Steve recently gave a guest lecture to Professor Shruti Sharma’s introductory class on optimization for graduate students at the NYU Tandon School of Engineering, in the Technology Management and Innovation Department. Following are lightly edited excerpts.
Scalable
“Scalable" is something that everyone your age has heard a bunch of times. For my generation and older, it wasn’t a concept—we didn't hear about “scalable.”
Today, scalable to me means you have a couple of people at a startup in Silicon Valley and you’ve created something, but you’re kind of doing it by hand. Say your startup is like Uber but in the garage you only have three drivers, and you have five customers, two of whom are actually your friends. You don't need fancy stuff. Maybe your app looks good, but in the back you’re just doing this whole thing in your hand. That is not scalable.
As a launching point, optimization lets you build something that scales. You can always buy more computers; you can always put them in parallel. Optimization solves the problem the same way every single time. You don’t have to hire 50 people with headsets in a call center and then create rules and regulations and list them in training manuals. Not only does that take a lot of time, but maybe the rules for training when you were five people are different when you’re 50 people or when you're 5,000 people. What about when you go into different countries? In that case, with optimization software, the interface may have to change, but the actual core doesn't. Optimization is smarter and faster—and scalable.
Optimization is a Subset of AI
My colleagues and I think of Artificial Intelligence as a subset of data science, not as an alternative path in the road, but a subset. Not everything data science is AI. Running a time-series statistical analysis, that’s not AI. Why not? AI is concerned with automating intellectual tasks.
In my view, optimization, similarly, is a subset of AI. Not everything in AI is optimization. “Is this a picture of a cat?” is one the gateway AI applications; it is not optimization, which essentially makes recommendations, such as: “Here is what I think your next move should be”; “Here is how I would arrange this”; “Here is what I would do.”
Now, I'll grant that a lot of the new talk about AI is around certain algorithms like ChatGPT. I beg you, don't confuse the algorithm with the field. You are learning in this class about linear programming and mixed-integer programming and those kinds of techniques. They aren't large language models but, in my view, it doesn't matter. If you're trying to find a trend, you evaluate which statistics to use. It depends. You have to look at the problem and find the proper technique. There are areas where large language models, as we're all finding out, are crazy brilliant, but there are also a lot of areas where they are completely unable to help.
When I explain optimization to a business audience, I discuss “assets,” which I define as things of value that you have some control over, even though you might not own them. For example, people are one of the most important assets, whether they are employees or customers, but you don't own them. There are different classes of assets and you are making various decisions about them.
In a textbook or an introductory class about optimization, there is typically presented a simple problem like the traveling salesman, and the objective is to minimize miles. In the real world, that's not how a company thinks. Executives want a high-quality solution that minimizes waste, yes, but also, for example, yields nimbleness. It is much better for most businesses to come up with a quick answer that is pretty good than the perfect answer that takes an hour. When I get on my Uber app and request a ride, Uber doesn’t have an hour to decide the perfect mix. The ability to make fast decisions under different conditions is critical. If you had to pick just one business objective, most executives would say, “Well, at the end of the day, it’s about profitability, isn’t it?” I don’t know--Amazon for many years didn't run a profit. My point is that it’s a multi-objective that you’re typically trying to achieve.
Decision or Rule?
I hope that you here today either go into our field and do this kind of work or you become important executives and greenlight projects from people like us and we’ll be working for you! Either way, we’re at a very exciting time with everyone talking about AI. In my view, that is what you’re studying here in this intro to optimization course: this is AI.
Making better decisions is essentially optimization. If you ask most practitioners what we do, it's hard not to use the word “decisions.” Most people would say that a decision is something you think about, and then make a choice. However, consider that there are a lot of instances where people do not perceive a decision is being made, because there is an action based on a “rule” that someone created artificially at some point in the past. You might be thinking I’m splicing things thin, but I can tell you that if you ask businesspeople to show you decisions they are making so you can use optimization to help, you will find a lot of times the decisions are trivial at the end. Often, there is not a lot of savings. So, you will have to wind back.
For example, before Uber, if you weren’t hailing a cab from the curb, to request a ride you called a taxi company. In many cases, the dispatcher would get on the radio and broadcast that a pickup was wanted at your address. The fleet’s drivers listened to their radios and one would respond that he was available. Accepted wisdom at that time was that the dispatcher was not making a decision, he was just making a ride available. In fact, there was a decision being made--which driver should be assigned to the ride--based on a previously established rule.
If you become a practitioner, you will look at businesses and search for what might be possible, not simply relying on the way they see the problem decomposed at the present. You will look at the whole problem. That's where Uber came from. As I was saying before, it used to be that if you were a dispatcher, you didn't know where the cars were, so how could you assign someone? But now, everyone's got a phone and now we know where everyone is—that gives us an opportunity to better assign and route, to carve up the problem. I believe that you all are going to find over the next five or ten years lots of opportunities. They might be subtle, and you might have to take a very naive view of a problem from the beginning to end. If you find a case similar to old-school taxi dispatching where there was no assignment, no decision, only a rule—you may well find there should be a decision about an important asset, and that optimization can improve it and yield tremendous value.