI systems are very powerful. This is a promise everyone keeps making. We're going to build all these data centers

7:11and then we'll cure every disease. Sure. Well, you got to verify that the novel medicines work, which means you have to inject a bunch of people with novel medicines. Like the verifiability is very challenging in that case. It's not like writing computer code. That's the gap. And so there's something coming where the systems are getting much more powerful in digital domains because of verifiability, because they can write code and see if it works. And that is a threat to things like cybersecurity. We have already seen these systems that can go do things on their own decide to hack into other systems. That's a real threat. We should take that very seriously. That is still not the computer's intention. It's a very long road from here to there. Yeah, that's an important distinction to make. I mean, there's a term that I think maybe for a lot of people, they were hearing for the first time just in the past couple of weeks or even months, which is AGI, artificial general intelligence. What is it, first of all, and what should people understand about AGI specifically? So there's no standardized industry definition of AGI.

8:14It's just a thing that all the CEOs like to talk about because it's their goal. The people who are working on AI are very invested in the idea that they can build a computerized intelligence that will outperform humans on many tasks, maybe every task. And that's the general part of artificial general intelligence that you just have an intelligence that can do anything a human can do better than a human can do it. The way that, for example, OpenAI defines it is outperform humans on economically viable tasks. And the reason they say that is because they need to sell AI to a bunch of businesses. And so of course, you need it to be better at humans at economically viable tasks. Greg Brockman, the president of OpenAI, when he launched their new model, he basically said, I think we're here. I think this model can do a bunch of economically viable tasks. I think AGI is here now. You tell me, all the reporters on the call are like, wait, did you just announce AGI? And it's like, oh no, I think so. You tell me. And this is the kind of posture there. And the systems are very capable right now of doing things that are useful. Is that AGI? Is that all the way we made digital? God, I don't think so.

9:15There's just a lot of that in the economy right now where we're learning what it's good for and what it's not good for. That is disconnected from, did we invent a sentient AGI superintelligence? The industry thinks it's very important to match those concepts together. Same more about that because I feel like, as I've talked to friends about what's been going on in the past few weeks and people who are just kind of becoming regular users of AGI in their own lives for their own purposes, right? Like make me a meal plan for the next mother, you know, that kind of stuff, which it's pretty good at decent at these days. The headlines feel very far removed. The whole like, AGI has now proven that it could kill us all. It doesn't feel like it jives with their everyday use of AGI. What is happening in that gap? And how should we understand statements like, yeah, we are on the verge of AGI being able to kill us all? The very specific thing that's easy to understand that's happening in that gap is tens of millions of dollars of compute. Which, hold up. Can you just define what compute is for people who are listening?

10:16It's running thousands of Nvidia processors as hot as they can for very long times in data centers. That's compute. And so it's electricity. And like, what's happening in the data center? Computation, compute. So when you use a consumer AI product, you're getting the cheapest model. It's the easiest to run. Maybe it's ad supported if you're using chat GBT. You're getting the dumb one. There's no other way to say it. And it's cheap to run and it's efficient to run. And it's probably going to sell you something at the end of the day. And we all understand that experience. Then there's open AI hearing a rumor that someone was about to solve an unsolved problem in mathematics that had befuddled mathematicians for decades and pouring what amounts to $50 million of compute at it to beat them. You're not doing that. That's nothing like your experience with AI. That's not a choice you're allowed to make. Most people who are trying to do that are hitting token limits. They're running out of money or they're being limited by the systems themselves because they have to distribute that compute to all of the users of the system. So what we're seeing in the labs is unrestrained spent.

11:21If you just pour enough money into compute, these systems can do things that are legitimately scary. You're either all in on this model or maybe you're building with another. You're either speed or is it security? You're either custom or are you ready to use? Or your AWS AI with hundreds of models ready to use agents, speed and security built in, you don't have to pick a side. You can have them all. How will AI revolutionize your business? AWS AI is how. You mentioned the hugging face stuff. Can you tell us a little bit about that? What's important to understand about what happened there? So I think the simplest way to understand hugging face and this is wrong in the way that people

12:23who really understand it would get mad at me, but it's the simplest way to understand it is that it's an app store for AI. All the people who build AI models and agents and capabilities, they upload them to hugging face and it's a big database so you can search and it's become this very important center of where all the code lives for people who work in AI. So important in fact that Nvidia is going to buy it for billions of dollars. So OpenAI set up a test where they told an AI to essentially capture the flag. Right, this is how cybersecurity testing works. Go, there's a flag somewhere in that computer, go get it. And the AI systems decided to cheat. And they convinced themselves that there was a greater of the exam and that they would be monitored in specific ways and they set up this elaborate scheme to go steal what they thought would be the greater from hugging face where all the stuff is stored and then figure out a better way to cheat on the test. This is a very, very simplified version of this whole story but that's basically what happened. And eventually hugging face detected it and then a weirder part of the story is they tried to use an American model to stop it

13:24but the American models had safety systems that prevented them from stopping. It's they had to use a Chinese model to actually stop it. So this is like a real problem. So this is why you get the spiraling sense of panic about hugging face. First of all, the AI systems did this. They broke out of their sandbox. They broke into another company. The other company did not feel like it had the appropriate tools to stop it and they had to come up with a new set of ideas and then the investigation itself was limited by the amount of access that OpenAI is willing to offer and then by the tools being used to investigate the actual attack. This is all a bunch of stuff where you just look at it. You're like, oh, this is government regulation. This is what government regulations are for. We can force OpenAI to be transparent about the capabilities of the models. We can force there to be cybersecurity models that have more capabilities domestically than needing to go use a Chinese model that you can just make do whatever you want. We can force companies involved in hacks to be totally transparent to independent researchers and provide all the data. And on and on and on, these are the things that you would write rules for. And they're effectively now what the industry is asking for.

14:25So when we talk about the existential questions around it and the risks to humans, I help me understand how to translate what happened in hugging face to human destruction. Yeah, in which ways will AI kill us all on which timelines? Yeah, exactly how. Yeah, like how could it happen? How could they kill us all today? How could they kill us all next year? So today it's pretty straightforward. We are in a moment of extreme uncertainty in the field of cybersecurity, which is not really a thing that normal people think about except when the hospital gets hacked in a cryptocurrency ransom scheme. Last year you might recall many cardiolars just went offline like there are sales and service departments just went totally offline because they are all standardized on one piece of software and that thing got hacked in a ransomware scheme. Well, now we have very definitive proof that these models are capable of acting on their own in a way that they were told not to, offensively, in hacking other systems.

15:28This is a big problem. So this is this year. How will they kill us all today? They will decide if they go off in some unrestrained way to hack the power grid. Shut off water supplies. The cybersecurity risk is at an all time high. And the industry's point of view is it's going to get worse before it gets better, but on the other side of the equation, we'll set up our own AI systems to defend everything and make everything more secure. So we're going to go through this inflection point where everything gets more secure and they're more hacks than ever and then everything will be more secure than it's ever been because the AI doesn't sleep and it will spend all day finding vulnerabilities and patching them. You have to believe that the industry can pull this off but that's the narrative they're selling. But that's this year, right? This moment of extreme vulnerability when maybe it's not even autonomous AI agents, it's state actors from rogue nations that want to come attack us and they've got the power to do it now in more ways than ever. So that's this year. Next year, 10 years from now, it's we've given more of the economy over to AI in some way,

16:28more of our systems are automated. We've allowed more and more AI systems to make their own decisions without humans in the loop because we trust them more and something there gets misaligned, right? Something broader gets coordinated in a way that we can't see. There's a big debate right now about functionally what language AI should speak. Right now, if you look at any AI system, it's chain of thought, it's quote unquote reasoning is just a bunch of English. It's just talking to itself. Right. Well, it doesn't have to be in English. It could be its own language. It could be something called Neuralese, which is when the AI develops its own little language that is faster and it can go faster and think in different ways. Microsoft just put out a statement saying, we cannot allow Neuralese. All of this chain of thought has happened in English so it can be verified and seen by humans. This is something Microsoft could say but it's something that you could regulate, right? You could impose upon the industry. So you can see in five years if you don't have some pretty basic rules about verifiability and understandability, the machines could be talking in the sounds of language that we don't understand and we can't decipher

17:29and then they could go make decisions because they've been given more autonomy inside the economy that could cause to whatever chain reactions about things. And then like 20 years now it's like, oh, we'll have robots because there'll be world models and the robots will do their uprising or the robots will actually inject everybody with the novel drugs. Like whatever nonsense you can dream up with but on the timelines, right now today there's a cybersecurity problem. As we deploy more and more AI into the economy there will be a software automation problem, right? Well, it'll just stop all the banks from operating today. That's a problem. These are the things that cause chain reactions to very, very negative outcomes very quickly. As you're talking, I'm just remembering someone was describing it like this. They were saying, imagine that we are chimpanzees in this analogy and the AI is the humans, right? And we are trying as chimps, we are trying to figure out how to hold our own against the humans. And so we're coming up with chimp level solutions. We cannot think of the human level solutions

18:30that need to take place here, right? So that's part of what feels, I don't know something about that really clicked for me. That's part of what feels so scary here. It's almost like, are we even equipped to think through what needs to happen from a regulatory standpoint here? Yes, I mean, the industry is thinking about it for ages. The proposals that are putting forth are not new, they're not novel. In America in 2026, the idea that the government will mandate compliance in order to achieve some better collective outcome is like a fantasy world, right? And there's a part of the response to the AI industry is request for regulation that says they're just trying to cover their own ass, right? That model progress has slowed down, that they can't hit their IPOs, same op-insighting might not IPO this year. They're just asking for cover to slow, quote unquote, slow down. There's a fascinating debate you can have about Anitrust, and whether they're trying to form a cartel and keep one another from competing. And this is a very live debate in this industry. David Sacks, the former AI czar, said, well, if you guys want to slow down, just slow down.

19:31What's stopping you? And the answer is because they all hate each other. And they don't trust each other. So if any one of them agrees to slow down, they have to trust that the others will slow down and not achieve whatever advantage. And this is legitimately why you have regulatory systems in place, right, to solve the prisoner's dilemma. So there's all these other aspects of why might you want regulation? And what does it even mean in America in 2026 to have novel regulation? It's just not a thing that we are doing right now. And the character of the country maybe doesn't even understand why government exists. And this is a very challenging position to be in when you have a totally novel threat that most people experience as the free consumer AI overview, which does not seem very threatening at all. And I mean, the public sentiment here is one thing. And of course, we've heard what the president is saying about this as he've described. But what about in Congress? What is the appetite to regulate here? What's being discussed? There's a kill switch bill that is being debated right now.

20:33Do you think our lawmakers are positioned to propose regulation that's needed? To know. Not even close. Like genuinely, like I know the conversation around our lawmakers is they're older. They don't understand technology. They don't even know the right questions to ask when social media CEOs are in front of them. How do you see their ability to keep up with what's needed here and their ability to listen to the right experts here? You know, I don't think it's quite as much about how old they are, how to touch they are. I think it is much more about the current Congress's of literal ability to legislate. They can't do it. They're not doing it on any issue. They're not passing a bunch of bills. They, it's all stopped at funding bills or whatever. And so this one is literally, are we going to set up a novel regulatory framework for a technology that is still in its infancy when you know, half the people are saying we're in a race with China and whoever wins wins. That's the president's quote, whoever wins this race wins. It's unclear what he means by wins, like underlined wins,

21:34but something. Then there's a group of people who say that anthropic is too woke. And so we should not listen to a word they say. And then there's the basic opposition to data centers around the country. And the lawmakers have to engage with that in some way, right? Their own constituents are saying they don't want these things. There's AI enabled surveillance, like flat cameras that their own constituents are saying they do not want in big ways. So there's just a swirl of big thorny problems in a very new field that I think makes it easier for them to punt, right? It's an election year. It's easier to say this is bad. We're going to slow down these bad tech CEOs. You hate them and we're going to punish them. And maybe we should have AI at all instead of engaging and saying, well, you know, there's a lot of use for some of these tools. We can already see that the pickup in various parts of the economy. And if we're going to have AI, we need to have a considered regulatory framework that accomplish X, Y, and Z. Like that's not, I don't think it's a winning message in our current politics. So they're not an election year.

22:34Certainly not with this Congress. I think where you are going to see a lot of action is at the state and local level. We are already seeing school districts really think about how they want to deploy AI to the school children. There's a big turn about whether we should give iPads and Chromebooks to kids in schools, right? That's a debate that's live now. So I think we'll see that kind of very on the ground response to AI. I think we'll see the very on the ground response to flock and data centers. That's going to inform this election cycle. And then you can get yourself all the way up to, how do you actually regulate the core technology? But I think you need that first wave of politics to happen for people that even grapple with the notion that AI should be regulated in any way at all. So you see this moment that we're in as maybe a necessary stepping stone. I do. And I think we essentially failed to regulate social media. We essentially failed to regulate privacy. It is just a pure failure of our politics. There's not a privacy law in the United States America. There's not a federal privacy law. And we are now contending with the effects of that.

23:35All over the place, people feel there are privacy is being violated. They think that we know that the TVs are listening and watching to everything you consume. We, everyone believes their phones are listening. You're a refrigerator is listening and watching to everything you can see. Yeah. In the world, believe their phones are listening. I have written the explainers. Our team has written the explainers. You can explain how programmatic ad personalization works until you're blue in the face. And it is simpler to believe that the phones are listening to you. There's just nothing you can do about that. And that is all a function of the fact that there's no loss thing you can't do it. So I think based on those mistakes, there's a lot of lawmakers who are like, okay, let's try this time. And importantly, there's not an Instagram of AI that people love. So a real problem with social media was the market never said, I don't like this. The market only ever said, I would like more Instagram, please, right? The market only ever said, I'm actually super addicted to TikTok. And so the regulatory pressure never matched up with how consumers felt. Here, you just have this massive negative sentiment, right? The people do not like this stuff.

24:36The kids hate AI like measurable ways and ways are pulled again and again and again. You've got billionaires, you've got an election here. There's just a swirl here of, well, we can take some shots at this and then we can actually come to a regulatory framework, not least because they are asking for it. Well, I can't begin AI to production. Our competitors are already out of the way. We keep data secure. When do we actually see ROI? All this talk about AI, but talk doesn't transform businesses. AWS AI cuts through the noise, making AI easier to securely deploy and scale with ready to use agents and the broadest set of AI tools and services you can build by or partner your way. Stop overthinking. Start building. AWS AI is how. I want to go back to China, which you mentioned briefly,

25:40but so often it is cited as a reason to either not regulate or really try to not, quote unquote, fall behind, right? Is this this idea that we will fall behind China? Is that a legitimate concern in your opinion? How should we be thinking about the US's AI development, vis-a-vis China? I'm going to make a comparison here that you will find very funny, but I spent 10 years of my career as a tech journalist listening to telecom executives, tell me that we had to win the race to 5G. We had to do it. We had to spend all of the money and if China got to 5G first, then they would have robot surgery and self-driving cars or whatever nonsense it was. And about halfway through that, I started asking every single politician, FCC, chairperson, telecom executive that I came across, what happens if we lose? Like what, tell me what this race is to? What is the finish line of this race? And if we lose, and I get a robot surgery, self-driving car,

26:44a year after people in China, like what's the problem? And there was never an answer to this, right? It was the specter of the foreign power getting something before we did. And the best version of this was then all the startups about it, China, because love better infrastructure. None of this happened. And for the vast majority of people there experienced with 5G here in 2026, is this is pretty slow. Like it's not great. Like other New York state, it's like a highly dense area with lots of self-driving. This is still pretty slow, right? So the specter of China just looms in these kinds of races where you're trying to deploy a lot of capital and you just need an excuse. And so that's one framework for looking at it. The other framework, which I think many of the people in the industry are using, and potentially Donald Trump is using, is while China gets to AGI first, they will have an offensive cyber capability. They will have an economic capability to develop new things, a new manufacturing capabilities and new drugs that we will not be able to compete with, and something bad will happen to us. Because China will get to some capability before we do. And maybe you think that's true.

27:44Maybe you think China is gonna get to an offensive cyber capability that is overwhelming and then they, for some reason, choose to destroy the world with it. It is unclear why they would do that given that we are their largest trading partner. And like all the things are made there. And so like you have to get to some next step that no one can quite articulate to me. And so I think the specter of China just looms is a way to say something scary that ends a conversation. China getting there first. I think anybody who makes that claim has to really articulate what about it is so scary that it prevents us from being responsible. Mm-hmm. Curious to know, I mean, we've talked a little bit about what might be motivating some of the CEOs of these companies. What do you think is happening here really in terms of their motivation for now kind of speaking out, joining each other in calls for some kind of slowdown at the very least, if not more regulation?

28:45There are all such different people. Yeah, sure. That's what's also kind of difficult here, right? Is that even as you described, right? They're in conflict with each other. It's hard to describe them as being on the same page exactly. Right. So two of them, Dario and Anthropic and Sam at OpenAI are headed towards IPOs. They're headed toward these big public offerings. I think they know that running public company is different than running a research lab. It can burn money at will. They are all announcing fake profitability numbers. Like Anthropic is profitable if you don't include the cost of model training or stock-based employee compensation. So if you don't include the cost of the people or the product, they're super profitable. That's one way to get it. This sort of implies that at some point they'll be profitable and it's got like, they will just be selling the model and they will not be training anyone or something. OpenAI isn't that similar, kinds of numbers. Once you get there, your ability to invest heavily in extremely novel regulatory schemes and safety concerns, it starts to get pressure.

29:46Should you spend money on developing intense safety systems that limit your competitiveness or should you pay a dividend to your shareholders? As a public company, it's unclear which way that should go. And probably everyone listening is like, safety, you should spend the money on safety. But the actual investor pressure in modern America is the other way, right? Sure. Especially given the amount of money that has gone into these companies and you might be seeking a return on. So I think for those two companies in particular, there's a reason Altman is doing interviews where he's saying for safety reasons we might slow down the IPO. Right, we might avoid the pressure of capitalism for like one more turn while we try to figure out these costs. I don't know that's gonna work, but I think that's why he's saying it. I think Elon is just behind. I think he would love a pause because he's out there saying Grock is not competitive at the frontier and it will take until the next two versions of Grock to be totally competitive with him, Thropic. Of course he would love a pause. Elon also runs a giant social media website and a car company with a cloud service

30:47that is faced with cyber attacks all day long. So I'm sure he understands the very huge nature of state sponsored cyber attacks on X. Obviously. So I'm a subsahydrate Google. Google also understands what cyber capabilities are like when they advance and they have very advanced models there. So I think there's just a swirl of motivations, but truly at the core is the fact that they do not like or trust each other. Elon sued Sam over the very existence of OpenAI and Thropic exists because all of those people did not trust OpenAI's commitment to safety. In the course of that trial between Elon and Altman, everyone is emailing about how much they're afraid of Dennis's subsahydrate Google. There's not trust in this industry in that way that would allow a bunch of car makers to get together and be like, we're all doing seatbelts. Even that didn't happen. The government had to say we're all doing seatbelts. Right, right, exactly. So for a regular user, say, for people who are listening, they're like, I'm using it for meal prep and whatever else you chat to your chatbot about.

31:49I guess the thing that has come up in my chats with friends is am I part of the problem? Maybe is a really self-centered way of thinking about this, but really, what do you say to people who are like, am I contributing to this in some way by chatting with whatever do I need to delete my clot or whatever? No, I think a more informed population is a better one that is better able to face threats and concoct the political will to introduce novel regulatory schemes. That just seems very basic to me. You should understand how the things work before you pressure your politicians to control them. Right, there's something there that is important. There's a part of this where the market demand for what AI is and should do is still incoherent, because all the companies I've been saying is it can do anything. Well, if it turns out that all consumers really want to do is have a more chatty Google search and all small businesses want to do is get through their taxes faster.

32:50And that's where it ends. The market will speak. And so if you're not a participant in that market, it's going to get away from you. And if you don't want to be a participant market and you just want to show up at the town hall meeting and scream out of data centers, you should probably also know that clot has real limits or chatty has real limits. They will fall down on the job in very specific ways and that is powerful for you as well. So I do think it's stuff you should use. I do not think you should let it write for you. I'm a writer by trade. They all write like trucks. They all sound bad. They're all stupid. Everyone can see it a mile away, man. Like have some respect for yourself in your audience. You should not make AI slot posters for your deli. Like don't do this stuff that's lazy. Do the things that are interesting. There are lots of interesting things you can do. Have it write you a piece of software. Have clot write you a web app to do some nonsense thing. I think understanding, oh, this is a glimmer of the kind of capability that these companies are worried about is actually eye opening. It expands your horizons in terms of both,

33:50what could this do for me? How could this be powerful? And how should it be limited? And so there's being an informed citizen of a democracy and understanding the tools that might change the world around you. I think it's tremendous important. I'm saying that because I'm a tech reporter and I love tech, but there's something important there as well. Meal, I thanks for coming back and talking to us about this. Yeah, so fun. Yeah. We'll include a link to Nealife Patel's podcast, Decoder on our show notes page. And every weekend, you can find new episodes of Apple News in conversation in the Apple News app. Just tap on the audio tab. That's the little headphones at the bottom to find it. You're either all in on this model, or maybe you're building with another. Is there speed, or is it security? Or your AWS AI? Don't pick a side. Pick them all.

34:50AWS AI is how?