Microsoft’s Top Scientist: AI Is More Powerful Than We Think
Eric Horvitz, Microsoft's Chief Scientific Officer, a medical doctor, computer scientist and former White House adviser, talks to Autria Godfrey, Stephen Horn and Laila Rizvi about how far and how fast AI will change society. He argues that today's systems are more powerful than people realise, that some investors will lose heavily even as the technology transforms the economy, and that organisations need to change how they work to take advantage of it.
Transcript
Electricity, the car, the internet. Every technology that has reshaped the world took decades to see societal impact, but can we afford to wait decades for AI's transformative influence?
It's uncomfortable to be living in a transformative era. There's lots of uncertainty ahead.
Eric Horvitz, Microsoft's chief scientific officer, is one of the few people who might just be able to tell us not just what's coming, but when.
Are we just all going to blink one day and be living in a different world?
The psychosocial implications of being in a world where other intelligences that are on par with human intelligence, we really don't understand the possibilities yet.
A trillion dollars has already been spent and it needs a return quickly.
I'm not sure we have five, 10, 15 years to work this out. I'm not sure how many trillions of dollars can be spent before the return comes through.
There may well be overinvestment in particular companies, architectures, and applications, and that's common during any major technological transition.
So how big will the transformation be and how soon will we see it?
We will all be surprised how our great-grandchildren are using AI technologies, and it's hard to predict right now what that'll be like.
Eric Horvitz is a medical doctor and a former White House advisor, and he says that AI won't just be another technological shift, but it will change the trajectory of human existence. So do we think he's right? And if so, when will it happen? All right, Stephen, my question to you is, do you actually want to see a world where we are changing the human experience?
I think it's inevitable. We have this conversation all the time about will it, won't it, start, stop? Is it a flop? And I think it's on its way and it will transform the human experience. How it does that, we're going to have to wait and see. Maybe Eric knows.
I feel like even though I am on this show, I talk to so many people about AI, I still feel so in the dark because I feel as if there's every other day someone else says, "Oh, by the way, I'm going onto the other side where you can't see anything, which is to the frontier labs." And it just seems to be just the US government and the frontier labs that know what's going on. And are we just all going to blink one day and be living in a different world? And I think you have to take people along with you, and I hope that's what Eric can explain, that how do the rest of us come on this journey?
Right, we're crying out for a bit of leadership, a bit of vision, aren't we? I was going to say, I feel like we keep getting these promises of it's going to transform society, it's going to completely change the human existence, yet there are no real-world tangible, this is exactly how your life will look different five years from now.
Yeah, and we also have the blindfold on, and essays aren't going to take that off, unfortunately.
Eric has made some very bold claims about the future when it comes to AI, so let's see if he can give us some of those tangible examples and a vision.
You've made some pretty bold claims about AI changing the trajectory of human existence. Can you give us some specifics on what you think a world that is powered by AI would look like just five years from now?
My comments typically get into the realm of cosmic. I gave a Tanner Lecture on this a few years ago where I reflected about AI being as powerful as a factor as shaping our lives, our civilization as human language was over centuries. And I do think 800 years from now, this time that we're in, including the five years you asked about, will be written in the history of humanity. It's something very important.
We're seeing early diffusion and impressive early results and transformation in areas that I would say are almost amenable to the touch of AI, the digitalized areas where information's already digital, where it's relatively easy to inject AI solutions into the workflows. And this includes the sciences. We're seeing really, I think from a long-term view out, very fast-paced influences in the material sciences, biological sciences, mathematical sciences. It's so fast that those communities are now grappling with core questions about what does it mean to have scientific integrity? What does appropriate citation mean? How should machines be brought into recognition in, for example, creative works discoveries along with humans?
Areas where things are slower, but with this high expectations include healthcare, where we have to think deeply about not just the core capabilities, which are really impressive of today's multimodal models, for example, but we have to sort of understand how do you integrate these tools into a larger system that's more intelligent where you have challenging questions about where you actually inject the technologies into fluid daily workflows, for example, of how healthcare practitioners make inferences and deliver therapies.
There's a lot of people around the world who are uncomfortable with the AI. They're uncomfortable about what's likely to happen to them, what's likely to happen to their jobs, what's likely to happen with the data centers being built, with climate change. And I would respectfully say that what I don't see coming out of Silicon Valley is a vision.
Let's just talk about sentiment about AI, the possibilities and the rough edges and the dangers are communicated. I'd have to agree, we, the AI community and folks interested in the possibilities as well as the potential challenges coming at us quickly given the rate at which this technology is diffusing and the rate at which we see the core capabilities changing and growing. We could do that a lot better. I think we need to be frank about the fact it's uncomfortable to be living in a transformative era. There's lots of uncertainty ahead tend to refer to as the deep currents of the technology.
I and a few colleagues wrote a paper about a year and a half ago on what are 10 priorities for AI research and better understanding of the core science of intelligence and what we need to worry about that and possible catastrophes of various kinds that we've heard about for years from various folks who look at the worst-case scenarios with where AI might go. What are the frog on the frying pan, the warming of the pot that might lead to significant changes socio-technically in how we educate, how we think, what's the status of human agency, our ability to make our own independent decisions as humans, the psychosocial implications of being in a world where there is other intelligences that are on par with human intelligence. I think we really don't understand the possibilities yet and we called for careful research and making deeper investments in understanding early trends, possible valuable mitigations, ways to direct, for example, economic impacts to shared prosperity notions, thinking ahead of the curve.
I know you talked about the uptake of ChatGPT when it first came out and how that was one of the most quickly adopted technologies ever. Do you think what started to happen in terms of some of the backlash, but also the fact that people are not as hopeful about AI is because since we had ChatGPT, people had AI in their pockets, something that they could talk to and something they could interact with, and that was so profound and so revolutionary, whereas they've just kind of become used to that now. And do we need something else where they can actually interact with the frontier to help them understand that vision? Because I feel as if even I'm someone who does interact with AI and I feel like the frontier is really agentic AI, but most people still just see it as the original chatbot. And often it's just government and frontier labs that have access to the most advanced technology. Do you think we need to do something which is more than just saying, "We're going to cure these diseases," but actually show it to people and allow them to engage with it?
We need to separate the way people are perceiving, as you say, AI in the pocket as a conversational agent from the core technological leaps in building systems that can do human-like cognition, abstraction, generalization, creative associations, polymathic abilities across topical areas. And the phrase I like to use is technological overhang, which is that these systems are much more powerful than people think. It's the interface that is rather limited and monocultural in some ways right now in terms of what people are thinking these systems can do. And there's so much possibility in shaping the interface to humanity, to humans, to individuals, to organizations that we are not exploring with the same vigor as, oh, my gosh, there's so much uptake with conversational systems. Let's put more GPUs and more data centers to handling the conversational interaction technology.
These systems have to figure out how they can ground with human to get on the same page, to share context, to jump ahead by understanding what the actual situation is even physically in an environment. We need to basically think about complementarity explicitly. What is it that these systems can do that people are not as good at and vice versa so these systems can shape in and be used in ways that really augment their talents and abilities and their desires for agency, for example, and their values and goals which come from people. And third, coordination. Well, what is the coordination model for how the intelligence is delivered? The if, when, and how, the back and forth. Conversation is a kind of kind of coordination, but it's just one form.
Look at general-purpose cognitive talents, and then look at that as we drive as humanity and as creatives, the shaping of how that touches people and how that touches organizations because that's the bottleneck right now and that's where such so little work has gone on, given all the frenzy and maybe valid hyperbole about the conversational particular way things are being delivered.
Well, we certainly want to know what we should do when the situation arises where AI capabilities are moving at a faster rate than the institutions are able to integrate them. So coming up after the break, we're going to ask Eric about the impedance mismatch. Stay with us.
Back with Microsoft's Chief Scientific Officer, Eric Horvitz. And Eric, I want to know what is your suggestion to the companies or to the people who are trying to implement these AI technologies, but they're really kind of drinking from a fire hose, they can't keep up?
It's always good to look out for best-case uses for surprisingly different applications out there. Keep your eyes open on the frontier of creative innovation with AI. Discover what is being learned and shared by these various forward-deployed engineers, for example, who are basically focused on figuring out how best to leverage AI in different enterprises. It's also very interesting to think about talking to your own IT folks and your own experts.
It's often intriguing to me that when I work with hospital systems, there's often somebody I talk to quite frequently who is like a top expert and we resonate deeply with frontier things happening in medicine. But when I show up to visit or check out what's really going on at the hospital system, there's kind of an insulation or kind of a wall between that expert and the organization where there's not the right level of trust or the right sway given to that individual to make change and to communicate even with the right folks. I would guess that most organizations have already hired or already have top talent in this space, but they're not creating fast enough and wide open enough the channels that let those people that really sort of empower them to engage and to help shape the future for those organizations.
So Eric, we had your very own Matt Firestone on the show just a few weeks ago, and he was basically saying to us that the vast majority of businesses are not taking up the AI tools that are available. And the reason for that sort of touches on what you've just said, it's really how business processes interact with the AI model. So how the actual companies themselves are shaped to take... And if you look at any industrial revolution, that's the killer app, if you like, is how the company organizes itself around the technology. Now I run a company, and I'm lucky enough to have some people around me who might know what's going on, but how does that piece happen? Because there are millions of companies out there, and if that piece doesn't happen, then the payback for the investment doesn't happen.
You mentioned the Industrial Revolution and what we saw back then, or even waves, Second Industrial Revolution, the electrification revolution, it took a whole bunch of time to go from this, and it took decades before this idea that you can have an actual high-powered motor at every personal station. You didn't need to do the kind of distribution. And that analogy I think holds right now. I also, it's just hard to move aircraft carriers and to get them to organizationally shift fast enough.
I do think there'll be a shaking out where there'll be best practices and orgs that will be taking those new motors and putting them in the right places and understanding how to completely refactor their operations and workflows to take advantage of the best developing technologies. Some will be universal in terms of, for example, how does AI augment the intellect in certain areas of your work stream? How can you reduce the loss of time and effort based on certain kinds of errors that show up and so on? There's a whole set of categories like this, but it's only been like three years, three and a half years since we saw some of the limbers of these sparks of artificial general intelligence showing up, which surprised everybody. How quickly can people really integrate the implications of that in an organization that's been around for decades? Even if you look at this and you go, "Huh, wow. Uncertainty, what does this mean?" There's still, I think, a lot of emotional, at the emotional level, skepticism that this is as big as it's going to be.
I buy that on two levels, right? I buy what you say about almost a tipping point where the new organization best practice, if you like, becomes apparent and then companies are either with that or they're not, and the ones who are with that, it's a shakeout, as you say, a tipping point. I buy that. Where I differ with you on this one is the size of the bet, the scale of the investment. I'm not sure we have five, 10, 15 years to work this out. I'm not sure how many trillions of dollars can be spent before the return comes through. And I believe that the return only comes through when that investment pays off and why I'm keen to figure out that tipping point of yours.
We have smart business folks that are trying to do their best to navigate this risky decision under uncertainty as to when and how do I invest and change given I might be too early and I might be too late. I see that again and again in our research developments. What's the right timing for the investments? There may well be overinvestment in particular companies, architectures, and applications, and that's common during any major technological transition. But I think it'd be a mistake to infer that there is overinvestment in the underlying scientific and technological opportunities, the exploration we need to be doing.
There's some historical analogies. I think with railroads, electricity, automobiles, and the internet, all these big changes where technology started to interleave into society in ways that everybody knew would be big someday, but wasn't sure when and how. It's this idea of thinking through what does it mean to be smart about when to capital investment when you're potentially running ahead of immediately realizable value. There'll be investors over the decades when this kind of thing has happened who lose heavily, even though the infrastructure they financed ultimately transformed the economy.
Let's just say that I love the realm of AI in the open world, it's my research area, decision-making under uncertainty, actions under incomplete and uncertain situations, that there is uncertainty. And some companies are leaning in now with some intuitions and insights that now's the time to go big. Some people will be hurt by that.
The thing about AI is that it's constantly shifting beneath our feet, whereas I feel like those other technologies were almost more static in a way. I found that AI functions in such counterintuitive ways comparable to counterintuitive concepts in quantum mechanics. Like when steering the AI, it's actually performs worse because it delivers me what I want to hear instead of what is actually useful to me. So how much do you think that's going to slow us down and how much do you think that's going to impact society? And are there any other counterintuitive things that you've seen emerge in AI as a technology?
We need to think about what our core design principles, what it is that we want from AI, which might be durable over time versus the technology itself as it's changing in its abilities and how it's being leveraged. I think we have to design for continuous learning and adaptation. That means we have to evaluate capabilities repeatedly as they change. We have to monitor how they behave in real-world settings. It means we have to also update safeguards when new risks appear. We have to be willing to revise workflows and norms as we learn.
We need to do some, as back to my comment, preserve human judgment, preserve human agency, pursue principles of augmentation of human cognition, make sure that systems are designed that can adapt themselves to different organizations over time, even when they change, that help us to integrate them. I think when technology is changing so quickly as AI will be doing over time, I think resiliency comes from having the right feedback loops, not necessarily fixed predictions, even though predictions at any moment can be guiding us to take the next steps. I think the answer is, given rapidly changing AI, is not to try to predict every capability in advance. It's going to become better at detecting change, learning from it and responding to it.
How do we do that en masse and not just in the frontier lab? Because to transform society, we have to take everyone with us. So how does everyone keep up with that and those changes are happening under our feet?
I've always been a fan of getting in place a diversity of studies, evaluations, and perspectives, whether it be including civil society groups like the Partnership on AI, government agencies that are charged with different countries with understanding and helping with safety and usability of these tools, and the companies themselves, the organizations charged with delivering technologies. I think third parties, the consumer unions and reports of the world someday will be weighing in more heavily to help people understand to do these kinds of analyses, other kinds of auditing programs.
Let me just say that I'd like to put it on the technology producers to do more to make these systems better declare what the capabilities are to better educate end users in how they work rather than putting up a cursor and saying question, please or say anything you want. We, Microsoft and other companies and even the open-source providers can do a lot more when they build these interesting applications that touch people in different ways and that provide different kinds of tools and skills with best practices.
As the chief scientific officer, I would love to hear your thoughts on where you see the future of AI going with regard to do you see world models being the next big thing? Do you think that LLMs have a short runway or will it be getting LLMs to the point of mimicking human reasoning?
We've all been surprised by the power of what we're seeing with multi-layered neural networks. I think that core component is here to stay and we're seeing the ability to learn general input-output functions of life, predict distributions of various kinds and to generate. I think that's going to be not going away. However, the architecture we have now, the transformer and how it's being used, I think will evolve. We'll have different architectural approaches. We'll be learning more from data about the real world and that will induce richer world models.
These systems, I believe, are already inducing rich world models from their training. I think that's the most impressive and surprising aspects of their power that they seem to be... We're pushing them to sort of, I think, inflate and learn about the world in ways that we don't understand yet. So I'm not a big fan of saying we don't have world models, we need extra world models. I think we're already seeing some of this magic, and I quote magic because not even the leading AI experts understand what these systems are doing inside that's giving them the power they have. That's why in our top priorities, 10 priorities, the number one was we need to invest heavily in understanding these systems and figuring out what it is that they're doing when we push them in basic ways with optimization algorithms. That will unlock more alternate architectures and new approaches that might take us further.
You've spoken about surprises, and we've had a lot of those in the field of AI, and you've also spoken about the edge of doability. What do you think is the edge of doability now, and what are we going to be surprised by?
One thing I'm sure about is that we will be surprised by where AI goes, not just the core technology and capabilities, but how it's being harnessed by people in society. There'll be some several big surprises ahead, I hope mostly fabulous ones. I think we often sort of prepare to guide the technology and to regulate it. We will all be surprised how our great-grandchildren are using AI technologies, and it's hard to predict right now what that'll be like.
Where do you think people will draw the line on inviting AI into their personal lives?
We're already seeing in folks using these tools in writing lines in the sand about my ideas, my core flows of what I want to say, my thinking versus polishing and helping me with communication. So I think that we'll see standards of practice growing as to when it comes to communications or writing. People will not want these tools thinking for them. Just a line in the sand that I think we want to not just celebrate, we want to build tools that understand that too.
What's something that's universally accepted in your field that you actually disagree with?
That these systems have nothing to do with human cognition in terms of how they work.
What are you optimistic about?
The long-term future of humanity with these tools. It's going to be these are game-changing tools at the level of human language, and I think that we have no idea how big this wave is going to be, but it's going to take hundreds of years, if not thousands.
That was wide-ranging and fascinating. I feel like we touched on a lot, but there was still a lot more we could have gotten to. Your biggest takeaways from our conversation with Eric?
I think as in Eric, in terms of the vision that he has for the future is a hopeful one and a much more hopeful and large scale one than I think most people do. As in, I think what I got from that is the skepticism that people have around these models and the fact that we are going to have to take another path. It's not worth going down this path. It's going to be really, really difficult. I think he doesn't agree with that necessarily.
Did you get a vision out of him?
Yeah, so I think Eric's obviously a super smart guy, right, and he has a big vision, but he has a big vision for his great-grandchildren, right? And we don't seem to have a vision for how we get to his great-grandchildren. And that, for me, I think is the challenge.
All right. Well, we appreciate Eric's time with us today. Thanks so much for watching, and be sure to leave your comments below. What did you think about our conversation with Eric? Do you think that we are headed toward a life-altering human existence? Weigh in, and while you're there, be sure to like and subscribe. We'll see you next week on Agents of Tech.