John Williams Feature

JOHN WILLIAMS: Hi, I'm John Williams. I'm a professor of information engineering in the Department of Civil and Environmental Engineering at MIT. And my background is basically in physics.

But I grew up in the UK. I'm from Wales, which is, basically, steel and mining. So I ended up in Oxford University, played rugby there for Oxford. So I got an Oxford blue, they call it. And we happened to beat the Springboks, who were the world champions at the moment in rugby.

But that was a long time ago. And so my journey went through UCLA, where I studied physics, basically, got fascinated with how we could simulate physics. I ended up going back to Wales and did a PhD in Swansea University.

And that was the place where they invented the finite element method. A guy called Zienkiewicz, Professor Zienkiewicz there was the world renown in finite elements. And so I became embroiled, if you like, and fascinated by simulation.

After the PhD, I went into industry for eight years and did two startups, one in Colorado. And we basically were simulating the oil rigs up in the Arctic. So I spent some time in the Arctic, which is a place where the one thing you want to do is get out of there. It's great for about three days, and then it becomes pretty hellish.

So I spent time on offshore rigs there, and we were simulating, in our company, the impact of ice on offshore rigs. And the big thing about ice mechanics is that it's not continuous, that you've got-- basically, most of the mechanics is based on this continuum theory. And the assumptions of continuum theory don't hold for ice, that you've got it fracturing, you've got pieces of ice.

And the key problem with the offshore rigs was that the ice would ride up on the sides and possibly take the deck off. The ice was about 13-foot thick. So these are not small pieces of ice. This is the height of a room.

And, when they hit you, they shake the whole rig. My first night on the rig, I was really scared. It was shaking. It was like a leaf.

But the structures, they've done a pretty good job, I guess, of designing them because they survived. But that's how I became embroiled in or fascinated by non-continuum mechanics. So this is where you've got pieces like-- you might have a soil or sand or powders, so drugs, for example.

And, in the company, we invented something called the discrete element method, which is basically lots of particles. And the trick is how to figure out how they interact. So, computationally, it's quite expensive. But we made some-- we developed a commercial code, actually, called Sea Ice that was sold to the oil companies.

And so, after that, we sold the company. Didn't make a lot of money. Had lots of shares that we could paper walls with, but, eventually, they came good.

But then I applied to MIT for a job in MIT, and I became a research scientist here. So that's how I ended up in MIT. That's the short story.

[MUSIC PLAYING]

[MUSIC PLAYING]

JOHN WILLIAMS: Basically, we developed a code that handled bodies that were interacting. So it turned out that we could actually not only simulate things like soils and ice, that we could simulate computer systems. So we ended up working for Ford on simulating their global network of computer systems.

They wanted to know whether they could move files around the world fast enough that the engineers would have, say, a two or three-second delay. It turned out that they couldn't get within an hour. They couldn't get up-to-date data. You could get the bits across the world in a second or so. So from Australia, say, they'd have a design team working there. And they'd ship the designs, then, say, to Detroit.

But the problem was to actually put that data into a database, that took time. Because it needed indexing. And that was the bottleneck. But we were able, then, to simulate systems, pretty complex systems. And I got put in charge of something called the system design and management program here at MIT.

And the goal behind that was we wanted to understand how systems behaved. It was clear that at the component level, we knew how to design things pretty well. But when you start putting them into systems, it's not clear how those systems behave.

And that comes to the problem today. We're building systems, say, with generative AI. And the question is, how will those systems behave? So it's not only talking to one AI. You have AIs talking to other AIs. And we don't really understand how those systems will behave.

Now, the world leaders in systems at the Santa Fe Institute-- and they came up with the idea that there are four kinds of systems. So there are what we call simple systems, that we can reproduce experiments on them. They're highly predictable.

So you can imagine a clock pendulum swinging. It's highly predictable. We know how to do that. The planets, to some extent, are predictable like that. We know where the sun is going to be. We know where the moon is going to be.

And so there's this idea of the clockwork universe that's very predictable. And I was brought up with that kind of science, Newtonian physics. But it's clear that there are other systems.

So the second type of system is what we call a complex system. It's kind of predictable. So you might consider, for example, fluid flow. Where it becomes unpredictable is, say, when it flows over an airplane wing, and you get turbulence. And it turns out that turbulence is not predictive in the same sense. Overall, you might get within 10% of the correct answer. But the details are not so clear.

The third type of system is what we call complex adaptive. And that's a system that is never in the same state twice. For example, planet Earth is never going to be in the same state as it's in today ever again, that things will change. And they're difficult because you can't reproduce experiments. Because they're never in the same state.

So once you're at complex adaptive systems, you've got real problems that you can't apply traditional methods. To some extent, science is difficult. And some of those problems, they term them as "trans-science," that you'd think they're kind of scientifically approachable. But it turns out you can't do the experiments. Climate would be an example that we'd be very reticent to do experiments on the planet, for example, pushing CO2 or SO2 up into the atmosphere to cool the planet.

And so complex adaptive systems take us into a very different area. The worst kind of system is chaotic. Those kind of systems, minor perturbations end up in very different results. So the problem today is that we're building systems, and we often don't know their properties. Generative AI would be one.

If we have these AIs talking to AIs, we won't know exactly how that system is going to behave. Perhaps you get hallucination. Maybe you don't. And so it's clear that these systems, the problem is, how are we going to test them? How are we going to align them? How are we going to constrain them? And that's the problem we're wrestling with today, is how to simulate these complex or complex adaptive systems that may be chaotic at some stages.

[MUSIC PLAYING]

JOHN WILLIAMS: We teach a course on digital transformation. And basically, it's from the computational perspective-- computers. Software is changing the world. And companies are struggling with how to implement this digital transformation. You can even see it here at MIT.

We're running a data center that really is not an up-to-date data center. And that we need to understand, Should we use cloud computing or not? So let me give you my condensed version of what's happening in the world that in the early 2000s-- sorry, in the late '90s, 1990s, we had the World Wide Web. So this is Sir Tim Berners-Lee, that sits across the hall here, invented, basically, the World Wide Web. So he invented HTML. There's a protocol called HTTP so computers can talk to each other.

And we had the ability then to get documents across the world. He was sitting in CERN, and scientists in the US needed to get the data from CERN in Switzerland. And they couldn't. They had to phone him up. And so he invented this system whereby they didn't need to do that. He invented an addressing system called URLs. This is the HTTP stuff.

And then we have the ability to have one computer ask another computer, give me this data. And so the World Wide Web was the start of an explosion. And we end up now buying almost everything on the web. It's changed the whole of industry. It's changed how we advertise products. It's changed how we sell products.

Based on top of that, Google came along and started to map out all these URLs so you could look up information. And Google realized that it had spare computing power. So along came the idea of cloud computing, where you have these big data centers and they're not utilized 100%, so they can rent out their computation. So that was started in about 2000, 2001.

Since then, cloud computing has become dominant. In the US, we have Google, we have Azure, which is Microsoft's cloud, and we have AWS, Amazon's cloud. Amazon is the largest by-- it's probably double anyone else. Azure is the second biggest, and Google is the third biggest. But these clouds allow us now to get rid of our desktop computers, our servers under the desk. We can just go to the cloud.

So this allowed us to manipulate large amounts of data. So the largest file I'd ever manipulated two years ago was about a 100 million line file. Now, with Google and their BigQuery, in my last demo I did, I could search a 55 billion line file, so 55 billion rows. So large data now becomes, we can handle it, we can analyze it. It's still a problem.

Most companies are setting up data pipelines so that they can make decisions based on data, the term data-driven decision making. And you'd like to have control over what data you're producing within your company so you can see how you're performing. Now, if you think of us with our students, what data do we have? Not that much. Whether they had A's or B's in classes, but we know very little about them, and they're our product.

So companies wanted master data. And it turns out, now that you have large data, you can apply machine learning. So machine learning comes along in 2012. You've got Geoffrey Hinton in University of Toronto that basically shows that we can identify objects and images. So machine learning took off and has now resulted in generative AI. So that was the next step.

At the same time, we had cybersecurity coming in. The cyber attacks became prevalent. The World Wide Web and the internet wasn't designed really to be secure. In the early days, it was designed so that it was robust, that you could take out a router, and you could still get the messages. They'd find their way around.

So we have these technologies, and on top of these, in about 2000, the Internet of Things was invented here at the Auto-ID Lab. So Sanjay Sarma was running it initially, and then he handed it over to me because it became more involved with software than with actual tags, the RFID tags that they were putting on products to track them.

So you have all of these technologies coming into play with companies. So the digital transformation is basically about how to handle these and how to handle people.

[MUSIC PLAYING]

JOHN WILLIAMS: SPEAKER: So you come back to a systems view of how people operate. And it really started back with the Toyota production system. So this was Toyota came. They didn't know how to build cars. They saw what was happening in America. They decided to understand how to build automobiles.

And by the late '80s, they were outperforming the US with automobiles. And their method of controlling it, they called it the Toyota production system, became what we now call lean engineering. And there's a famous book here by Dan Roos and James Womack called The Machine that Changed the World. And really, lean engineering became the prevalent way of manufacturing goods.

And it was kind of based on just in time. Instead of pushing your materials onto the shop floor, you pulled them. So you had less spare materials. It was more efficient.

What's happened now is that we've begun to realize that the world is changing so quickly that we can't apply the same design methods that we used to apply. So we used to apply waterfall methods, they call them, where you come out with a specification of what you want. And then you figure out the functional requirements, they call them. And then you start to design it, and you come up with a design parameters.

So this is basically, this famous professor here, Nam Suh, that came up with axiomatic design. So you separate what you want from how you're going to do it. And that has become prevalent in the software area and gave rise to what we call DevOps, where companies now can deploy live code 10, 12, 20 times a day.

So you've got companies-- like, the poster child would be Netflix. They can deploy live code and make changes in their pipelines. So they have about 100 pipelines, data pipelines. So this is the problem that companies are faced with. They need to set up data pipelines to have the data that they need to make decisions.

And then they need to understand. They're usually applying some machine learning to that data because it's so large, that humans can't actually look at it. I can't load that 55-billion-line file into my browser and look at everything. So we've got this problem now of the large data pipelines that we need help with to understand how they behave.

Again, you've got companies like Uber, that can make real-time decisions about-- I order a cab. It's going to tell me in a couple of seconds the cost of that cab. They're going to do a calculation of what kind of cars they've got around, what are available. And they're looking at the road conditions. Are there bottlenecks? Is there traffic congestion? And they'll come up with a price in a couple of seconds. And they're doing millions of calculations to do that.

So this is the problem of digital transformation. Depending on the size of your company, maybe it's easy. It's easy if you've got six people to say, yeah, we're going to become agile. If you've got 10,000 people, you've got a different problem.

And so corporations are struggling with how to make these changes. Part of the problem is that the C-suite now, it's always been the command and control type structure, that the top people make the decisions, and the people lower down, execute. Turns out now the top people don't have the information or don't have the background to, say, to make the decisions lower down about, what software do we use in this pipeline? The CEO isn't going to be able to make that decision.

So there's a change happening where more and more companies are going away from command and control to actually pushing decision-making down. Now, that itself is difficult to do, to have decisions being made and percolated up. And companies are struggling with that at the moment.

It's pretty clear that the customers are changing. If you think about the younger students, they're very different to a generation ago. And it seems that the smartphones and social media have really had an impact on that, the Gen Z. And there's several books now out about these changes.

The mental problems that these young kids are having are double what they used to be. And it's not the kids' fault. It's the fact that they're having a very different upbringing due to cell phones. Particularly, young girls, they're doing self harm.

And this is part of the problem, I think, that we're seeing at the universities, that these changes are very real, that the students are coming out without the preparation of social engagement that we used to have. We didn't have cell phones, let alone smartphones. So you've got all of these things contributing to this dynamic of OK, how do we adapt to these technologies?

[MUSIC PLAYING]

JOHN WILLIAMS: Yeah, you know yourself that it's really difficult to give up knowledge that you've acquired. So when I was doing my PhD, there were certain bottlenecks. And you knew that you had to challenge those. Now, the bottlenecks have changed. For companies, they really are struggling with how to manage their people.

And it's people, processes, and stuff, technology. And the people are the big problem. They're the solution as well. But it's really difficult now to motivate people. People move companies pretty regularly.

We're used to a high standard of living. And so we've got issues with how do you balance your life and your family. And all of these things are kind of interacting and in ways where we've got very little data about-- like I say, the Gen Z are coming through. We're just recognizing the problems now.

And so to some extent, in a lot of areas, we don't have the data to make the decisions. So you're in the position of having to learn. So a lot of the products we're putting out there that now are to learn from, that there's some kind of-- there was a couple in Singapore that wanted to make a bread maker.

And so what they did was they consciously decided they were going to embed software so that they can talk to this machine, that they could have data fed back into how the bread came out so that the user could hit a smiley face and give them feedback. Anyway, basically, they set up a learning system where they learned better than anyone else how to make bread.

And that's what we're doing with a lot of products today. We're trying to push them out and build into them the kind of facilities so that we can learn how they're being used and can adjust. So the strategy today is to go out with minimum viable products that you get out there, say, with your product, and then you learn from it. And hopefully, you can learn fast enough that you're going to outperform your competitors.

But that kind of model doesn't allow for this five-year plan. And of course, the people at the top of the company, they need to be able to make some plans, at least for a year or two. You can't not make the plans. But it's pretty clear that it's more like you set off on a journey, like driving across, say, from Boston to LA, that you know you better head west. But if the road is closed, you'll go around. And you'll go down a different road. And you can still get to your destination. But it's not like you can plan every step of it.

And so most companies now are in this mode of where you know you've got a general direction, but you can't specify the details exactly. It's very difficult. Imagine now with this generative AI that you want to develop a product for your corporation that everyone's going to use in your company. So this is enterprise scale. It's not just group scale or personal scale. It's going to be that kind of product.

How much is it going to cost you? How you do that calculation, I don't know. So you can't really cost things out in the way that you used to be able to. So a lot of it is about making the best decision you can. And we're seeing all kinds of companies struggling. GE, for example, spent 4 billion on a product that didn't go anywhere. It's really tough out there to use these tools in a sensible way that you can survive.

If you look at the data about the top 500 companies that now, the average period they're in, the top 500 has gone down from 20 years to something like 5 years or 10 years now. So it's the defenses that large companies used to be able to put up are almost impossible to put up these days.

You've seen it with Microsoft, missed the whole of the smartphone. They nearly went out of business. Apple nearly went out of business. They managed to recover, partly because that Microsoft kept them alive, so there wouldn't be antitrust. But Facebook, they went after the metaverse. And it looks like maybe it's going to be there in the future, but the timing was off. Now, they're doing generative AI pretty well. So they'll put their eggs in that basket.

So my sense is that all companies are being forced into being agile in some way. I don't like-- in software, agile has got a specific meaning, the agile methodology, is what all the consultants sell to you. It's got the scrum method. So it's not so much that as being aware that you're dealing with a world that's changing quickly.

[MUSIC PLAYING]