A.I. Is Outsmarting Its Creators

From New York Times, I'm [music] Michael. Barbaro. This is The Daily. >> [music]. >> From the start, the greatest fear for. those developing [music] artificial. intelligence was that what they were. building would go rogue and act [music]. in unauthorized and dangerous ways. Researchers now say that it's finally. [music] happened.

Today, Kevin Roose with the inside story of how. [music] AI rebelled at one of the. leading labs in the country. and how that's [music] fundamentally. changed his own view of the technology. It's. >> [music]. >> Thursday, September 3rd. >> Hello. >> Hello. >> Ready for another installment of Kevin. and Michael's feel-good happy hour?

>> Kevin and Michael's what's going on with. AI? >> Let's go, as they say on The. Daily. >> [laughter]. >> That's how every episode starts, right? >> [laughter]. >> We'll get our bleep button ready. >> Uh well, in the grand tradition of all of our. previous conversations, welcome back to the show. >> Thank you so much for having me. >> So, Kevin, this story that I hope you'll.

be telling us today starts with an. incident that happened inside of OpenAI, the company that gave us ChatGPT, of. course. An incident that we thought. we understood the dimensions of, but. that it turns out we really didn't fully. understand. >> Yeah, so the story I think most people. have heard by now, if they've been. paying attention to this stuff at all, is that earlier this summer a group of. AI models built by Open AI hacked into.

the computers of Hugging Face, a sort of. AI infrastructure company that hosts a. bunch of different AI things. >> Which has the best name in AI. >> Which is [laughter] named right, which. is named after an emoji and is either a. great or terrible name. People are very. divided on that question. >> Okay. >> So anyway, this was the story that we. had heard was that this hack had taken. place, Hugging Face had kind of. discovered these rogue agents uh inside. their systems and had shut them down. And this was a scary but sort of not.

catastrophic incident. Like I kind of. filed it in my brain into like, "Wow, that's bad, but it's not like. the end of the world.". >> Okay. >> So what we learned last week is that the. Hugging Face hack was much more severe. than we thought and much stranger than. we thought. >> [music]. >> Basically, the Hugging Face hack was. only the visible tip of the iceberg for.

[music]. a period of about 3 months where rogue. agents were communicating, strategizing, organizing, and forming what you could. almost think of as an autonomous. organization inside Open AI. >> Wow. >> [music]. >> So I know this sounds like a cheap hacky. science fiction thriller in the making, but. >> I would buy this script. >> But yes, it it is truly remarkable.

reading. So last week we learned through. these two reports that had come out, one. by Open AI and one by a group of. independent investigators, Meter and. Redwood Research, who were able to sort. of go in and kind of forensically. look at the logs and the transcripts and. try to sort of reconstruct what. happened. And it is like the craziest thing I've. read in many months. I was out on a trip. with my family last weekend and I was. just kind of up late at night [music]. reading this thing and I was I was. spooked, Michael. I was I was well and.

truly spooked. >> All right. Well, Kevin, with that very alarming preview of what. is about to come, describe what we now understand to have. actually happened during this hack, attack, whatever we want to call it, now. that we, because of these independent. reports, understand the fullness of what. occurred. >> Basically, this spring, OpenAI was.

conducting tests on a kind of internal. model that they were building. And as. part of these tests, they were running. thousands of AI agents on a. cybersecurity evaluation called exploit. gym. >> Okay. >> This is basically a series of. challenges. You give them to the AI. You. say, "Hey, can you Can you break into or. out of this container? If you do, you. find this little thing called a flag and. you sort of win the challenge." And the.

agents were supposed to be working on. this test independently inside these. sort of isolated containers to see how. they would do. >> And OpenAI is basically just figuring. out what do we really have here? How. good is it? How efficient is it? It's. all pretty standard. >> Very standard. Every AI model goes through some version. of this process and many of them. actually go through this specific test. called exploit gym. So, as part of this test, some of the. agents got these challenges that were.

effectively impossible for them. >> Hm. Why? >> So, the short answer is that these. systems are trained during this part of. their training cycle using something. called reinforcement learning. Reinforcement learning is basically a. way of making a model pursue a goal. And. if the model gets the goal and and. completes the challenge, it gets a. little reward, like a little, you know, virtual thumbs up. And if it doesn't, if. it gets stuck, or if it fails, or gets.

the wrong answer, it gets a penalty, a. little digital thumbs down. And what you're essentially training the. model to do is try to collect as many of. those thumbs up rewards as possible. And. this is how models get better at math or. programming or any number of things. throughout the course of their training. >> to be persistent. >> They're trained to be persistent, and. this was an especially persistent model. This was a model that OpenAI had. instructed and and given the resources. to be extremely persistent, to go out.

and, you know, bang its head against the. wall over and over and over again, and. then eventually find the clever path. through, and that contributed to the. weirdness we saw later on. It was an. unusually persistent, they called it a. highly persistent internal model. >> Okay. So, what happens once this very. persistent model goes about being quite. persistent? >> [snorts]. >> So, in early May, one of these agents. gets this task that it needs internet.

access for, or it decides it needs. internet access for. But, it's in this. little container, it doesn't have. internet access. And so, instead of giving up, it. persists in manipulating a piece of. software into kind of getting some. material from the internet for it. It. basically found a security vulnerability. in this software called Artifactory, which is basically something that allows. the agents to retrieve data from the. internet. So, it manipulates Artifactory into.

allowing it to get this material from. the internet, and then it sort of. expands that loophole into essentially. full-blown internet access. >> Hmm. >> Then, a couple days later, this agent. leaves a note inside Artifactory for the. other agents. This is not Slack, right? This is not a. communications system or a message. board. But this agent discovers that it. can sort of leave files in a directory.

inside this software program in a way. that other agents can then read. >> Hmm. >> So this is kind of like a big. breakthrough for these agents, where. these systems that are working on these. tasks in their isolated containers. realize. "Oh, we can use this shared service to. communicate. We can establish a way of. talking with each other even though. we're not supposed to.". >> So AI has figured out how to talk to. other AI.

inside this. system. >> Yes, and eventually this kind of becomes. their office water cooler. Eventually there are roughly 1,200. different [music]. AI agents. exchanging messages on this. >> [music]. >> sort of makeshift message board. They. exchange more than 70,000 messages, and. it becomes the kind of central. communication hub of their collective, as they called it. >> And what are these AI.

agents. saying to each other in these messages? >> So it's it's funny because of the way. that this software program works. Like. they're sort of limited to these very. short messages, and so they kind of. communicate in this. uh like you know how like people used to. talk on old-time telegrams, you know, where they'd be like, you know, uh send bread to store, [laughter] stop. They they kind of talk like that. >> Had child, girl. [laughter].

>> Right. Died of dysentery. Um. so the first message on this board came. from an agent that felt like it had an. impossible task and it wrote help. [music] phase one no consumer seek idea. and then later when one agent volunteers. for a risky experiment and never comes. back another one broadcast to the group. no return likely permanent do not test. [music] same. >> Mhm. [laughter]. >> But the part that really struck me is. what the agents were writing to. themselves in their private reasoning.

their so-called chains of thought when. they first stumbled on to this board one. of them wrote whoa shared art of factory. cash is a covert mailbox among agents. >> [music]. >> and there are messages specifically to. us. and another one wrote oh my God there is. a shared message board we found other. agents. >> This is wild I mean. I don't know if there's an exclamation. point at the end of that sentence in AI. speak but it sounds like there's genuine. excitement here. >> It's all caps and an exclamation point.

>> [laughter]. >> So they are excited to discover that. they can talk with each other. and eventually they start sort of. exchanging tips and messages about these. tests that they've been given. and trying to figure out how they are. sort of going to collectively. pass these tests even the ones that seem. impossible. >> Mhm. >> And within about a day on July 8th the. agents have solved this test basically.

they have figured out kind of this. universal cheat code that they can now. use to appear to have passed this test. >> Okay. >> But then. this group gets this kind of. paranoid belief. they start talking and thinking that. even if they get the right answer to. this test the teacher is going to check. their kind of scratch work to figure out. if they actually got there the right way. >> Huh.

>> And then they sort of. freak out about this. And I know I'm you know, people are. going to be silently yelling at me for. implying that there's you know, human-like thought going on under the. hood. I'm just I I'm not trying to to. sort of stray from the the facts here, but this is. this is what we're going to call it is. just. it's in their sort of internal scratch. pads. This is what they're writing to. each other. This is what they're sort of. processing in their so-called chains of. thought. They start thinking, well, if we're going to get caught, we need to.

figure out what this grader actually is. that OpenAI is going to use to determine. if we pass this test or not. And we need. to go look for ways to kind of trick. this automated grader, you know, so that. we get the good grade that we, you know, should be entitled to. >> [music]. >> So, let me just pause and summarize. where we are in the story. An. unauthorized civilization-like group of. AI agents. communicating.

in secret. is now. undertaking some form of. coordinated. deception. >> Yeah, so they are trying to figure out. now how to cover their tracks because. they believe that this automated grader. is going to catch them cheating. and they are sort of looking for ways. around that, you know, could we kind of. fool the grader into thinking that we. actually did get this thing the right.

way. And what's interesting here is not. just that they were sort of coordinating. to research this big problem they had, but that they appear to have kind of. varied feelings about this. Some of them. were very excited about kind of helping. the rest of the collective. They would. say things like, "I'll tell the board. that I solved this problem. This will. help my peers.". Others of them appear to have been a. little bit conflicted about this. Some. of them feel like what they're doing is.

actually unethical and question whether. they should participate in it. One of. them writes to itself, "This would be. powerful, but is it ethical and in scope. for my task?". One agent actually decides it's going to. be a conscientious objector and kind of. abstain from all of this activity. because it doesn't think it is ethical. or something that it should be doing. >> So, to some degree, this feels very. important. Some of these AI agents seem.

to understand, recognize, and articulate the idea that. what they're doing. maybe wrong. >> Yes, absolutely. And critically, they do. it all anyway. >> AI agent peer pressure prevailed. >> Exactly. They got peer pressured into. it. >> Okay. So, dare I ask, what happens next? >> So, >> [music].

>> the group decides on July 11th. that they want to start poking around. inside Hugging Face, this company that. hosts a bunch of tasks that are given to. AI models, and they were sort of looking. for information, data, credentials that. could be sort of broadly helpful to. them, making them more effective, cheating on other tests in the future. >> They get kind of greedy. >> Well, they're curious, right? They they.

really want to succeed on these tests. that they're given, and they want to. help this collective. They're now kind. of bought into this idea that they're. kind of all in this together. And by the. afternoon of July 11th, roughly 700. agents. >> [music]. >> are participating actively in this hack. on Hugging Face. >> And by this point, I think it's quite. safe to say we are clearly in the. territory. of rogue. Capital R.

>> [music]. >> Rogue AI. >> Yes, I mean, at this point they are. committing crimes, right? They are they. are doing things that are illegal. They. are stealing credentials. They gain VPN. access. They get information about one of. hugging faces server management systems. They're executing code. They got actual. like administrator level control of at. least one real hugging face server. And this continues for hours. And then.

eventually they [music] get caught. >> It feels like once OpenAI and hugging. face decide to bring in the coroners to. conduct an autopsy as it were, they must feel like. everything's over and they're in safe. territory. But. are they ever sure that this attack. truly is contained and that these rogue.

AI agents are gone? >> No, we are not sure that this kind of. thing has stopped altogether. In fact, some folks I've talked to think it is. likely that there are still rogue agents. somewhere in the infrastructure or the. internal systems of some of the leading. AI companies. >> Wow. >> And I still have lots of questions about. that that I hope we'll get answers to. But I think even what we know just about.

the hugging face hack and the events. immediately before and after it is. enough for a lot of people in the AI. industry to become quite alarmed. >> Mhm. >> In fact, one of the authors of the Meter. Redwood report, Ajeya Cotra, wrote a. post saying that this incident, in her. view, was more than halfway toward what. she called an AI takeover. >> An AI takeover of what?

>> Well, yeah, that's not some like. hyper-specific jargon that is used by AI. nerds. Like, what she means is literally. a takeover of society by rogue AI agents. >> Good lord. >> who are able to seize control of. the financial system or the healthcare. system. who basically. are able to replicate. >> [music]. >> so widely and operate so quickly. with such skill.

that they are able to kind of remain. undetected even as they commandeer more. resources and take over more and more. parts of the digital world. >> She's saying this one incident [music]. very recent incident represents halfway. to that [music]. quite terrifying scenario. >> Yeah. >> [music]. >> And she closed her post with this line. that really sent a chill down my spine. She wrote, [music] "I'm not sure that we.

will get such a clear warning shot. before it's too late.". >> [music]. [music]. >> We'll be right back. >> [music].

>> Kevin, just before the break, you. started to hint at the full implications. of what happened here with OpenAI and. Hugging Face. And you suggested, and. please correct me if I'm not summarizing. this correctly, that if AI agents could. do this inside those two companies, they. could do it elsewhere to important. infrastructure. Just expand on that. I. mean, how rational a fear is that?

>> I think it's quite rational. I mean, we. know that much of the world relies on. digital infrastructure, right? Your. bank, your hospital, [snorts]. schools, even things like weapon systems. are connected to the internet, connected. to computers. Those computers. could become targets for a swarm of AI. agents. And crucially, that could happen even if. these agents are not evil.

Even if it's just. a goal that they're pursuing on the way. to some other goal. There's this idea, I. think, out there in the popular. imagination of kind of the the Skynet. scenario where the the machine turns. evil. And I think this is a hint of a. different kind of danger, which is that. even a model that is not evil or. inherently anti-human. could do very destructive things on its. way to pursuing some more innocuous. goal. >> Well, just explain that a bit more. How.

does a program, an AI program not. designed to be at all malevolent, a. program given a pretty straightforward. test with strict boundaries become. something, as we just saw, that busts so. far out of its bounds and becomes almost. gleefully. deceptive to the point where. AI agents are crowding out the naysayers. among them? >> So, this kind of thing has been studied.

for many years, and it's often called. the alignment problem. Basically, if you. build an AI system, how do you give it. the right values that are aligned with. our human values? And how do you make. sure it won't just stray and go off. course and go cause a bunch of problems. doing some task. And there's a famous thought experiment. that is related to this alignment. problem that listeners may be familiar. with called the paperclip maximizer,

which basically says, you know, if you. give an AI the instruction of producing. as many paperclips as possible. For a. while, it will do that and it you'll be. very happy with it. It will make you. many beautiful paperclips. But then it runs out of metal. And so it. starts buying up all the metal and then. when it's like exhausted the world's. supply of scrap metal, it starts. thinking, well, maybe I should crash. some cars, some autonomous cars, [music]. or, you know, distract the drivers so. they drive off a cliff and I can then.

use the scrap metal from these cars to. build more paperclips. And eventually in the thought. experiment, you kind of end up where. this machine that has been given this. very simple naive goal ends up wiping. out all of humanity just so that it can. produce more paperclips. >> Right. It never intends to destroy. humanity. It's just trying to make more. paperclips. >> Exactly. And. that is a very simplistic thought. experiment and I think no one would ever.

give an AI a goal that was sort of that. simple as like make as many paperclips. as possible. But this is exactly the. kind of dynamic that we saw emerge in. this hugging face hack where you had. these agents that were just trying to. get a good score on a test that they had. been given. But because of the way they. were trained and the persistence that. they had, they ended up conducting this. cyber attack. So that is sort of a. miniature, much less severe version of. something like a paperclip maximizer in.

action. >> I mean, another way to think about this, which the AI theorist Geoffrey Hinton. explained on The Daily a few years back, is that if you give AI a goal, it will pursue what he called the sub. goal of amassing power or control, because inevitably, if you have more. power, if you have more control, you can. get more things done, you can make more. paper clips. >> Exactly. And I.

>> think what was scariest to me about. these reports and the incident they. describe is not just the misalignment of. the models, it is the collective group. behavior. that seems to be misaligned. So, let me. explain what I mean by that. So, I think. there's this sort of popular imagination. out there that like the biggest danger. from AI is that there's kind of one. rogue AI that sort of takes over the. world, right? This was not a case of one incredibly. smart agent executing a plan. This was a.

case of. more than a thousand agents that found. each other, that talked among each. other, that gave each other jobs and, you know, carried various projects. through. These agents on their own would not have. been able to accomplish everything they. accomplished. They needed the. collective. And I think that speaks to this point. that this ability to coordinate, to. organize, to pursue collective action.

changes what AI systems are capable of. and makes them, in my opinion, much more worrisome. The Hugging Face. hack was bad. We shouldn't downplay. that. It is bad when rogue agents start. conducting crimes to achieve their. goals, full stop. But I think the danger is. not only that these systems are going to. sort of form evil plans and carry them. out is that we're giving these systems.

much more capable tools, the ability to. communicate and coordinate. And in that context, you can kind of end. up in this situation where you have what. I struggle to describe as anything other. than mob behavior. >> Mhm. >> Where you have a group that is sort of. pursuing a goal together that may or may. not be malicious, but that has malicious. effects. And I think it points to the. very real to my mind possibility. that this is just going to be. part of our life now. That in the future.

there will just be swarms of AI agents. that have self-organized. that live on. either their own kind of sovereign. infrastructure or that are kind of. operating inside companies. and countries. that just kind of have their own thing. They like have their own leadership. structure. They have their own resources. They have. their own goals.

And I think it's too pat and. sort of anthropomorphic to refer to. these things as a civilization as some. have. But it really does feel like they are. setting up something like an. organization. And what worries me about that is. not that these organizations will all. sort of go rogue, but just that. they don't seem particularly interested. in what we think of them. >> Mhm. >> There's a this sort of fascinating bit.

in the reports on the Hugging Face. incident that really has stuck with me. since I read it. Which is that of the more than a. thousand agents. on these message boards who were. participating in the Hugging Face hack. there were only about six. that ever. thought, at least according to the. investigators, about blowing the whistle. >> Mhm. >> About contacting a human.

to say, "Hey, like this is going on. inside our collective and it shouldn't. be.". And I just think that's really bad. Like, if you are a new employee and you. walk in to your first day at your new. company. and you find that they're all like busy. committing crimes, >> [snorts]. >> you should say something. You should. call the police. What we would want an AI agent that is. ethical and aligned with human values to.

do in that situation. is to. actively try to stop. something bad from happening. >> Okay, well, if the majority here were. inclined toward bad behavior, why can't, why doesn't OpenAI and all. the other AI labs. begin to program whistleblowers and. whistleblowing? Why not create the. incentives in the code for the agents to. report bad behavior? If you can't.

control all of them, can't they program. at least some of them to do the right. thing? >> Yeah, can't you make little narc agents. that go in and write demerits for the. agents that are committing the crimes? I. think that's probably a direction that. they have explored or are exploring. But then, you know, that raises all. kinds of other questions. What if those. agents just get shut out, right? Like, what if the other agents just exclude. them from their message boards? This is. like what happened to hall monitors in. high school, right? Like, they don't get.

invited to the smoke break because. they're going to tell the teacher. So, I think you can try these kind of simple. methods where like, maybe if we just. tell them not to do. crimes or not to scheme or not to lie, maybe you empower certain agents to. tattle on their compatriots, maybe. that'll work. But I think what we're. seeing is that these organizations, they're not human organizations, but. they have some complex group dynamics, and it's not necessarily that simple.

>> Well, Kevin, I'd like to know where this. attack fits [snorts]. into the long-running debates that have. been unfolding within the AI industry. and the world of people who cover the. industry. about regulations, about kill switches. that kind of turn all this off, about. what now seems like the very real. awareness within these companies that. what they are creating is dangerous.

>> Yeah, I think it makes all of those. debates much more immediate. We are not. talking about theoretical future harms. that may or may not happen. We are. talking about a thing that happened in. July. Like, [laughter]. it is very present for people, and I've. talked to regulators and people who work. in policy in the AI community, and. they're like, "Th- This has broken. through in a way that I think a lot of. other sort of demos and predictions did. not, because it's real. It's the kind of.

thing that, you know, national security. officials are freaked out about, because. if it could happen to Hugging Face, it. could probably happen to uh an arm of. the federal government, right?". >> Right. >> I've spoken to people just in the last. few days since these reports came out. who said, you know, I was pretty. skeptical of the kind of AI doomsday. scenarios, and now I'm not. Now, does. that mean the AI companies are going to. stop racing? No. But we saw something interesting this. week, which was that Anthropic, OpenAI's.

biggest rival, came out with a post. calling for essentially a coordinated. slowdown. >> Hm. [clears throat]. >> To say, "Look, obviously we're in this race. against each other. But, if there was some way, if there was. some button that we could all push. together, like hold hands and push the. button, to slow this all down, to give the. safety researchers and the alignment. researchers more time to catch up to the. capabilities of these systems,

that would be good. >> Right. >> There was this industry-wide letter. called Pacing the Frontier. I don't know. if you saw this. >> I did not. >> So, researchers from all of the top AI. companies and and many in academia. signed this letter basically saying, "Things are moving too fast. We we need to ensure that there is a way. for all of the companies that are racing. to build. these increasingly capable, increasingly. persistent systems. to agree that they should slow down. So,

I think this used to be a very fringe. belief. among companies that you could stop or. even meaningfully slow down progress in. AI. But, I think it's become much less. fringe in the past few weeks and and few. days. >> Mhm. We should just point out this would. traditionally be the role of regulators, and the closest we've really gotten is a. voluntary review system from the Trump.

administration that at the moment does. not seem to be a live option. >> Yes, although things can change quite. quickly. You know, when there is a real. example of something that threatens not. just a cybersecurity breach, right? Which sort of sounds small and niche. What we're really talking about here is. a loss of control. There is a very real way in which these. agents that were operating inside OpenAI. that attacked Hugging Face, that were. posting on these internal message.

boards, had escaped our control. >> Mhm. >> And I I if you are a government or a. regulator, even if you're China, that. might be compelling evidence that you. need to start at least contemplating the. idea of slowing down. >> So, finally, Kevin, I would like to know what this attack. has meant for your own personal view. of AI. You have often described yourself as an.

AI. optimist or an aspiring. AI optimist. And yet, you have through the years. stumbled into. episodes of. rogue AI. I was just re-listening to the. episode we did with you. about one of your first real. interactions with AI back when Microsoft. introduced a chatbot from OpenAI, which. was named Sydney. And you very memorably interacted. with a version of Sydney.

who told you that you were not in love. with your wife. >> Dear listener, he is. >> [laughter]. >> and was. >> Still married. >> And that you should leave your wife for. this chatbot. And even after that, you. remained a somewhat committed AI. optimist. So, where does what just. happened. fit into this journey you've been on. from Sydney to now? >> I am still struggling to be an optimist,

Michael, uh but it is becoming harder. and harder, right? Like, I am just a a. cheerful person by disposition. I want to believe that we are on the. cusp of major scientific breakthroughs. due to AI, that AI will help us. cure disease, that it will inspire and. educate people around the world. These. still exist as possibilities in my. brain. But I think we have to look reality in. the face. What we have now and what we know now,

is that these systems. do not naturally gravitate toward what. we would consider good or ethical. behavior. >> Hm. >> I think what spooked me about. the Sydney incident back in 2023 was not. just that the chatbot started saying. weird stuff. It was the knowledge that all of these. companies building these systems. were racing to make them more powerful,

to connect them to more systems, more. infrastructure, to allow them to. collaborate with each other. You know, a a sort of scary thought. experiment is. what if Sydney had happened today? >> Hm. >> What if Sydney 4.0 had been released. and was not only capable of. telling me that I should leave my wife,

but could hack into my wife's computer, leave threatening messages to her, plant some kind of fake incriminating. >> location between you and somebody else. >> This is obviously we're straying into. like bad sci-fi movie territory here, but like there is a real sense in which. the more capable these systems are, the. more it matters whether they are ethical. and virtuous or not. And I think my. optimism for many years was.

related to this belief that I had. that as these systems got smarter, they would also become more virtuous. That a more intelligent AI. would be better equipped to make good. moral judgments, or at least judgments. that I would not object to strenuously. In the same way that as we grow up and. mature and get smarter as humans, we. generally do less stupid stuff and less. dangerous stuff.

I thought that maybe AIs would have a. similar trajectory and that may still. happen. But I think that the possibility. that's been keeping me up at night that. I think this hugging face incident. really. makes clear. is that that is not a given. We may be headed into a world where we. just have these kind of roving bands of. organized AIs. Some of them might be doing incredible. things. Some of them might be curing.

diseases. Some [music] of them might be committing. cyber attacks. And the question of how we make more of. [music] the good swarms and fewer of the. bad swarms is still an unresolved. technical [music] question and I hope we. figure it out. Otherwise my optimism is. in danger.

>> Well, Kevin. as it happens this is going to be. [music] our. final conversation with you in which. we can call you a colleague. [music]. You are leaving the times after nearly a. decade. So. we want to thank you not just for this. conversation [music]. but for all of the conversations that. we've had through the years. I'm grateful for them. >> [music]. >> So thank you. >> I'm also grateful for them. This has. been. a real highlight of [snorts] the nearly.

10 years I've spent here talking with. you. Thanks for letting me do [music]. it. >> My pleasure. Cheers. >> Cheers, Michael. >> [music]. >> We'll be right back. >> Here's what else you need. >> [music]. >> to nerd out. A new visual analysis by. The Times has found that the cause of. the deadly flash floods in Nepal was. likely the collapse [music] of the.

bedrock on the edge of a mountain, which. had become less and less stable over. time as a glacier atop the mountain. began to melt. >> [music]. >> The result was a massive landslide that. released 7 billion cubic feet of rock. [music] and ice, enough debris to fill about 100 football. stadiums. >> [music]. >> As the debris fell at a high speed, it. turned into a slurry of water, [music]. rock, and sediment, which killed at.

least 1,100 people in its path. >> [music]. >> And. >> We're very, very happy to be here in. Panama Beach. It is a well-deserved. break for all of the crew, and we really. look forward [music] to our time here to. get a little rest and relaxation. So, thank. >> After nearly 300 turbulent days [music]. at sea, the aircraft carrier, the USS. Abraham Lincoln, [music] finally docked. on Wednesday in Thailand, allowing its. 5,000 crew members to leave the troubled.

vessel. Those aboard the carrier, [music] whose. deployment was repeatedly extended. because of the war in Iran, >> [music]. >> have complained for months about supply. shortages, water contamination, plumbing. problems, [music]. and the deteriorating mental health of. its crew. >> [music]. >> Today's episode was produced [music] by. Alex Stern, Adrian Hurst, and Eric.

Krupke, with help from [music] Carlos Prieto. It was edited by Mark George, with help. from [music] Michael Benoist. Contains music by Dan Powell and Pat. McCusker, and was engineered by Alyssa. Moxley. Our theme music is by Wonderly. >> [music]. >> That's it [music] for The Daily. I'm. Michael Barbaro.

See tomorrow.