Hello, and welcome back to The Future in 5 Questions. This Friday, we have Neil Thompson — director of the FutureTech project at the Massachusetts Institute of Technology’s computer science and artificial intelligence lab and a principal investigator at MIT’s Initiative on the Digital Economy. Thompson’s expertise lies in the intersection of industry and the next generation of computing, which is especially relevant now, given the burgeoning resource footprint of AI. He has advised businesses and governments about the future of Moore’s Law — an empirical principle that explains the speed and capability of computers over time. Read on to hear Thompson’s thoughts about the physics limit of current chips, how the U.S. can keep its lead in computing power, and how researchers use large language models. Responses have been edited for length and clarity. What’s one underrated big idea? A big underrated idea is that the pace of computer progress is slowing down. This, I think, is very counterintuitive, because we keep seeing results in the news about how incredible AI is and what it's doing. But the underpinnings have changed. Almost all of this stuff that we're doing with artificial intelligence is based on this technology called deep learning. Deep learning has this characteristic where you spend a ton more to do a ton more. That's different from how we historically got improvements in computing, which was through Moore's law and improvements in computer hardware. In that case, you didn't have to spend more, to get much more. So now, there's an intrinsic cap on what you're able to do with deep learning and just spending more. Suddenly, you get systems that cost so enormously much that you don't really want to keep investing in the “more.” For example, we're already hearing that GPT-4 cost more than $100 million to train. And people there are thinking, well, actually, there probably won't be another step in the same pattern because of the escalation in costs. That's notably different compared to the decades and decades of improvement we've had from Moore's law. And I think we're going to be reckoning with that a lot more in the coming sort of 5-10 years. What’s a technology you think is overhyped? Blockchain. In the discussions I've had with people, I keep coming back to this idea of why is this better than a database? There are some cases like cryptocurrencies where you can make a strong case for it. But in many of the things that people are talking about, the distributed nature of blockchain actually has some drawbacks. Because you often actually want to have some way of making changes. You want to be able to say, well, that transaction actually is fraudulent. Like, it may be that people on both sides of the transaction agreed to a transfer. But if the thing wasn't actually delivered, you need some way to reverse it. And so I think many of the benefits of blockchain are really benefits that come from databases. It's only a small number of cases where you really get the incremental benefit that comes from all the machinery of blockchain. What book most shaped your conception of the future? I want to answer in a slightly different way than you've asked. I want to talk about a graph that shaped my view of the future. This was a graph that Horst Simon, who used to be the head of computation at Lawrence Berkeley National Lab, presented and I saw it in a seminar one time. It was this graph of Moore's law as it evolved over time. Moore's law is meant to characterize the sort of doubling of computing power every couple of years. But actually there's an underlying trend — which I think is even more important. Probably twenty years beforehand, Richard Feynman identified miniaturization as a key thing that could be done. Turns out that the miniaturization of parts of the computer has many benefits. One is that you can just get much more on the chip. And that’s just geometry, right. If you shrink it, the same-sized chip can hold more. But perhaps a more important piece is that as transistors get smaller, they actually use less power, even proportionally. And that meant that you could run chips faster. And that's hugely important. In fact, for decades, we had exponential increases in speed. And Horst Simon’s graph was showing that even though we have continued to miniaturize, the physics that meant we used less power stopped. So that power use has plateaued. That means that the speed of our chips has plateaued. This is why chips today — and computers today — run at about the same speed they did even 15 years ago. And that really suggested just a complete change to me. Horst Simon had shown that the nature of Moore's law had gone from a whole bunch of things improving all at the same time to the speed capping out. That led me to realize that actually, Moore's law — which used to benefit everyone who's running programs — was now really differentially benefiting people who had parallel programs and much less the people who were doing things serially. And unfortunately, you know, most of the things we're doing are serial. That led me to do my PhD work showing that, indeed, there was a break there. What could government be doing regarding tech that it isn’t? This really comes back to my first answer about the pace of computer progress slowing. We need to be investing a lot more — and by a lot more, I mean ten, a hundred times more — in trying to figure out what the next era of computing looks like. That's because computing has already been slowing for more than 15 years. But it's worth a ton to the U.S. economy. And right now, we're not investing anywhere near enough in thinking about how to improve those things. We recently started to take good first steps with the recent CHIPS Act. But there's much more we need to be doing in terms of post-CMOS technologies [the circuitry designs and fabrication processes needed to make the next generation of semiconductors.] The history of computing is one that the United States has been incredibly dominant in. A huge proportion of the algorithms that have pushed computing forward have come out in the United States. Many of the biggest supercomputers have been here. That overflows into all these other areas of society and gives them benefits. We're actually really losing that lead. We need to make sure we have good secure factories that can produce cutting-edge semiconductors. The CHIPS Act covers that. And people are starting to invest in some of these post-CMOS technologies — but it just needs to be much more. These are incredibly important technologies. This is not just about AI —these are some of the fundamental things that define what the next generation of computers will look like. Could we make them 1,000 times more power efficient, 100 times faster? Those kinds of questions. What has surprised you most this year? I'm sure this answer you get frequently, but it would be how effective large language models have been. It was clear that these deep learning systems were growing very rapidly. And indeed, we have done some work in the lab on those. What we've seen is that there's enormous escalation in the data — and therefore the cost — of these models. And, boy, has that escalated quickly. OpenAI recently said GPT4 cost more than $100 million. So that's pretty shocking. But at the same time, it is truly remarkable the number of things that they can do — how useful that is to people. I'll give you a very practical example that I thought was really interesting. One of my students, when he's trying to make an argument in a paper, he will put the paragraph in ChatGPT and say, “What would a reviewer in economics say, if they read this and wanted to critique it?” And so it will then give an argument. Sometimes he'll agree with it and sometimes he won't. But he said, typically, he'll get at least one or two good ideas of ways to better communicate what he’s written. And that seemed to me like an incredibly interesting usage that I hadn't ever thought of. |