Return to site

Wharton Professor & Microsoft Azure Chief Economist Dr. Amit Gandhi Discusses Life & Tech

· Microsoft,Acadamia

Wharton Professor & Microsoft Azure Chief Economist Dr. Amit Gandhi Discusses Life & Tech

broken image

RR: Could you walk us through your personal and educational experiences?

I was born and raised in Ohio. My parents were physicians from India. Unsurprisingly, as a child of immigrants and physicians, there was a lot of pressure to become a doctor. However, sometime around high school, I recognized my own tendency to analyze problems and topics for extended periods of time and from multiple angles.

A typical high school course works by progressing through a sequence of topics, whereas I would get hung up on potentially the first lecture or chapter and I would often be tempted to research the history of an idea rather than keep pace with the class! My grades would suffer at times and my awareness of my own curiosity made me realize I was a) not especially efficient at plowing through work and b) this would bode badly for a future career as a physician if I spent all my time on the first patient!

This overall pattern spanned from the obvious subjects like math, but also history. I also discovered the philosophy of science and how knowledge progresses (e.g., a history teacher exposed me to Thomas Kuhn’s “Structure of Scientific Revolutions”) which is not in a linear progressive fashion but rather marked with short bursts of immense upheaval in how a subject is understood. An appreciation for epistemology and philosophy of science definitely influenced my collegiate and graduate school journeys.

I also placed an emphasis on what later in life as an economist I would understand as “option value” - I did not want to commit to a particular path such as being a doctor or lawyer. That’s what led me to want to pursue a liberal arts education to really make sure that I had time to think and canvas the space of disciplines and domains. That was a tension in my family because I was becoming sort of a rogue actor to them - by not committing to a path, I was taking a risk, and that is a hard situation to contend for first generation Indian parents.

Once I got into college, my interests shifted towards business tracks and undergraduate business courses. But even there, I was hesitant to commit to business school, and wanted to hold onto that scientific mindset to think through business problems from first principles. Economics ended up being a great blend of that liberal arts mindset, history, society, yet math, computer science, and business. It had all the complexion of what I wanted in terms of the narrative, methods, and tools.

At some point, I hit another crossroads when I took an impactful course at Harvard with Professor Garry Chamberlain, who recently passed. He taught the course very simply: how do you make an economic decision with some data using economics and statistics to arrive at an optimal solution. It was very practical, but also real and profound. He recommended I attend the University of Chicago business school; they had built a fortress of applying Bayesian methods and tools to business questions.

One thing I have learned over and again as you reach different life milestones is that you get to a point, and start from scratch and go through the same cycle: you climb up during your undergrad, then you’re shot down to size during the first year of your graduate program where you are amongst the best and brightest in your field, and you climb again. The same cycle happened when I became an assistant professor to when I got tenured.

Then I took a journey into the technology sector where I went to Microsoft to bring an economics program and capability to their cloud business, and was humbled again by working amongst a new set of experts. Throughout all these experiences, I’ve used the compass that I want to keep pushing my own personal growth and avoid stagnation and complacency with past success. This is challenging in academics because it’s very easy to become an expert and become comfortable with what you know. But I’ve always wanted to expand my understanding and apply it in different areas as a testing lab of that understanding. The only way to grow is to put yourself in uncomfortable situations.

RR: Were there any courses that stuck with you, whether it was in economics, math, or some other field, that pointed you in the direction you wanted to pursue?

There was a high school course in history. The challenge with history classes is that you read these incredibly thick books, but the key to understanding was finding the underlying thesis of the book. There was a systematic approach that my teacher had picked up when he was in grad school on using a small set of conversational questions to ask yourself while reading to extract the thesis. It was a very structured form of thinking that actually allowed you to reinvent someone else’s work as your own based on your appreciation of these questions.

This process was effectively reinterpreting someone’s scholarship, which amounted to a research mindset. You’re constantly consuming state of the art knowledge, but also synthesizing and transforming it into your own.

That ended up being very powerful, and whenever I got lost with seeing the forest from the trees on a topic, would often go back to this script.

RR: Are there any things you pursued outside of your work or academics that you find dear to your heart?

With economics, statistics, and econometrics, it becomes such an important part of your psyche because to pursue them you have to love the subjects and that can be obsessive - there is no end to knowledge! Having balance is very important, and especially now in this pandemic, people are seeing diminishing returns to loading their days with work.

One important release for me that has stuck since youth is sports; I love playing tennis. It’s one of the sports where you can be humbled quite easily yet feel as though you are trying out for the high school varsity team each time you step on the court. The other beauty is the spontaneity of it - you never quite know where the ball is coming or where you might hit it until a split second before the shot. It’s hard to be spontaneous in life as you get older but the tennis court is a really amazing blank canvas.

The other thing is music, which is very important to me. I used to play a lot of guitar when I was younger, but I picked it up again. Working those other parts of the brain is really important to keeping balance.

I find myself increasingly trying to impart my passion for knowledge and education to my children. It pains me when they form a negative relationship with school and feel like I can uniquely make it fun for them. I think as adults, we struggle with understanding the right allocation of time between directly imparting values and lessons to our kids and pursuing our passions and leading by example

RR: From your time at UChicago, what is it that led to Microsoft?

Coming out of the University of Chicago, I became an assistant professor and then a tenured professor at the University of Wisconsin, Madison. In that portion of my career, I developed this research program around the economics of demand. When you have a product, how much are people going to want to buy it? If you change the parameters of the product, for example the price or its features, how will demand change? It’s one thing to see it in an economics textbook, it's another thing to measure this theoretical construct in real data.

I became really obsessed with that question. Demand and supply are beautiful, but unless you can measure them in the data, they’re not real. How can you touch and feel these things with data and software as opposed to theoretical imagination? It's hard to do that.

Go to the grocery store, what do you see on the shelves: hundreds and thousands of products. It's not enough to measure demand for one thing, you have to think about the demand for hundreds of thousands of things that simultaneously depend on each other. It’s complicated. It turns out that in the economics profession, the empirical practice of demand estimation, although amongst the oldest topics in economics, was still relatively new as an applied discipline. The research profile I created that got me to tenure was centered around the issues of taking demand models and ideas from economics to real data so they can be measured and analyzed for making decisions.

At that point in time, the question became: what can you do with this perspective? You can publish papers, and that’s great. But, who do you want to influence with that thinking? There are two general areas you can influence.

You can influence governments, who need to regulate industries and need to understand what demand in those industries look like. I increasingly found that government clients, like the Department of Justice or the Federal Trade Commission, weren’t ready to consume too much data. The data had to be very light; there is only so much that the government can do, with high level statistics, maybe a few pictures at most. This felt much more like consulting than building intelligent data products with economics.

So then I thought - who is a potential customer for econometrics and causal inference of around supply and demand? The customer must be someone who has a lot of data, someone who has a lot of computational resources, and someone who has big, heavy, hard questions. It turns out that tech companies are those exact places. Google, Amazon and Microsoft are really the big three, I would say, in that domain. It turns out that Google and Amazon had already made some investments in developing the economics around their product.

What’s interesting about Google and Amazon is that their primary customer, as companies, are for the most part consumers. They are people, like me and you. I search on Google, I have a Google phone, I buy things on Amazon. Microsoft is a harder problem because although I might personally use a Windows OS, the much larger customer base for them are businesses or enterprises, so they are really a business-to-business entity.

Being a tech economist in a business-to-business environment, or an enterprise environment, was a completely, wide-open, brand-new space. I had an opportunity to develop economic thinking in this green field and so took it. Satya had just made Microsoft’s identity a cloud company once I got there, and they were just starting to try to figure out the business, the economics of it, and the data.

One thing that made data and economics potentially valuable to Microsoft cloud is that they started the journey in a rather different position than their windows business - they were not number one. If you are already number one in your industry and if you have never used economics in running your business, then you don’t need an economist to tell you what to do! You got there without needing an economist. But in this case Amazon (AWS) was the market leader and Amazon had made a very public high profile investment in using economics to run its business.

One thing challenging about being a cloud provider for a technology company is that it is expensive - you have to buy all this stuff. You have to buy data centers, and servers, and other infrastructure, so you don’t make 100% margins like you do being a pure software company. So the whole DNA, soul, and story of the company around the identity of being a cloud product had really not been formed. In that environment, it was fascinating to bring economic perspectives and tools to help shape the story.

However the first thing I wanted to touch to use economics to shape the story was data. What you quickly learn is that consuming data for economic modeling in a large complex corporations to apply even the simplest economic model is a challenging new problem. It’s not so straightforward - it's a what all companies are facing and part of what makes the cloud so exciting.

I think the biggest lesson for me was that the world, society as a whole, is just at its infancy in using data to make its fundamental daily decisions. To be a data-disciplined organization is still a very new undertaking and a very new proposition. We talk a lot about machine learning and AI, and I think it’s interesting, but the set of areas that can be influenced by data go well beyond building vision and speech processing.

How do you price a product? How do you incentivize people to buy a product? How much inventory for a product to hold? We have these very basic, mundane classic questions, but influencing them with data is still very new and still very exciting.

RR: You went from teaching, to Microsoft, and back to teaching. Obviously, those are very different industries. What would you say is the biggest non-obvious difference you noticed between your time in industry and in academia?

It’s a great question, a really important question, and I struggle with this duality, to some extent. What’s exciting about industry is that there’s always this “reach for the stars”, “go for broke”, “zero to infinity in a second” kind of ethic, especially in the tech space. You can white sheet a bold, brilliant idea, and in a couple of weeks it's real, it's a thing. People are using it; it has oxygen.

Whereas, in academics there’s no quick win. You have to develop ideas slowly and systematically and bring the scientific community along with you. If you are going to be an academic, you have to be obsessed about your problem and the process of studying your problem, because you have to be willing to endure, struggle in obscurity, to some extent.

And that’s a tough divide, and raises a key challenge as to how you are supposed to apply scientific thinking in the setting of corporations and organizations. One thing you encounter in industry is a need for a quick win, outcomes, action, and impact on a much shorter time scale than is possible for a scientific view on a topic. This is also where data has a very powerful role to play because you can develop small experiments around data and bring a rapid form of the scientific method to bear on business problems.

In some ways, industries and corporations have filled that void by throwing more people at problems. If you have a hundred people working on something, there must be some truth that emerges is the business bias. But with headcount comes entropy and complexity - the question is whether we are creating work for its own sake or adding value.

But that’s really where, I think, data changes the kind of relationship you can have with the business problem you are trying to tackle, because it really opens up the depth of what you are trying to study. In some ways the perfect world is a little bit of both, because you do want deep thinking, and academics is the only “industry” that supports multi-year investments in problems.

Corporations would be hesitant: “Give me X investment, and in five or six years I’ll give you an outcome, I’ll give you an answer”. Very few people are willing to make those investments, and academics is really the only thing we have in society for them. Universities are very important in that regard.

What can work better in the modern era is the cooperation between universities and businesses. The other thing, going back to machine learning and AI and economics, which is where my own interests lie, is if you are going to have an impact in those fields, what do you need? You need data, but you need, not just any data, you need lots of it. You need big data, gargantuan, petabytes of daily data. Universities don’t sit on that data.

Who sits on that data?

Technology companies and born-in-the-cloud organizations. In a sense, they have the cutting edge problems and questions, and academics has the thinkers and designers. What we really need to be doing is putting those two together more. That’s another area that interests me as an entrepreneur. How do you build those bridges between those realms?

RR: What roles do AI, machine learning, and cloud computing play right now? How do you see them advancing in the future?

First of all, these technologies are incredible. You have to be awestruck by what they have achieved. Because ultimately, you can argue the science or the techniques are relatively well understood, but it’s about putting it into motion and practice in an end-to-end way. That’s where all the rubber meets the road. Can we have models that work on their own, that make decisions without any human intervention to do tasks like image recognition?

There were generations of scientists that got closer to the core milestones on this problem, but once the sword in the stone got pulled, it left the science domain and scaled in the business domain. This brought a combination of a little bit better technique, a little bit better data, and a little bit better hygiene around the practical tools to make it functional. I think what happened for speech and image recognition can happen for human and organizational tasks more generally in a way that blends human and machine interactions.

Even if you’re not doing image recognition, or speech recognition, or even high-dimensional hedge-fund-style trading, one of the powerful effects of the AI/ML revolution has been making organizations invested in their own data. In some sense, even if you don’t know why it’s useful, you know you have to collect it with the migration of business to cloud. Suddenly, everyone is realizing they have to be systematic about the data intelligence on which they are going to live. More than anything, it’s created an ethic to use that data in actual, real, practical, business context.

So there’s been the ironic effect of the AI/ML revolution - Even if you’re not using AI/ML tools, it has made everyone look at data now to use even pre AI/ML data technique (e.g., linear regression!) Some of the most influential data science can be looking at an average of a collection of numbers - the measure of central tendency in a variable. Powerful, right? But, what’s the hardest thing? It’s to understand the right variable to measure. How do you get businesses to collect information on the right variable? That is a design problem and takes foresight when building a data system.

In some ways, part of the collateral effect of AI/ML has been getting businesses willing to have that conversation. “Okay, I’m not measuring the right things.” In some ways, measuring what matters can be far more influential than what you do with those measurements in realizing the first order value of data. I can say a sample mean is a very simple form of AI. You take an average, I train my sample, there’s something that comes out of it. You go from an average, to a regression, to a neural network; that’s all part of the same spectrum.

The big journey is starting to develop simple models around your data. And that’s, more than anything, the industry movement that the AI/ML revolution has really pulled out and is revolutionizing industries and economies - applying very classical thinking with modern tools.

The other thing is that it’s created all this democratizing of these tools. Nowadays, on my own laptop, the kind of tooling I have available through R, Python, and Tensorflow, and just being able to play... These frameworks are just incredible, and in some ways, they frameworks themselves are the main event as a way to unpack problems as opposed to the engineering creations that come out the other end. But, they should be part of the citizen data scientist tool kit for everyday life and work. Even in academics, I wouldn’t be approaching these tools in the way that I do today unless I saw their immense, direct value in an enterprise, data science environment. I think it’s been very profound, but people are still catching up to it. There’s a lot of room to grow still.


RR: What role do you see cloud computing taking in the future?

We need to first define cloud computing at its economic level; what is cloud? It is fundamentally selling and providing a capital product at a variable cost to the end consumer. The cloud provider is investing in a capital product, a data center or server, and selling it to you so that you don't need to do that capital expenditure yourself. This is not unique to cloud. It’s no different than electricity, for example.

At first, people needed to generate their own electricity in their backyard. At a certain point, it became more economical for someone to provide that service to people at a variable cost so that they don't need to incur the heavy capital expenditure and they can consume electricity at a variable cost.

This raises the question of what kind of capital services can you replace with cloud. What kind of expenditures are people doing in their daily lives that you can replace with something as a service? I think what we found in the cloud economy is that that range of things can actually be quite broad; the core cloud model is to sell infrastructure as a service.

Instead of you buying a server and putting it in your own company’s data center, the cloud company will build that service and stream that service to you online. This just reproduces your current functionality in the cloud, presumably with a reason to do so because there are efficiencies to this model. At some point in time it's cheaper for large technology companies to grow your own IT investments instead of you taking on the capital costs on your own.

You can start to expand this business model out: think about any expensive investment in technology with uncertain returns. Anything including: ‘I want to experiment with this crazy AI technology,’or ‘I want to scale up an exotic computation to see what it yields,’ or ‘I want to hire a team of economists and operations researchers to help design prices for my product.’ All of these are capital projects (physical and human capital) with uncertain returns. So, what happens if these projects with uncertain returns become services on the cloud.

Effectively, the cloud will create some service or function for you to experiment with in your own enterprise and if it fails you can just walk away from it. You don't need to pay for it any more. But if it succeeds, you can scale up and this really dramatically changes and expands the set of risk reward trade off to almost every experimental uncertain business investment you can imagine.

This is where I really see the vast potential of cloud computing. It's not so much about what you are provisioning - whether it is a VM or some other infrastructure, but it is how you are provisioning it. Where it can go in the future is business model solutions, algorithms, and ideas all being transmitted and exchanged in a medium that allows for rapid adoption to the everyday business and person. That's really one of its core attributes and strengths. In many ways, we see this today with all these platforms like Uber and Airbnb.

Uber is effectively transportation as a service. Instead of buying a car I can stream a car service through this ride sharing economy. The economic model is thus old but what is quite new is the sheer economic scale it's operating on. I am not sure there is any economic good or service that is not subject to impact from Cloud. That's really immense.

RR: Regarding that can you actually describe your favorite project as chief at Microsoft?

I have worked on many interesting projects, but a favorite of mine was using a simple economic principle to make a big point. Going back to the cloud problem described above - it's one thing that customers can get elastic cloud computing on demand that they can scale up as much as they want, but for that to happen there actually needs to be capacity on the other end. So who is buying all this capacity? The cloud providers. But how much should they buy? Elastic cloud creates a perception of infinite capacity, but in reality there is finite cloud capacity in the world.

The answer to a cloud provider’s capacity problem can’t just be to buy more capacity to avoid any risk of a capacity outage because that would mean infinite capacity! Instead it needs to answer: How much, where, when, how, and why to maximize business and customer value?

You need to understand all of these dimensions before you can make a successful capacity decision. Then I started to realize why a company might make sub-optimal capacity decisions. It was an application of Conway’s Law, a very interesting business principle that you should all know. It says in a nutshell, that the product you ship always has some underlying reflection of the organizational structure from which it is born.

The organizational structure always impacts your product, which is why organizations really do matter. I happened to be in a position where capacity was an organization within itself; all cloud capacity was kept under one organization.

To go back to Conway’s Law, if you are in an organization with a cost, and the capacity to Microsoft is a cost, then organizations will be incentivized to keep costs low. And what's the best way to do that? Well it's to not buy capacity. In some ways, one of the things that I discovered was that problematic capacity decisions can arise if there are not organizational incentives to enable the profit maximizing solution.

Diving into this problem really helped me understand that organizational structure, concepts, and communication fundamentally matter as much as the science of the variable you are trying to optimize. If you call capacity a cost, you aren’t going to buy more of it. Let me give you a different economic interpretation of capacity.

What if I told you that it was actually an investment? By buying capacity, you are investing in your eventual customer success. This is now a very different business variable than if it was considered a cost. Then, in order to know how much to buy you need to know how to value the asset. Until you know how to measure this investment value, you will never know how to solve the problem.

To unpack that even more, what's hard about being a cloud provider is that provisioning capacity can take a long time. These are large scale, complex capital projects that can take over many years to evolve. Any cloud provider needs to think many years ahead in terms of its capital investments in order to be successful and have the right amount of investment on their books to support their product.

Now think about that, it's like trying to predict how big social media would become on like day three of Facebook. Could anyone have guessed the size? That is the challenge with cloud - the business models in the future have not even been written. There is a ton of uncertainty into how much risk, pricing, and competitive pressure will exist in the marketplace for this product.

My favorite project is using economics to make complex capacity investments in the face of massive risk/uncertainty if you had an irreversible investment with a large amount of risk and uncertainty, you cannot just sell the asset if things don’t work out - a portion of the investment is a sunk cost. There's something that's always going to be lost which means it's an irreversible investment.

The heart of those business and economic problems is that there are a lot of options to value and manage that uncertainty, just like Black-Scholes is a financial model that businesses use when facing problems of investment under risk and uncertainty. That means supply chain and capacity decisions can have a Black-Scholes mechanism in their investment, because what they really need to do is manage risk. For example a great risk management tool is land - it does not lose value and has a lot of option value to grow cloud if the demand is there. But if you don’t hold the land, then you can't suddenly respond to demand if it arises out of the blue.

Working on the option value of capacity investments including land and energy was quite fun. It wasn't doing some crazy high-tech machine learning thing, but using clean economic principles with a smattering of data to guide an investment decision. Using option value which is a beautiful idea to a practical business problem was immense fun

RR: To wrap up, we just want to touch a little bit on your research. You talked a lot about all the interesting things that problems you dealt with at Microsoft. How did you bring this to your research that you're currently doing in industrial organization and econometrics? What is your favorite question that you are dealing with there?

So you can totally see that I took my economics bias and applied it to Microsoft, and I have done the same thing where I am taking my Microsoft bias and applying it to economics. This is because I found out firsthand that to make those decisions, we needed to get into data. We need to use this data to influence many stakeholders, which meant not only researching this data but turning it into a professional product. You realize in the course of writing papers, you need to make many small microdecisions.

Which variable should I use?

Which cuts of data should I consider? How should I write down the function for this? When I'm writing a paper, it's fine for me to manage all of these decisions in my hand because I retain that right as the researcher to make those decisions. If you are doing this at scale in a business enterprise, you need higher order principles that allow you to manage the data complexity so that you're focused on the strategic concept of what you want the data to demonstrate rather than the low level details of how the data gets cut, transformed and applied to the question. This is where I believe there is a lot of value in Machine Learning, AI, and automation to how research is conducted. This is where the human and machine interaction is potentially powerful.

We have this problem in business and social science that is sometimes called credibility revolution. It is very hard to reproduce scientific findings. Why is that? Because all these things are done by hand. All these little transformations got done by hand and there is no versioning or traceability. It was all done by humans putting together a paper. When I first came to Microsoft I used to always talk about my model and how my model of economics would work. Very few people could really appreciate my model but if you change the angle, I was really just building a product orientation. I wanted my product to work and to be functional.

If you go back to how we write papers, when we write down models, we want these models to be effective and functional. This means that they can be reproducible and ported into different environments so that they can grow and be scalable and robust. I believe ML tools can be used to productize research thinking more readily to automate the things that don’t need to be done by hand. If you go back to something as basic as building a demand function, there are many elements of that question that I as the analyst should not necessarily do by hand. I should let ML and AI do a lot of things for me while guiding those tools to make sure they are not just solving a computer vision problem but actually solving the economic problem of interest.

This melding of AI machinery into how economic structures are researched in data has been a very exciting area for me to begin to unpack. Just take something as basic as running a regression of Y on X. What are the X variables? Which ones and what are they? Up until now the smart economist got to pick the X variable, but is that necessarily the right thing to do. Maybe we should be letting computers tell us what to do. These are some of the interesting things I am trying to study.

Written by Karina Thanawala, Thomas Braun, Gihyen Eom & Edited by Alexander Fleiss