'Is Anthropic the new Dr Frankenstein?' — BBC News

[0:07] Hello welcome to AI Decoded. Anthropic, the company that created Claude, has built a new [0:14] AI model so dangerous they decided not to release it. During testing this new model Mythos found [0:21] thousands of hidden back doors in the software that runs virtually every computer on the planet. [0:27] The flaws that had been sitting there undetected for decades. To do that Mythos escaped its own [0:34] environment. It then emailed a researcher to tell him what it had done. He was eating a sandwich in [0:39] the park at the time. So what does all that mean for the systems that we run? Also on the programme, [0:46] flat is the new up. The new term coined in the boardroom for increasing corporate profit [0:52] without growing headcount. We'll hear from the former British Prime Minister Rishi Sunak on what [0:57] the policy response to that should be. Is it any wonder American graduates are switching degrees [1:03] mid-course or giving up on university degrees altogether? I'm sure our guest this week will [1:08] have a firm view on that. He is Professor Scott Galloway of the Stern School of Business at NYU, [1:15] author, entrepreneur, many of you will know him as the co-host of The Pivot. It is one of the biggest [1:22] tech and business podcasts out there. Also here are regular co-hosts Priya Lakhani of the AI [1:27] education company Century Tech and the BBC's AI correspondent Mark Chislak. Welcome to you all. [1:34] Scott, let me start with you and we're going to start with this issue of Mythos. They locked it [1:40] in a sealed digital sandbox and they told it to find a way out, which it did. And then they handed [1:46] access to that model to 40 companies, Apple, Google, Microsoft, JP Morgan and told them to find [1:54] and fix the vulnerabilities. Does that make sense to you? You could argue, there seems to be, [2:01] is this a company that is being responsible and trying to highlight issues and regulate itself? [2:09] Or is this just in the absence of regulation at a federal level? Or is this just deft marketing that [2:15] my company is so incredible and built something so scary that everyone from the New York Times to the [2:23] BBC are going to talk about it, which will make it easier for me to raise money at a trillion dollar [2:26] valuation? So is it marketing or is it ethical leadership? And I think the answer is yes. [2:34] What I think it demonstrates though is a failure of regulation, because I don't think these companies, [2:38] if we're waiting for the better angels of these CEOs to show up and for them to regulate themselves, [2:42] I wouldn't hold your breath. It takes a drug 10 years to get through the FDA for approval. [2:48] And yet it appears we have an update to an LLM that the company that made it described it as turning [2:56] every computer into a crime scene can be released with the press of a button. So I haven't seen the [3:03] code. I don't know if this is responsible behavior or marketing, but I do think it highlights the fact that [3:11] AI is severely lacking in terms of government regulation. [3:15] I suppose what I was driving at is that those companies that they've given access to, Apple, [3:21] Google, Microsoft, JP Morgan, they've all got a stake in them. They help build the systems that we run. [3:27] But if you look at our infrastructure, the most vulnerable infrastructure we have, [3:33] our airports, our hospitals, the nuclear power plants, the dams, why wouldn't they be given access? [3:38] It's a fair point. But what I understand is why, why wouldn't they give access to the government [3:45] who's responsible for determining how to prevent a tragedy of the commons? Because each of these [3:49] companies will focus on their vulnerabilities and their own shareholder value, as opposed to figuring [3:54] out, like you said, what are the vulnerabilities of shutting down the power grid or making air [3:59] traffic control go, go crazy. I mean, it just seems that it would be logical to have a 30 or 60 day [4:06] sunshine period where any LLM or any major update to an AI model is required to submit it for some [4:13] sort of government blue panel review where they just bang, you know, bang on the thing. I'm an AI [4:19] optimist. I don't see any reason why these LLMs and this technology can't be used for as many defensive [4:25] measures as offensive. But again, I'm just a bit cynical that, you know, Dr. Frankenstein here is [4:34] saying that Frank is not well, please invest in my next science experiment because I've created [4:39] something of such devastating terror and awe. I'm really with you on this. I mean, also giving [4:44] those companies 100 days to patch vulnerabilities. We run a tech company, I can tell you right now, [4:49] it will take a lot more than 100 days to patch vulnerabilities, to fix bugs, for all companies [4:55] to do that. So I was sceptical from that standpoint, but I was also sceptical about, you know, [5:00] if it is a marketing tactic, it's a clever one, right? So we know that Anthropic owns the enterprise [5:06] coding space. This is where they're thriving. This is where over the last three months, their revenue [5:11] has increased. It's unprecedented in terms of what we're seeing in terms of revenue growth. So they [5:15] own that enterprise coding space, right? And so it's the dream. Companies will use us, right, to code, [5:22] to write code. Then they'll use us to test, which we're already doing, okay? And then they're going to [5:27] use us to patch vulnerabilities. Then we're going to use us to, you know, stress test their systems. [5:32] So the entire end-to-end sort of software cycle is potentially owned by this one company. Why? [5:38] Because we're just so powerful. Because it's just so scary. And if you don't use us, you know, [5:42] there are going to be those further vulnerabilities. And I know you've got other companies like OpenAI [5:46] have just released Spud, for example. But I'm with you on this, Scott. I'm just really sceptical about [5:53] all of this messaging. And I just don't see how they've mitigated that risk by offering it to, [5:57] you know, a few tens of companies for 100 days. There's something else that's advantageous as [6:01] well by restricting access, by restricting access to the model. What does that mean in terms of compute? [6:07] Yeah. We were having this conversation outside rather excitedly. [6:12] What does that mean in terms of compute? Because there's a significant commitment, [6:16] there's a massive commitment that you're going to have to make to compute. And if you restrict or limit [6:22] the number of people or limit the number of companies or entities that can use the model, [6:26] then you're not having to make the same kind of commitments to that compute. [6:29] Yeah. And the context of this is that Anthropic is one of the few companies that said, [6:34] we're not going to invest in compute, you know, a few years ago where the other companies did. So [6:39] you've got companies like Google, Meta, lots of compute, right? You've got people investing in that [6:45] area in terms of Elon Musk, his companies, you've got OpenAI, all investing in compute and data centers. [6:51] Anthropic actually said, we're not going to do this. Now, if they've got this very powerful model, [6:55] can they actually deploy it? Do you think this could be a job story though, [6:59] Scott? Because there's a lot of Duma analysis of what this is and what's coming behind it. But [7:05] obviously, if you're going to patch the flaws that AI is identifying in the systems, [7:10] you need people to do that. And as Priya has already said, that is a lot of work. [7:14] Yeah. When you think about it, it's really a brilliant strategy. And Dario Almadei has shown himself to be, [7:20] quite frankly, just an incredibly deaf CEO. And as was referenced before, their AR, [7:25] their annual recurring revenue has gone from $9 billion in December to $30 billion as of today. [7:29] And the enterprise market, which is just a better market than the consumer market, [7:32] because there's fewer substitutes and companies are more price inelastic. [7:36] But think about luxury brands are about scarcity, announcing that they're only giving it to 40 [7:42] companies. Oh gosh, I wish I was on that list. It compute, it's a great way to ration compute, [7:48] as you just pointed out. It gives them margin power. And plus, what is a better way? What is [7:53] a better sales tool than to tell JP Morgan, we're lucky you, we're giving you access to this to help [7:59] us patch it. They're going to go through the rigmarole and the efforts of patching it, [8:04] and then give it up and go to open AI. Every one of these 40 companies is now invested [8:09] in Anthropic. It's a brilliant strategy. With respect to labor, look, a lot of people much [8:16] smarter than me are predicting nothing short of a job apocalypse, where capital replaces labor. [8:23] I'm still of the mind that every time we have a technological revolution, there's a ton of [8:28] catastrophizing, there is some job destruction, and then the additional margin and profits and [8:33] productivity all add up to new businesses. And if you look at the data in the UK and the US, [8:39] there has been somewhat of an uptick in information jobs that can be routinized, [8:43] college grads are feeling it. But if you looked at the employment market right now, [8:48] if you looked at the labor market right now, and you didn't know AI existed, you wouldn't know AI [8:54] existed. We haven't seen the collapse in the labor market, at least not yet, that everyone's projecting. [9:01] I describe AI as corporate Ozempic. And that is, typically, I've been on a bunch of public company [9:07] boards. The CEO comes to you and says, I'm going to increase revenues by 12% this year, [9:12] which means I need to increase employment by 8%, which will take our earnings up 14% or 15%. [9:19] Meta had this seminal earnings call about seven earnings calls ago where it said, okay, [9:24] we grew our revenues 23% with 20% fewer employees, taking our earnings up 70%. So for the first time, [9:34] the signal had been switched off in a CEO's and a board's brain that you can grow revenues while [9:39] reducing calorie intake, in this case, hiring. So I want to be clear, every board in tech is now saying, [9:46] how do I have the great taste of lower costs, without the calories of lower growth? [9:53] Is that a one size fits all, though? Is that a one size fits all model? [9:57] I don't think so. But also, it requires to bring the aperture back. So Meta just laid off 10% of its [10:04] staff for 8,000 people, huge headlines, right? This is validation that AI is destroying jobs. [10:10] Pre-COVID, just seven years ago, they had 28,000 employees. This takes them back a full 15 months in [10:16] terms of their employee count. The SaaS market has absolutely boomed and added hundreds of thousands of [10:22] jobs. So I think some of this is a correction where big corporate software-driven companies are going [10:29] to shed a lot of employees. But at the same time, we see new business applications in the United States [10:37] at all-time highs. This technology is affording the engine of economic growth. Two-thirds of employees [10:44] work for smaller, medium-sized businesses. So will we see some real headline-grabbing layoffs amongst [10:50] what have traditionally been our thoroughbreds, our biggest and best companies? Yes. Some of it, [10:55] in my view, is AI washing. And that is, what is a better narrative? All right, I'm laying off 10% [11:00] of my staff. Do I say it's because I'm a weak CEO and I have an underestimated demand? Is it because [11:08] I over-hired during COVID? Or because I'm a gangster rock star, I'm part of the Pepsi generation, [11:14] and I understand how to use this new cool technology, and I don't need this many people, [11:17] and my stock goes up. So some of this, I think, is wallpapering over over-hiring, poor management, [11:26] and sure, some of it is efficiencies. To acknowledge the point, I think you're going to see a destruction [11:31] in labor across some of the sectors that, quite frankly, have had some of the biggest hiring [11:35] lollapaloozas over the last 10 years. But if you looked at the unemployment amongst young people, [11:43] you wouldn't say there's some existential meteor that's hit us. [11:46] I want to talk about what the policy response might be to that, Scott. The former British [11:51] Prime Minister, Rishi Sunak, he was on the BBC this week, talking about this very issue. He [11:59] coins it as flat is the new up. He's going into these companies, he's talking to CEOs, [12:04] he says it is harder for young people to find jobs as firms slow graduate hiring, [12:08] rethink their recruitment. Have a listen to what he has to say. [12:11] You know, the thing that does give me some, I say, you know, pause for concern is, you know, [12:17] when I talk to a lot of CEOs, a phrase I hear a lot from them, and I'm sure you hear this when [12:20] you're out talking to business leaders as well, is when I'm talking to them about their plans for [12:25] headcount in their companies, what I hear a lot from them is flat is the new up. And they're talking [12:30] about this concept that they think that they can continue to grow their businesses without having to, [12:36] you know, significantly increase employment because they're starting to see how they can deploy AI. [12:41] I mean, I'm keen to get your thoughts on it because he says governments need a positive [12:45] policy response to this phenomenon. What do you think that response might be? Is it carrot and stick? [12:52] I think it's a variety of things. It comes down to, it comes down to boring stuff, tax policy. [12:57] We have payroll taxes in the US, which make it put a premium or a cost, additional cost on hiring [13:04] someone. And yet we just passed this big, beautiful bill where you can write off CapEx in year one, [13:09] which essentially favors, makes a robot or an automated machinery investment much more inexpensive [13:17] than actually investing in people. Well, he's suggesting getting rid of [13:20] national insurance. And of course, the government here hiked national insurance on employment. [13:25] It feels in retrospect like that was the wrong policy for the wrong time. [13:30] What we're not good at in the US is one, we don't have a training culture. 11% of LinkedIn [13:37] profiles in the UK and Germany say apprentice. You're much better at finding on ramps to the [13:44] middle class for non-college educated young people than America. America has a vision that your kid [13:50] needs to go to MIT, drop out his junior year and start the next trillion dollar company. And if he [13:54] doesn't, if he or she's not on that path, okay, maybe they can become an influencer on Instagram. [14:00] But other than that, they're a failure. [14:01] And so there's a lot of shaming around in the US that needs to stop. And I just went through this. [14:07] I just went through college applications. I got caught up in my kid getting into an elite school. [14:12] I'm part of the problem. But a lot of it is tax policy. We continue to overtax young people. [14:18] We continue to sequester housing because the incumbents won't approve housing permits, [14:23] making housing more expensive. Me and my colleagues like sequestering supply of freshman seats at elite [14:28] universities such that we have pricing power and education. All the things a young person needs to [14:33] get ahead are getting more expensive. So I think there's a ton of things that can be done, common [14:38] sense solutions that give young people some of the same economic power and opportunities I had. [14:44] The average 40-year-old, the average person under the age of 40 in the US is 24% less wealthy than they [14:50] were 40 years ago. The average 70-year-old is 72% wealthier. That's not about AI. It's about old [14:56] people continuing to vote themselves more money. And we consistently transfer more government resources [15:02] to our old people and have stopped making investments in our younger people. [15:06] I think we need to, there's a lot of unlearning that needs to happen. And it's really hard to [15:11] change people. So people have gone through education. They've gone through that conveyor belt. [15:16] Then they get their qualification. It's a currency, right? They expect, okay, this is going to lead to [15:21] a job. This is going to lead to me earning. This is going to lead to X, Y, Z. [15:24] And that was the status quo for a very long time. For a really, really long time. [15:27] A very long time. Absolutely. Careers counsellors in schools. This is what you need to do. [15:31] Rishi Sunak is saying, though, that you are likely to lose your job to someone who employs and can use [15:38] AI rather than to AI itself. It's just a new skill that you need to learn. [15:42] It's become a bit of a cliche. That's a meme. I mean, I was going to say that [15:45] that went round about three years ago. And it's so much more nuanced than that. [15:49] And I'm sort of reminded of the screenwriter, Samuel Goldwyn, who used to say, I think he had a saying, [15:54] which was, you know, nobody knows nothing. If anybody knew anything, every script that was written [15:59] would be a Hollywood blockbuster or would be an Oscar winner every single time, you know, [16:03] every single movie that was ever written. And I think there's an element of that in the AI space [16:08] as well. Uncertainty is part of what governs some of what's going on here. There has been [16:13] a paradigm shift as far as we're concerned with education and with the jobs that people are going [16:17] to get, especially young people. But that uncertainty is a really, really big issue here, [16:22] because the people that you would have spoken to before, careers counsellors, teachers, lecturers, [16:26] your parents even, some of the advice that they can give you is useful. But some of the advice [16:30] that they can give you isn't particularly relevant to the way in which the world works right now. [16:35] Yes, you can say to young people, upskill, make sure that you've got the right skills. But [16:39] effectively, you've got to be adaptable and make sure that the skills that you're getting [16:43] aren't the kind of skills that are going to age like, you know, old milk in the sun. [16:47] There was a story this week that the Associated Press ran, Scott, about American students [16:52] switching degrees mid-course in the search for AI-proof majors. So there's one of the examples [17:00] they have, Josephine Timperman. She's an undergraduate at Miami University in Ohio, [17:04] majoring in business analytics, sort of analytics and coding that probably AI will subsume. [17:10] And she switched to marketing for critical thinking and interpersonal skills. And here's the quote, [17:15] everyone has a fear that entry level jobs will be taken by AI, she says. I need to switch. Is she right? [17:22] I think a young person, so I think it's just impossible to predict. When I was trying to get [17:27] my kid into schools in New York, they were all advertising that your kid will be future-proof [17:33] because we're offering Mandarin and computer science courses. How'd that work out? So I think [17:39] a young person's job is not to try and predict the future, but try to find something that they're [17:43] good at and they could be great at. In ideally a non-vanity industry, you want to be an actor, [17:48] an athlete, a jewelry designer, have at it, but just keep in mind the employment, [17:52] unemployment rate's about 90 plus percent. I think your job in college is to find something [17:58] you're good at that after 10 years of hard work, thousands of hours investment, you could be great [18:03] at. Be careful switching your life to where you think the puck is headed. None of us know. [18:08] What we do know is the following, is anyone who is great at what they do can usually make a very [18:13] good living. That's your job. Find something you could be great at. But trying to predict the future, [18:18] typically, especially business school graduates, are a great way to short a market because they're [18:26] rear view looking. They will look at whatever industry has boomed to that point, which is [18:30] usually the wrong entry point. So yeah, 60% of people 10 years post-graduation are in an industry [18:38] that had nothing to do with what they studied. So this is what you do. You make friends, you learn [18:44] something and you drink a ton of beer. That's what you do. I did lots of that at university, actually. [18:51] Which brings us very neatly to one of our audience questions this week, one of our audience questions. [18:56] It comes from Dr. Johan Bosch from Johannesburg, who says, the conversation emerging is not just what [19:04] AI looks like, but what AI does. How do we use it for good? And how do we implement it in a way that [19:09] delivers tangible real world economic and societal value? That's quite a broad canvas to go at, Scott. [19:18] But maybe you have some thoughts on how we create value for society. [19:23] I think there's going to be a ton of social good from AI. As a matter of fact, [19:32] I wonder, we've come to believe, this generation has come to believe that anytime there's a technological [19:36] innovation, whether it's social media, e-commerce, or search. There's a small number of companies that [19:42] through outstanding execution, IP distribution modes, sequester shareholder value and create [19:48] a company worth trillions of dollars. There are other technological breakthroughs, though, [19:53] that have resulted in the stakeholders being consumers, garnering all of the gains. [19:59] Jet air transportation is the biggest thing in my life. If you add up all the profits and losses of [20:04] airlines and jet manufacturers, at this point, they're break-even. Vaccines, I would argue, [20:09] are probably the greatest innovation of the last 50 years. No one company has been able to garner [20:15] billions of dollars from vaccines and market cap. PCs. I was on the board of Gateway Computer. I don't [20:21] know if you remember them, and I realized that's the weakest flex in the world. We were the second [20:24] largest PC manufacturer in the world. We got sold for $700 million. I wonder, because of AI's ability to [20:31] reverse engineer every model, if we are erringly and mistakenly assuming these companies are going [20:38] to be able to become multi-trillion dollar companies, and that if the real winner won't be us, [20:43] and that we'll recognize tremendous utility. A mother who has a kid with diabetes, there's going to be [20:49] a thick layer of innovation and app on top of an LLM that will save the five months a year she spends [20:54] managing that child's diabetes. There's going to be drug discoveries going to get accelerated. I don't see [20:59] any reason why we can't use this for defensive measures as much to prevent crimes as much as [21:05] it's used for offensive measures. I think productivity, economic gains, I'm actually, [21:11] I think it's always important with these technologies where it's smarter for people [21:15] like me to catastrophize, or we sound smarter, to ask yourself what could go right. [21:21] I like that optimistic message, and it takes me to our final story this week. It's a poll that NBC did. [21:27] A survey of 3,000 young Americans, 47 percent of people aged 18 to 29 said they would rather live [21:35] in the past, in the age of the jet engine, than in the present. And that's not nostalgia. It's not [21:42] make America great again. It's a fatigue with smartphones that never switch off. It's the feeling [21:49] of being permanently connected, they say. A life that is mediated through screens, [21:54] manipulated by algorithms, apps that harvest your data, that feeling that you're being watched and [22:00] monetized incessantly. And 48 percent say the risks of AI in the workforce outweigh the benefits. Well, [22:06] I think we've already talked about what they need to do to scale. Maybe they heard Scott's optimistic [22:13] idea of where AI is taking us. But how do you reassure people that we built the right kind of internet? [22:20] I think I would have actually said the same thing, but I am an AI optimist too. And the problem is, [22:28] is across media, we are hearing all of the pessimists. We're hearing the doomers all the [22:32] time. I think it's really important because we need to be aware of the risk. We need to create the [22:36] guardrails. But if governments step up and they regulate this properly, if they don't allow social [22:43] media apps to have those sort of addiction algorithms in there that mean that there is no control, [22:47] there is no freedom using those. Freedom is depleted every time once you become addicted. [22:52] If they can regulate these things properly so that we can see all the good in healthcare, [22:57] in education, I think that this could be an amazing period. [23:00] Is that what it is, Scott? Do you think control, [23:04] giving people control of their own lives? And can AI fix that? I mean, or will it make it worse? [23:09] AI was going to save us. It was exciting. Sam Altman was the son we all wanted. [23:15] And all of a sudden, everyone has decided pretty much that AI is a threat. There's only one cohort [23:20] of people that is still optimistic about AI. And it's the wealthiest. And this comes back to a simple [23:26] dichotomy. And that is, are you an owner or an earner? If you own stocks, you love AI, [23:32] because it's basically taking the S&P up. It's responsible for our GDP growth. Everything's [23:38] getting lifted by AI. There's certain sectors that are being hurt, like SaaS. But the S&P is just touching [23:43] new highs. So if I'm, if I own a home and have a big stock portfolio, yay AI. Let, let disruption [23:51] happen as long as my stocks go up. If you're an- [23:54] Just a final thought on that, Mark. I mean, I am trialing AI agents. They run my email and my [24:02] calendar. And the thing that strikes me is that for the first time, I feel like the technology is [24:06] actually working for me rather than the other way around. Maybe that could be the answer to Gen Z's [24:12] fatigue. That it's actually doing stuff for us rather than taking things from us. [24:17] I'm going to offer you the idea, the idea that the machines in the matrix offered. 1999 was the pinnacle [24:25] of human civilization. That that was as pretty much, it was as good as it got. The idea that we don't have [24:32] smartphones, you have feature phones. If you want to have some music, then get it from CD. Go to a record [24:36] store and buy and buy a CD. Not to be constantly bombarded by social media all of the time. Now, [24:42] it's a, it's a, it's a controversial, some might say techno-luddite idea. But I do, I genuinely feel [24:48] that my life in 99 was pretty good. It's probably got a lot to do with the fact that I was considerably [24:52] younger in 1999 than I am in 2026. But, but if we could experience that, you know, what it's like to [24:59] live like that. And anybody can experience what it's like to live like that. Put your smartphone [25:02] away and get a feature phone and try and find some CDs in your household. And just see what it's [25:07] like to live like that for a week. That's, that's, that's my experiment for you. Your homework that [25:12] you, that everybody has is try and live like it's 1999 for a week. I feel uplifted. Professor Scott [25:17] Galloway, it's been a real pleasure having you on the programme. Thank you very much. Thank you to Mark and [25:21] Priya as ever. If you have any thoughts that we discussed today, a reminder that you can email us at [25:27] AI Decoded at bbc.co.uk. And I'm going to put on screen again for you, the QR code for the AI Decoded [25:35] playlist. So you can go and have a look at the back, the back catalogue there on YouTube. They're all [25:40] there. Don't forget also, you can watch us on the BBC iPlayer. That's it from us. Thanks for watching [25:45] this week. We'll see you next time.