Remember that these large language models, they're not deterministic. They don't look for the truth.
The vast majority of population in the United States and in Europe think that their jobs are under threat. Where is this coming from?
I think it's a media story more than anything else. I think the number of jobs won't change. Technology is just as good as the people that are using it.
Hello, welcome. My name is Adam Jasser and this is a special edition of On the Record. My guest today is Werner Vogels, Chief Technology Officer at Amazon. Mr. Vogels, thank you very much for joining us here in Warsaw.
Thank you very much for having me.
So, I have a simple question to you at the beginning. Are you worried at night that AI is going to take your job away?
Why should it? It's just another tool. It will make us in some areas more efficient. I can imagine that a number of simple jobs which could be automated may change. But I'd like to go back to another type of technology. If you think about spreadsheets—spreadsheets were developed in the '60s, became commercial in the '80s, Lotus 1-2-3, Excel. And at that moment, take a company like Morgan Stanley would have a whole floor of accountants that all have a calculator. It's basically a big human spreadsheet. When spreadsheets came and became more commercially available, all of those jobs disappeared because basically were a big human spreadsheet and now were replaced by technology. Those accountants didn't disappear. They just got other jobs in a company where it's more relevant instead of just being a human calculator. They actually were forced to use their brains to start using their judgment and the things that actually accountants were really good at.
The interesting phenomenon is that while data seems to be indicating that jobs are growing on the back of AI, the vast majority of population in the United States and in Europe think that their jobs are under threat. Where is this coming from?
I think it's a media story more than anything else. We love controversy. We love things that are changing. I have been saying there is absolutely no evidence that jobs are disappearing. Actually, people may move to jobs which are much more critical to using human brains: judgment, insight, creativity, all these areas that we're really good in as humans versus just doing paper pushing. If you just have a paper-pushing job, it might be a job that gets automated. And in some sense, AI is not really that terribly much smarter. It's really good in automation, automating a number of processes. And you can imagine that, for example, in the legal profession, things where you used to have to do a lot of research, you can now use these automation tools to actually get the relevant information to you. But it's still the human that makes a judgment and look at the data. So I think just jobs will be shifting a bit, and I think humans will go to a higher-value job cycle instead of just having to do the paper pushing.
Right, but some of the tech executives involved in developing AIs, they were saying this publicly on many occasions, so in a sense maybe—and now they're changing.
Look, one of the areas where the stories go about is development.
Developers, for example, those jobs will change dramatically. We at Amazon, we're hiring as much junior developers that we were in the past. Things are not changing. You don't become a senior developer without being a junior developer first. Do these jobs change? Yeah, because there's a number of parts that get automated away. But it means that we can use humans in an area where they are more relevant, and as such, I think the number of jobs won't change. The content of your job may change. Some of the boring stuff will just go away and you'll be left with the things that are way more interesting.
Right. So is this just another case of technological progress not killing jobs? We know that from history, right?
Or is there some risk that eventually—
I don't think so. I think we just get better tools, we become more efficient, some of the boring, repeatable stuff will go away, and we'll be left with jobs that are way more interesting than that they are now.
Right. So let's move to other potential risks, such as losing control of AI. This is another sort of popular concern, like in some of the futuristic movies, such as Terminator. How serious is the risk of an AI agent or a group of agents going rogue in the physical world?
That risk is absolutely there, but that goes for anything that you automate. If you look at agents, agents are a bit different from the regular AI as we know it from the past years. First of all, AI is something that is 50 years old. The term was coined in 1956 by McCarthy and Minsky. And interestingly enough, McCarthy later made this statement that says, "As soon as it works, we don't call it AI anymore." So there's a whole body of work that predates from the past two years before these large language models came into effect. Now, large language models created text, and as such, this could be a meeting where you get text-to-speech and then you get a summary of it and things like that. Those carry little risk. You can maybe look after it at the summary of—
You can verify, as you made a point today.
That's accurate, yeah. But agents are different. Agents take actions, and as such, the other coin of that is risk management. If it's just a meeting summary, if you mistake a little bit, nobody cares that terribly much. If it's marketing content that actually may have commercial effect, maybe have a review cycle on it. If it's on the other side, if it's all the way over to medical systems and people's lives are at risk or personally identifiable information may be exposed, that's much higher risk and, as such, you need to have a much more controlled review cycle about what these agents are allowed to do. So it's really about what are these agents, what are they allowed to do? You set the guardrails and make sure that they cannot go awry. I think some of the examples that we've seen in the past weeks around and the German wiki and things like that demonstrate what the risks are if you don't have the right guards in place.
Yeah. So risk management is the crucial part here, not necessarily what the capabilities of these agents are. If there's high risk, you need to have much more higher level of human control and human judgment to be able to control those agents. But risk tolerance is higher when it comes to war, and we've seen that AI is increasingly being used in the modern weaponry systems: drones, remotely controlled swarms, and all of that. Is that a concern that you know—
Of course, but that goes for any area which has high risk. I mean, indeed, these wartime scenarios are high risk, and as such, human judgment should always be in the loop to make sure that these systems do not go awry or make the wrong decisions. Remember that these large language models, they're not deterministic. They are what I would call—they don't look for the truth; they look for plausibility. Basically, it's a statistical word generator. If this is a table and in the context there are chairs, it's probably physical table. If there's columns and rows in the context, it's maybe a database table or a spreadsheet. So these are just statistical word generators instead of something that has true intelligence. And so in that sense, these are systems that need to be controlled by humans. Now, some of these have very low risk and you can just let them run. If there's high risk involved, humans should be in the loop.
Do you believe the industry as such—and I'm talking about all important companies involved in this—do they now understand this risk first and do they handle it?
I think they understand it. Whether they handle it is a very different story.
Safety is a crucial concern. And of course, these days with such high competition, the desire to move fast is present. And we need to make sure that these companies do not cut any corners when they develop their technology just for the sake of competition. At Amazon, it's really important for us. So we have this environment called Bedrock. Now, for us, it's important not to have one model. We're not a company that sells one model to our customers. We have all these different models available. So for customers, it's really important to experiment to figure out what are the right technologies for the particular problem I want to solve. Do I want the very high-end, expensive model for this particular problem, which is actually quite costly, versus maybe a very cheap model, very smaller, but is good enough for the technology to use? Now, in all of this, guardrails around that are crucial. We need to build systems around it that control these AI systems to make sure that they don't escape the guardrails.
But some governments seem to be in two minds about this because, on the one hand, in the US, for example, the government obviously sees AI as a sort of growth proposition and no guardrails and, you know, let them invent and so on, so forth. But at the same time, the same administration realized at some point that it was risky and they introduced limits on AI. So is awareness among the regulators now at the level that you would consider comfortable?
No, I don't think we're there yet. I don't think the regulators—the most crucial thing in all of this is education, but it's been the case for many other technologies as well. When we started cloud computing, people go like, "Oh, no, you can't do that. You run on that other people's technology." Now it's acceptable. But it all came down to sort of educating the regulator about what the capabilities are and what the true capabilities are of the technology before they start making these kind of restrictive decisions. Now, problem, of course, with the American government at the moment is that they are a little bit unpredictable. And predictability, I think, is really crucial in all of this. I think there's two different models happening. In the US, it's the prototypical capitalist approach: big companies creating wonderful technology, selling it to enterprises and things like that; where, for example, in Asia, whether we think about China, South Korea, Singapore, they have a very different approach. They believe that AI should be fundamental technology that should be available for everyone, which means that very cheap—
It's a public good. And so in that sense, there's two sort of different competing models. We see these very fast, smaller models coming out of, let's say, Asia versus, you know, the big, heavyweight models that are created in the US. And it's just a different approach in both worlds. But we see that there's a lot of interest in sort of not only sort of regulating it, but making sure that this becomes a public good that everybody can use.
Right. So the bottom line on this is that the governments are still grappling with how to regulate, how to monitor, how to supervise and manage this risk that you mentioned.
Yeah. So, you know what? I've just taken a month off, and during that month, I didn't read any tech publications. And then you come back after a month and what has happened in a month, it's incredible. If I get confused, I'm pretty sure my customers get confused and the regulators probably equally so. And as such, really understanding what are the risks in what particular area—you can imagine that using this technology, whether it's military settings or healthcare settings or security settings, that is a much higher risk, and as such, you need a different approach there than if you just have a tool that sort of helps you manage your calendar.
So it's just a different level.
It's just a different level of risk. If for you at home, if you manage your kids' school days, if you manage your shopping, if you manage all these kind of small automation tasks that help you become more efficient, it's clear that that carries little risk. However, healthcare systems are a very different story, of course.
So health, military, biological—
High-risk systems, exposing personally identifiable information—anything that actually would have access to that kind of information requires a very different set of guardrails.
So regulators, as always, are a little bit behind the curve.
Well, you know, plus that I often think, if I look especially in the area of security, I think what actually the regulatory requirements are are often just a baseline. If that would be enough to protect ourselves, we would just all be fine just implementing the regulatory requirements. But I think we as technologists know by now that you need to go for a much higher bar if you really want to protect your customers and your business.
Right. So again, at the end of the day, how worried are you that, let's say, a group of agents or some physical actor uses these agents to, let's say, launch a nuclear attack on a country or release a swarm of drones that will kill people in some city?
I seriously hope not. But you know, technology is just as good as the people that are using it. Who sets off this swarm of drones? Who allows these set of agents to roam free? Who didn't put any guardrails in place? Responsible use of technology—and it's not just about AI, I think in general responsible use of technology depends on the humans, not necessarily on the technology itself.
