Is the new AI model really too dangerous to release?

0:00This BBC podcast is supported by ads outside the UK.

0:30what we call the quick and dirty so it's a recap of the biggest pop culture headlines that week whatever you need an explainer on we have got you and then on Fridays we do dumb stuff Fridays yes we're three times weekly three times weekly search for shameless in your favorite podcast app now

0:49it's almost like you're inoculating yourself against the blue light by just blasting yourself with as much light as you can get, first thing. That's so interesting. Welcome to The Interface, the show that decodes how tech is rewiring your week and your world. I'm Thomas Germain. I'm Nicky Wolf. And I'm Karen Howe. And today on The Interface, we will be asking, why do AI companies want us to fear their products? How, despite what you've heard, the blue light from your phone isn't ruining your sleep. And will we ever really own a piece of tech again?

1:32All right, I'm going to look into a crystal ball because I think I can predict what Karen's story is going to be this week. Yeah. You're kind of easy to read. I'm sorry. What's your guess? Well, there was big news from this AI company called Anthropic. Maybe people have heard of it. It's a competitor open AI about their new model, which is called Mythos. But Karen, you could probably explain this better than I can. Dang it, Tom. You got me again. Oh, well. So Anthropic announced a new model called Claude Mythos Preview. And it is basically their most advanced AI model yet, or so they say. Except we're not really sure because they haven't actually launched and released it to the world because they're saying it is in fact too dangerous to release.

2:24Way too scary. Way too scary. This is, we've created a monster. Lightning has struck the building. Everyone be afraid, be very afraid. It's not an exaggeration to say that they're warning us that this might destroy the world. Their evidence of why this is very scary is that they ran a bunch of tests with Claude Mythos And it supposedly autonomously found thousands of high severity critical vulnerabilities. And this is talking about the model's ability to review code and then find exploits, basically bugs in different software that hackers could then use to compromise the security of that software.

3:10So it can break into everything, essentially. That's what they're claiming. Right. Yeah, exactly. And Anthropic says, you know, if this model were just simply released into the world without all of the preparation that they're going to do moving forward, then it could seriously undermine economies, national security and all that good stuff. If you watch the reaction of both the media as well as a lot of people in the tech world to Anthropik's announcement and also just people, members of the general public, you would assume that we should all be extremely terrified. Like this is something that a lot of people within the AI world have talked about and warned for a long time that as these AI models become more and more advanced in coding specifically, this is something that could happen.

4:00It could develop really sophisticated cyber attack capabilities. And that is going to somehow bring our very digital software, different worlds tumbling down. Right. The problem is we kind of have no way of verifying these claims because Anthropic is simply just telling us about this model. And there is no independent verification. There are no independent benchmarks. And in fact, in announcing that there is this big scary thing that they've developed without releasing it, they are following exactly the same playbook that they did many years ago with a totally different model, which was GPT-2.

4:38Was that before the big Anthropic OpenAI split then? Exactly. So longtime listeners of The Interface will remember that Anthropic and OpenAI, two rival companies, have deep beef with each other. And part of the reason is because the executives of Anthropic currently used to be executives at OpenAI, and then they splintered off and started a competitor. But when they were at OpenAI, they were, in fact, the very same people that orchestrated this big PR campaign around GPT-2, which was the early model that OpenAI developed one and a half generations before chat GPT. And at the time, OpenAI, because of the very same executives, said, we have developed a model that is way too dangerous to release, but we are going to announce the entire world that we have developed it as a safety measure so that people know that this kind of capability could be on the horizon.

5:36And we're going to start working with all these partners in academia and other research spaces to try and shore up and test this model before we actually roll it out. And this is literally what Anthropic is now doing once again with Claude Mythos. So alongside announcing that they had Claude Mythos, they also announced this new initiative called Project Glasswing, which is this collaboration with companies across the tech industry as well as financial companies. This includes AWS, Apple, Google, Microsoft, JPMorgan Chase, Linus Foundation, NVIDIA, so on and so forth. And they mentioned that Anthropic is going to give them this early preview of the model, allow them to use the model behind the scenes.

6:22And collectively, this consortium of companies is going to somehow shore up all of the critical software in the world before this model becomes publicly available. so it's it's hard at some point not to see these ai companies as the boy who cried wolf right every couple of weeks it's not just like that one way back and this time every couple of weeks there is something some headline where an ai executive says our model has just become conscious we should be terrified of that like every single time and a lot of the time the media just reports what they say completely without any kind of critical thinking around it but not today but not well what happened yeah setting setting aside maybe this one is they've built some kind of monster that that can bring down the world and launch all the nukes and whatever at some point people are going to stop believing them right we're past that point right like everyone's heard this before like oh no, please, somebody stop me from making this product that I've been working on for the past five years.

7:33It's going to kill us all. Oh God, I wish someone could tie my hands behind my back so I wouldn't go to work tomorrow. And in the meantime, they're preparing for an IPO. Just for people who might not know, an IPO, initial public offering, is when a company goes from being fully privately owned and private hands to launching on the stock market and allowing people to buy shares in the company. The price that something launches at in an IPO is very important for the value of their company. So they want hype as much as possible for an IPO. Yeah. So they're saying, you know, be afraid of us, be very afraid, but here's the kind of stock price we want to launch out later this year.

8:06Right. We're not necessarily saying that everything that Anthropic is saying is total BS, right? Like this is the problem. There's just no independent verification of the claims. So they could be true and they could be false. So we can't really make claims at this very moment with such limited information about whether or not there really is a step change in the coding capabilities of Claude Mythos that would cause massive security vulnerabilities. But what we are saying is that it is also important context to understand that this is also a PR game. Right. Anthropic could have totally, they could have developed Claude Mythos and they could have just quietly engaged in Project Glasswing, you know, behind the scenes, just giving their model to various companies for preview and just never told the world.

8:59And it's like, why are they in fact announcing it from the rooftops? Well, as Nikki said, they are preparing or eyeing an IPO later this year. And also they literally just had a huge face-off with the Department of War, which we covered in another episode as well, where the Department of War declared anthropic or tried to declare anthropic, a supply chain risk. Ultimately, that was dismissed by the courts. But Anthropic is sort of in a situation where they would do well for themselves if they positioned themselves as a central node within the tech industry and also the financial industry and all connected themselves with all these other companies, made themselves very, very important to all these companies as a kind of shield of protection from potentially other funny businesses that the U.S.

9:49government might try to do to them. I think it gets pretty annoying for people after a certain point. Like you don't hear like McDonald's coming out and being like, oh, like we've made a burger. It's too delicious. Everyone's going to get so fat. We cannot release this burger to the public. It's just a strange way for companies to talk about themselves. But it's because if you make, you know, the AI bomb that's going to destroy the world, it means you've got the best tech in the world. So even though like you're saying we should be terrified of you it's kind of great for anyone who like owns a part of your company it's something we've heard from every single one of the major ai companies since all of this started at the beginning with the launch of chat gpt and it was like the point of open ai in the very beginning right was like they were going to be the good guys they were worried that you know if the big tech companies make ai it's going to be too powerful and destroy everything So open AI, they were going to be the ethical ones, a research lab that was going to release AI in a way that was good for the public.

10:55So Google specifically wouldn't ruin the world. And then they've sort of shifted gears. Now they want to be the one who's releasing the AI to the public. And Anthropic is like biting at their heels, even though they were supposed to be even more concerned about safety. It's just like all this in one sense of big game. But then also this might be true. Right. We don't know. There are cybersecurity experts and hackers, white hat hackers and other kinds of experts have all been in the information sphere commenting on this news. And honestly, it's pretty split between people who think that it is a huge step change and people who are saying actually reading some of the reports that Anthropica has put out thus far.

11:41There's just a few holes in what's going on. And so some of the holes, I mean, first of all, as I mentioned, there's no independent benchmark. So a lot of cybersecurity folks have pointed out they could have, in fact, used some of the well-known tools for detecting exploits as a benchmark to compare whether or not Mythos has exceeded existing tools. But they didn't. So that calls into question, like, what is exactly going on here? They also mentioned, like, they found or Mythos found autonomously thousands of high severity vulnerabilities. But when you look at the fine print, what that basically means is Mythos graded its own homework and Mythos declared that it was high severity.

12:28and when they manually they got a bunch of human experts to manually verify over 150 roughly of those and they said in the report anthropic says in the report we cannot definitively say that these are high severity but among the small pool that the human experts verified the human experts largely agreed with mythos's own conclusion so it's sort of like this strange roundabout like anthropic is creating its own homework and also claude mythos is just creating its own homework it's like what is actually going on here um and then there's two other interesting points that people have raised one is what's actually the false positive rate because when you have a tool that's supposed to detect exploits it's kind of only useful if it detects them very quickly and cheaply but one of the biggest exploits that anthropic has been touting around in all of its blog post cost around$20 ,000 and a thousand runs of Claude Mythos to find.

13:33And so people are saying, like, are people going to be spending tens of thousands of dollars to use this tool? Like a criminal probably doesn't have that kind of budget. Like, although there are major cybercrime operations where, like, if you could hack, you know, one of the biggest companies in the world, it would absolutely be worth$20 ,000. And also if there's something that can find critical vulnerabilities in national infrastructure, there's obviously nation states that would be interested in. Yeah, right. Right, right. North Korea. Yeah. And then the last kind of skepticism that people are raising is that open source models are kind of already close to catching up.

14:14So when Anthropic announces that this is a new skill that only they have and they're going to keep really close to the chest before it gets unleashed onto the world, is that really true? And interestingly, when it came to GPT-2, that was actually the thing that led OpenAI to ultimately release GPT-2 anyway. Because a few months after they announced this big thing saying it's too dangerous, an open source project immediately replicated GPT-2 and the world didn't fall apart. So OpenAI just released the model. Arguably, the world did fall apart, but not for those reasons. Right. Yeah. I mean, the world did.

14:52I fell apart personally. So I think the big question for people, if you've heard about this, is like, should you be freaking out or not? Are you guys worried about this? I mean, I got I got to be honest. There's like a part of me that is a little bit worried. And this this goes to show how powerful these mythologies are, because on one hand, they've cried wolf so many times. And on the other hand, one day it could, in fact, be more legitimate, like they could make more legitimate claims. There could be a more legitimate step change. And so these models are computational technologies. Coding is a very computational task.

15:34Like this is one of the dimensions where I could see it adding up. And that's kind of why it's like, great, we should have had some kind of setup where we have independent verification of these claims, independent testing, so that we actually know what's going on rather than trying to read the tea leaves every time to figure out if this is another case of PR drama or something real. Here's the thing. I will believe that one of these companies genuinely thinks they've built something uncontrollably dangerous when they stop funding against and start funding for campaigns for independent regulation.

16:14Until that happens, I'm not buying it. I don't believe them. or if one of these guys cried if like they got out in front of a microphone and started crying then I'd be like wow they're not I don't know I wonder so stay tuned on this we're going to be following this story very closely but I think we can comfortably say that it is not panic in the street times quite yet at least not today at least not today

16:45Hi, everyone. It's Zara McDonald and Michelle Andrews here. Hi. That's Michelle Andrews. And that's Zara McDonald. We're from the Shameless Podcast, the pop culture podcast for smart people who love dumb stuff. Yeah, on Mondays we cover pop culture analysis, think Netflix's Lamar Odom documentary, the great unfollowing of pregnant women, and the tangled web of celebrity GoFundMes. On Thursdays, Zaz, what do we do? It's what we call the quick and dirty, So it's a recap of the biggest pop culture headlines that week. Whatever you need an explainer on, we have got you. And then on Fridays, we do Dumb Stuff Fridays.

17:17Yes, we're three times weekly. Three times weekly. Search for Shameless in your favorite podcast app now.

17:26Okay, you guys, I can't sleep. I'm tired. Every day is a struggle for me. To the point that I have, over the past few years, I'm not kidding. That is a real problem. I'm so sorry I'm laughing. It's not, yeah, it's actually not funny, Karen. No, but I think a lot of people feel this way. And for the past 10 years, there has been this huge discussion. I think maybe we could describe it as a public freakout over the question of blue light. Have you guys heard about this? that there's there's something going on with your screens yeah with blue light of the spectrum and it's ruining your sleep yeah yeah for sure yeah like i've done i've downloaded like all the filters and whatever everything on the screen to try and change my blue light into warm light yeah yeah you can get apps to sleep that remove the blue light part of the spectrum on your phone which is supposed to help you sleep right they're built into apple devices where like you can set it up a couple hours before bed, it'll your phone will like the screen will let go a little bit yellow, supposed to be removing more of the blue light.

18:38Here's the thing. This whole conversation is totally off base. It is built around a complete misunderstanding. There have been people on the internet that are lying to you about blue light and what it's doing to your sleep and how it is and isn't ruining your life, but there is truth and science at the base of this conversation. So I wanted to know what's going on here. Is blue light making a difference or not? So I decided to investigate. I did an experiment on myself with the advice of experts for it's been about a month now. And I wrote an article. You can read this on our website. And what I found is the life from your phone is not ruining your sleep.

19:28Are you saying that this is a hoax or are you saying that it's just a more nuanced than people usually understand it to be? It is more nuanced. That's what it really comes down to. But I also think that there's been willful misinterpretation of this and it all kind of starts with a study, one particular study 10 years ago that was oversold. So the study was released where they got two groups of people and they put them in the lab and half of the people, they had them read for a couple hours before bed on a book, like a paper book. And the other half, they had them read on iPads and the iPad group.

20:11They went to bed later. They had a harder time falling asleep. The quality of their sleep was worse. they felt more tired the next day and their bodies produced less melatonin which is like you know where all everybody started taking these supplements melatonin it's like the sleep hormone yeah it's the hormone that your body produces before you before you go to sleep it right initiates sleep in the in the brain yeah so in a laboratory setting blue light can ruin your sleep and our screens are bluer. And the reason for that is everything uses an LED, right? You've heard this, the light emitting diode.

20:48They're much more efficient. They're cheaper. The problem is they cannot produce white light. So the way that an LED, the way that your screen makes white light is they have a bunch of blue LEDs and they coat some of them with a yellow chemical and the yellow and the blue mix together and it fools your brain into seeing white. Huh, I didn't know that. Interesting, right? The problem here is there is a protein in your eyes called melanopsin that is light sensitive. And it is one of the main things that controls your sleep cycle. And it is more sensitive to blue light. If you expose it to red and green and blue light, the same intensities, the blue light will mess with the melanopsin more.

21:33the problem is is your phone is not bright enough to make a significant difference so what actually matters and i read a ton of studies here i called up a bunch of sleep experts and scientists who were doing the research that this is all based on and they said what actually matters is the total dose of light that you're getting throughout the day. But what really makes the difference is that you're getting more light in, you know, the first two thirds of the day than you get towards the end when you're trying to go to sleep. But your phone is not bright enough to make a difference. In these studies, they keep people like locked in a lab all day and expose them to a very low amount of light.

22:25And then at the end, when they make you look at your screen, the total amount of light you've gotten throughout the day is much lower. So the blue makes a much bigger difference and it messes with your sleep. But you don't live in a lab. You live in the real world. There's light everywhere. And these filters that you have for your screen, they're just not doing anything at all. There's been a lot of research on this. They don't help. So what you're saying is if you get plenty of light during the day and then at night you're staring at your phone it's not going to make a difference your phone is still really messing with your sleep yes but it's messing with your sleep for different reasons it's messing because you're engaging actively with content if you're doom scrolling that is putting your brain into a state of sometimes panic especially you know with news at the moment that is wreaking havoc with your sleep it's not the light it's not the light and that's what you know all these doctors who are who i was talking to were so worried about is they believe that everyone has been misled to think they can put on their blue light filters and then they can just keep looking at social media they'll be fine no the social media is the problem yeah it's the content it is still the best advice you can possibly give is um get like a normal alarm clock sleep with your phone charging in a different room But what's going on here?

23:46Because it's also true that over the past couple years, our sleep has gotten much worse. So there's something about modern life that has changed here. And to figure it out, I tried to revert back. And I have taken, I think, some of the most absurd and extreme measures I've ever gone through for an article. And it starts with this pair of special blue light blocking glasses that I got. Oh, wow. Looks pretty cool, right? So these are made quite a lot like Bono right now. So I do, yeah. We got similar shaped face, I think, too, maybe. Bono sleeps very well, I guess. That's what I've heard. For people who are listening on the podcast, Tom is wearing very, forgive me, unfashionable sunglasses.

24:34What are you talking about? They're so fashionable. Yeah. And the lenses are deep, deep orange. and you've maybe seen people wearing orange shaded sunglasses around this is the thing so you can buy blue light blocking glasses uh that are just like sunglasses with you know yellow tinted lindas or you can get a coating on your glasses right a lot of the times if you wear glasses yeah they're like do you want us to put this special blue light blocking coating you're throwing your money away that does not help i talked to the experts on this that's also a waste of money. I put these on now for my experiment three hours before bed for like almost a month now.

25:17Oh, my God. I covered my windows throughout my house with blackout curtains. Makes everything pitch black. I go out. I turn all my lights off one by one. And I light my apartment with candles. Modern. Well, yeah, exactly. It's pretty sexy, especially with the glasses, you know, like it's pretty spicy around here. Candles produce a very low amount of blue light. Incandescent bulbs like the old light bulbs that we used to have produce less blue light than LEDs. Candles produce almost zero. So the reason I'm doing this is what you really need is to get a low amount of light at the end of the day.

25:57You want a ton of light in the morning. It's almost like you're inoculating yourself against the blue light by just blasting yourself with as much light as you can get first thing that's so interesting what the doctors told me is like the researchers who study this stuff you want to if you really want to work on this get up in the morning and go outside the big thing that's changed in the past couple years is post-pandemic fewer people are leaving the house, right? It used to be you'd get out of bed, you would commute to an office. And because of like OSHA, like at least in the United States, we have like special regulations for workers.

26:38You have to keep offices lit very, very brightly, like much brighter than your house would be. So you're in in the you know, the old way of doing things pre pandemic, you get so much light throughout the day. Then you'd go home. There would be less. It would help your body go to sleep. Now, so many of us, myself included, roll out of bed, walk to their desk and get basically the same amount of light all day long. So you want to go outside in the morning, get a ton of light. Then there are also studies that show if you can get more light again around like after three o 'clock, like three or four o 'clock, that's really helpful.

27:16And then keep things as dim as you can in the evening. But when it comes to your phone, the amount of light that you're getting the amount of blue light 24 hours of light from your phone is worth about one minute outside in bright sunlight your phone is just not the thing that's making a difference all right what kind of advice can we give listeners in the uk where it was raining for the first 45 days of this year and you could go outside and really very much not get this i asked this question And they said even on an overcast day, being outside even on a cloudy day makes an enormous difference.

27:59Another piece of interesting advice that I got was what do you do throughout the day, right? You sit and get the light from the windows during the daytime. And then at night you turn on the lights. What I heard was actually do the opposite. Keep it as bright as you possibly can during the day. Turn all your lights on during the day. And then as it gets toward nighttime, turn the lights lower. It's better to just go outside, but if it's really overcast or it's raining, you can kind of fake it yourself by adjusting the lights in your house. So interesting. So after a month living like some kind of very weird monk with my special glasses and my candles, the quality of my sleep didn't improve that much.

28:41but I did start to feel that I was getting sleepier around the same time. And it was like, I was more driven. They call it your sleep drive, like your desire to sleep. It felt like my sleep drive increased a little bit. That's subjective. I can't, you know, tell you scientifically if that's true. Like, I want to stress, this is not science, right? Like this is worthless information, but this is my personal experience. I'm not going to keep wearing these glasses though. This is absurd. this is not a lifestyle choice that any normal person can make but going outside in the morning will make an enormous difference it will help your sleep and put down your phone doesn't matter if you if you got the filter on that's not helping you one of the things people love to do before bed as you were saying is read and a lot of people read on a kindle and if you have a kindle model 12 or older no you don't anymore because the company is removing its support from it you have to buy a new one because that device no longer works now this is this is something that is as an absolute bugbear of mine it's called planned obsolescence it is this idea that you don't really buy something anymore especially in the tech industry you essentially lease it until the tech industry decides that it's no longer going to be supported.

30:05And then you have to upgrade because you have no choice. Nikki, what do you mean that like my Kindle would stop working? It means I turn it on and there's just no books. What's happening is it's stopping being supported. So it'll still kind of drag on if you're very careful with it. Your existing library will still work. If you at some point perform a hard reset, a factory reset on it, it will stop working. anytime it develops a bug or something. There's no support for it anymore and you will no longer, crucially, be able to access the Kindle store. So you won't be able to get any new books for it.