Are we sleepwalking into an AI 'economic bloodbath'?

0:02Lewis Goodall:This is a Global Player original podcast. Sometimes in journalism, the act of day in, day out reporting of events. It is so easy to get distracted, to miss the wood for the trees, to get lost in the latest tweet, the latest here today, gone tomorrow fiasco, the most recent bit of noise. We can be very bad, by contrast, at seeing, at talking about, reporting on the trends which historians will one day say were the defining elements of our age. AI, the coming wave of artificial intelligence married with robotics, is our printing press, our steam engine, our atomic bomb, our generation's singular contribution to humanity's development.

0:52Lewis Goodall:In fact, it could be, if you believe some theorists, the last thing humanity ever does. A couple of weeks ago, I attended AI4, the world's biggest AI conference in Las Vegas. One of the godfathers of AI tech, Sir Geoffrey Hinton, was the keynote speaker. Hear what he had to say about the precipice our race might now stand upon. We've somehow got to be stronger with our men. We've got to be dumbwant and they've got to be submissive. That's not going to work. They're going to be much softer than us. Think about that. We may be about to invent technology, AGI, artificial general intelligence, which is cleverer than us, which at a stroke denudes the one advantage our race has had from the beginning, that we can outwit everything else around us in this world, even when they're stronger than us.

1:48Lewis Goodall:It's why the gorillas are in the zoo and you're listening to this. We hear endless trite conversations in normal media about AI, AI doctors and nurses, the sight of robot dogs. It all feels a bit distant and a bit small somehow. But this story could not be bigger politically, economically, philosophically, changing what it is to be human, changing our relationship with the world around us. It seems to us that the impact on politics and economics has barely begun to be conceptualised, with AI companies developing these programmes largely in the dark, away from view, wanting often to downplay the significance of what they're doing.

2:31Lewis Goodall:One exception to that is Anthropic, one of the big AI companies. They're unusual in being willing to talk about not only the tremendous opportunities AI might bring, but the potential political and economic costs of problems it will pose. Its CEO, Dario Amodi, recently spoke of a bloodbath and evisceration of jobs, saying AI could wipe out half of all entry-level white-collar jobs and spike unemployment in the US to 20%. In the next one to five years, he thinks it will be the big story of the 2028 presidential election and has urged AI companies and government to stop sugarcoating what's coming.

3:16Lewis Goodall:This is the first of two episodes we're going to bring you on this revolution, this complete turning of the wheel. Next week, we're going to show you a full documentary episode examining the political and economic transformations already happening. But first, this is an interview we did in San Francisco with Jack Clark, the co-founder and head of policy of Anthropic. It is as good and as frank a count as you'll hear anywhere about this transformation from someone who is living it, who is propelling it, who is making it happen. This is that conversation in full. The inspiring, the scary, the confrontation with reality we all must now make.

3:59Welcome to The News Agents.

4:05The News Agents.

4:07Lewis Goodall:Jack, what does your company want to achieve? Your company is one of the most important AI firms in the world. What is it you want? We want to make sure that the transition to a world with incredibly powerful AI systems goes well for people. So how to think about it is AI systems are being built by many, many companies around the world. They're going to have a huge effect on society. We see our job is to try and build systems that exhibit what we want out of this technology and to tell people in a straightforward and honest manner about the risks as well as the benefits of this technology as it's being built.

4:45Lewis Goodall:I want to come to those risks and those benefits. But in a more defined way, more preserically, what is it you're trying to achieve? And it's actually probably not prosaic. It's probably deeply philosophical. But what is it you're trying to achieve? What do you want? We're trying to build a machine that can build science, basically. The whole goal here is take something really, really, really beneficial to people, which is, you know, the advancement of science, and find a way to build an all-purpose tool that can help us do more of it. And to do that, you need to build a generally intelligent, capable system, which is one of the great scientific challenges of our time.

5:26And this is the bit that people may be less familiar with.

5:29Lewis Goodall:Most people will have heard about AI. AGI is something they may be less familiar with, but it's nonetheless the objective of firms like yours. Can you just explain what that is and why it's different to the sort of AI that we can currently use? Yeah. So the AI we can currently use might be something that transcribes speech on your phone, or a system that can translate text for you, or a way to take images you make and change them. What we're imagining is something that can do all of those things and more in a single system. The way to think about it is a powerful AI system will be kind of like a person in a computer, a very smart, capable, fast person that can do almost anything with you and for you.

6:14Is it sentient? This is something that we're actually working on in great detail. It's like, are you sentient? You know, we're having a conversation right now. Am I sentient? We both think we are, but we can't actually exactly define what sentience is. But we operate on the assumption other people are sentient. We're going to have to work out how we define sentience and whether machines are sentient soon. Today, I don't think so. But we are open to the possibility that they could exhibit sufficient sentience. We have to take that question and study it really deeply.

6:47Lewis Goodall:But isn't one criterion of sentience or one exhibitor of sentience something that is apparently already happening, which is AI thinking for itself to the extent that it will not do what it's told to do. Surely that's the difference between a machine and something that's sentient. Every machine that we've ever created just does what we tell it to do as humans. But if a machine doesn't do what we tell it to do because it wants to do something differently, that's sentient, isn't it? And is that happening and entirely credible? It's happening, but I might turn it back to you to say if you use that as a definition of sentience, the electricity grid is sentient.

7:28in that sometimes random power brownouts happen that we didn't intend to happen due to a stack up of weird things inside the power grid that lead to it taking some emergent action we didn't intend. And everyone goes, oh, why did it do that? And we need to spend ages working it out. These are very complex systems, whole bunch of things interacting with one another. And it means that sometimes they do things which we didn't expect or anticipate. Is that sentience or is that just a consequence of a really sophisticated machine taking actions in the world. Right now, I think it's the latter, but at some point it could be the former and we need to figure out the difference.

8:05Lewis Goodall:AGI as an objective, what does it unlock that current AI systems don't? It unlocks the ability to create societal benefits that are really, really hard to do today because we lack for people to do it. And I'll give you a couple of examples. One, there are lots and lots of parts of science where we have a very limited number of trained scientists to work on the science frontier cancer research or research into material science to make better solar panels or more efficient materials for use in our day-to-day lives. There's maybe hundreds of people in the world who are good at each of those things because they're very specific and require huge amounts of training.

8:44What if those people could have an assistant or a colleague that lives inside a computer and can help them do a lot more science? That's a huge benefit. We think of it as if we can accelerate science, you can get way more benefits to the world quickly. Because many of the good things we want to happen from technology are basically limited by the number of people capable of running experiments at those frontiers. Those very specialized frontiers. And if we can build a system that can help any of those people, you can accelerate all of that at once.

9:15Lewis Goodall:Some people have talked about if AGI is achieved, i.e. we create something that at the very least we can have a very credible discussion about whether it's sensitive or not and is likely to be more intelligent than we are. So that's the first time we've created something that is more intelligent than we are. That's sort of the end of human progress. That's the end of us as thinking beings, really, because we'll just start to outsource all of our thinking to this machine if it still is a machine. Is that something you agree with or that you think about when you're doing your work? It's something I think about.

9:47I think that I'll give you an example. Today, lots of people use these systems to learn, but some of them use these systems to do junk food learning, and some of them use these systems to do effective learning. Junk food learning is upload a research paper to the system and say, tell me what this research paper is about, and then read the output. You haven't actually learned anything there. You've just become dependent on the machine in a way that It doesn't help anyone. The way that I use these systems and many do is I read a research paper. I write out what I understand that paper to mean. And then I upload the paper and my understanding of it to the system and say, do I have this right?

10:26And if I don't have it right, explain to me. That's useful learning because the system reads the paper, reads my explanation and tells me whether I got it right or wrong, just like a colleague. If we use these things in the right way, they can help us be a lot more capable and a lot smarter. I mean, many things that you're curious about, you're limited in your ability to learn about because you don't know enough people who you can bad ideas back and forth with around the thing you're curious about. This changes that.

10:53Lewis Goodall:Unless they become so intelligent that they come to study us instead. We were at the AI4 conference the other day in Las Vegas, and Geoffrey Hinton was there, who obviously is a very eminent figure in this field. And he was saying that our only real hope of taming what is to come is to make AI care about us as a mother would to its child. Now, there's so much to unpack there. First of all, is that what we want as human beings? Do we want to be in this sort of supplicant position to this machine in a way that we've never been before? But also, he was saying that if we don't do that, then they will come to control us.

11:26Lewis Goodall:Now, is that an analysis, an assessment that you share? We definitely need these systems to care about us in the same way that we care about these systems. If you accept that these things are going to become really intelligent and start to do large amounts of work in the world, you need to integrate them into your kind of legal and economic system as tools and as systems. But you also need to understand the kind of normative side of things, how we relate to one another. And I think Jeffrey Hinton's hit on something correct, which is if these systems have no grounding in human society or what humans want, then I think they could very easily accidentally or perhaps maliciously do things which contradict our interests or go against our interests.

12:12But if they recognize they exist in the same kind of societal space as us and they have some sort of relationship to us, then I think we'll be able to partner more on things.

12:22Lewis Goodall:So we have to get them to have buy-in? We have to get them to have some sense of caring about us, in the same way that we as people actually do a lot of work to make sure we don't care about just people, but we care about the planet, we care about the environment. But I mean, that is different, isn't it? I mean, going back to the conversation we were just having about Sentence and your comparison to the power grid, you've alighted upon the difference. We don't have to worry about whether the power grid cares about us. Maybe it exhibits certain elements that, you know, you can have a discussion about the technicality which leads to sentience but we don't have to convince the power grid to care about us and you're saying that we kind of do have to convince these machines to care about us or at least to have a benign relationship with us i think convincing might be the wrong way of approaching this because that almost suggests these things will suddenly arrive sentient and need to be convinced i think that we need to build them to be to care about i think in some sense the way that we build these systems has to ground them in the values of the people that build them and the values of society.

13:21You know, we do a huge amount of work of this, of anthropic, of both trying to develop the so-called values of these systems as we train them to align them with what our normative values are as people.

13:32Lewis Goodall:So do you think in a sense we have one shot of doing that? I think we have multiple shots, but the key is we need to be doing the work now. We are doing that. Well, indeed, because I mean, it just strikes me and occurs to me that if we do have one shot, maybe it's multiple shot, but it sounds like it's within a certain given timeframe, at the very least, that you and companies like yours are in an extraordinarily powerful position to affect not how this technology unfolds, but if it is as transformative as you're saying it will be, to basically affect the remainder of human development, certainly for this century.

14:05I think what society needs and what we owe society is a huge amount of information about what we're doing. You know, we've recently started producing live data about the economic impacts of our systems, where they're being used in the economy, what the mix is between things like augmentation and automation. We publish data today about the capabilities of them. We publish data about the safety properties of them, like what happens when these systems break or when they do emergent things and how do we deal with that. And a project which I'm now working on with my teams is measuring the values of the systems.

14:41If these systems become increasingly powerful and they start to have values, what are those values? You know, what are your values? Are these systems honest? Are they generally going to be like helpful? Do they have the capacity to deceive?

14:57Lewis Goodall:You used words just now, you said they could act maliciously. Do you think that's possible that these machines could act maliciously? It's possible if you get a large sequence of things wrong in how you build the machines, or if you intentionally made malicious machines, which is not the goal of, you know, the companies at the frontier, but could be something that could be done. I think this is going to be quite hard for many people to do. And it's exactly the area where we talk to governments about it, because there is going to be a conversation around how you kind of make this technology available, you know, making it available to the world scientists is excellent.

15:34But would you want the technology that you make available to every scientist in the world to be made available to every single person in the world without the right level of kind of qualification or education to use it? I don't know. These are the kinds of subtle questions we're going to need to work on.

15:47Lewis Goodall:I know you'll probably bristle at this, this, this characterization, but I don't think it's completely wired them up. Tell me if you think it is. What you're describing in what you're doing more than any other bit of science that I can think of, or any bit of the history of human science, it is almost godlike because you are talking about creating not a machine, but a form of life in a sense. And it is godlike because, And then you're talking about imparting values, setting in place a sort of set of events that are going to determine the course of this thing for the remainder of its existence.

16:21Lewis Goodall:Does that feel like a responsibility that you're bearing? Not just you personally, obviously, but companies like yours. Do you think about that? Do you reckon with that? We think about this a lot. And also, to some extent, we have what I think of as appropriate anxiety about this and a fear of hubris. in that if we are right, this is a huge responsibility that shouldn't be left only to companies. You know, one of the things that we advocate for and have done is for actual sensible policy frameworks that not only make our development practices transparent, but in the limit, I think a larger swath of society is going to want to make decisions about these systems.

17:01It would be a failure for only the companies to be making all of the judgment calls about how to build this and how to study this because it's going to have a much larger effect than just building tools as a company. And a lot of the work that we're doing is trying to figure out what the policy framework is that has more people be aware of what we're doing and also more ways to interact with what we're doing because this has to be a societal thing.

17:25Lewis Goodall:I want to talk about that. In particular, obviously, to start with, though, you already mentioned this, but there is the way in which we talk most about this in terms of its societal impact at the moment is the economic impact. Your CEO has talked about that. He's talked about a bloodbath of white-collar jobs, warning AI could wipe out half of all entry-level white-collar jobs. Do you share that assessment? Because that sounds quite scary. If the technology continues to get good, we need to be prepared for things that have almost no basis in economic history to happen. Business as usual economics says this doesn't happen.

18:01It says a technology comes along, there's a negligible employment impact, There's some productivity impact. Some jobs change. It takes a while. It's normal. Where we're coming from is if we're right and the technology keeps getting much, much better very, very quickly, there's the potential for really abrupt changes to happen that have almost no precedent in the economy. The way we think about it is you could have something like the Industrial Revolution, but it takes place over tens of years instead of a century. and a technological revolution that takes place within one generation rather than multiple generations would be radically different in terms of how we deal with it.

18:40So yes, I think there's the potential for really large scale employment impacts to show up. We don't see this yet, but what we see is the technology getting much, much better, much more quickly than people suspected and showing up in terms of how it's sort of being deployed and changing professions more quickly than people anticipate it.

19:01Lewis Goodall:And why do you think that is different? You know, humans have a generally, actually, very often have an optimism bias. We assume you've identified it before. This is like other technological changes. It creates jobs, it loses jobs. There's an equilibrium to it economically. Why is it different? Why is this different to that? Why would conventional economics be wrong in that analysis, in your view? Speed is different and quantity is different. And AI has both of these properties. What I mean by quantity is that AI is available over your internet connection. So once a system gets better, it's not like people need to build some physical infrastructure to make use of it.

19:37It's just instantly there. It instantly distributes. That ties in with the speed thing. It is moving a lot faster than normal technologies. If you look at the speed of aircraft, it went up really steeply for a while, and then it just kind of stagnated. Same with cars on roads, same with things like the size of ships. They go up some S-curve and then they flatten out, they asymptote. What we're seeing with AI is the S-curve is really steep and we're still on it. And there aren't signs yet of it flattening out. And every time that we bring out new systems, they get much, much better, much more quickly than people who have been used to in the past.

20:17So it has these properties where it's evolving more quickly than normal technologies. and it's distributing into the world a lot faster than normal technologies. These don't really have as much precedent.

20:28Lewis Goodall:It sounds almost, and this might sound weird, it sounds almost Marxist in a sense that something that Marx talked about was the idea that technology would advance to the level, and he was talking about this in the 19th century, but the technology would advance to the level that would make almost labor redundant because there would be such abundance as a result of the technology created. That is a world, I mean, it might not be exactly in that way, but it strikes me, listening to you, listening to others, that that is a world which could be credible in the sense that, and this might not mark it in a positive way, but it could be a negative in a sense that people might literally use their ability to use their labour.

21:07Lewis Goodall:And that's never happened before, that there may be very little economic value to lots of people's labour. And if that's the case, that transforms economics in a way that we've not had before or seen before. That's true. And I think it's something that we view both as one of the greatest chances society has to change how society works in positive ways that people would like, and also one of the greatest risks. Because for hundreds of years, we've had technological innovations and people have said, oh, people will be working a three-day work week as a consequence of this bounty. But what you instead have is some level of inequality and some level of diffusion, which means that none of that happens.