INFORMS Does AI

When I finished my Ph.D. in industrial engineering and management science in 1996, I remember debating what to call my new profession – operations research (O.R.) or management science? As I went into industry, I quickly realized I wouldn’t have to spend much time thinking about the right term for my profession – the customers and the market could not have cared less.

However, around 2006 the term “analytics” seemed to come out of nowhere and catch the attention of everyone in the business community. I was skeptical and resisted when the term started to catch on. I thought it was just another meaningless buzzword. But the buzzword wouldn’t go away. I had to pay attention because our customers cared about the term. They wanted to make sure they weren’t missing out on the analytics movement. I had a front row seat, watching software vendors (which I was part of) and consulting firms do a massive rebranding to show how we were all “doing analytics.”

Like others, I now give Tom Davenport credit for making the term popular with his great Harvard Business Review article and book, both called “Competing on Analytics.” Davenport’s work made CEOs and leaders realize that they were going to be left behind if they weren’t taking advantage of the data they had available to them.

Early on, I had the feeling that reporting, visualizations and business intelligence (BI) vendors were getting all the attention from the excitement around analytics. In other words, it seemed like a lot of companies felt like they had tackled analytics if they were doing a good job with reporting. However, INFORMS did a great job of elevating other parts of analytics by defining the three different types: descriptive, predictive and prescriptive. This allowed INFORMS to stake a claim as the all-encompassing analytics profession.

The Recent Rise of Artificial Intelligence

As we all know, this space moves fast. In 2012 there were some big breakthroughs in image recognition. Improvements in deep neural networks, extra computing power with GPUs and more data allowed machines to draw equal to people in their abilities to identify images. Since deep neural networks work much the way we think the brain works, this became branded as artificial intelligence (AI). This was actually one of many rebirths of the term AI, first coined in 1956. Other rebirths quickly flamed out as the hype faded into the realization that the efforts weren’t very close to real intelligence at that time.

This new movement quickly caught on because, in part, big tech firms like Facebook and Google realized they could add a lot of value with image recognition, translation and other similar applications. And, within a few years, the self-driving car seemed just around the corner.

As a result of those 2012 breakthroughs and the subsequent successes in specific use cases (image and video recognition, natural language processing, translation, voice assistants and similar areas) and hope for other areas (like self-driving cars), we were led to what I think of as the first definition of AI: the ability of algorithms (or machines) to think and reason like a human. That is, if a machine could recognize an image, be an assistant you could talk to or drive a car, isn’t this at the same caliber as human intelligence? The media, popular culture and many in the research community took this to mean that we were right on the brink of machines having human or even superhuman intelligence.

Although the hype has tempered a bit in 2020, it is this definition that allowed people like Ray Kurzweil to claim that the singularity [1] (when machines will gain intelligence that is unstoppable) is near. In response to the fear of the singularity or other bad outcomes, top researchers formed groups like OpenAI [2] to advance the research, keep it open and ensure its safety. For clarity, it might be better to apply this AI definition to the term that researchers in this field use: artificial general intelligence (AGI). This is still a very active area of research with many new breakthroughs – research that has even reenergized questions such as “what is consciousness?” and “how will we know if a machine has it?”

It is easy to get swept up in the hype around the media announcements regarding AI. To counter the hype, Melanie Mitchell’s book “Artificial Intelligence” [3] and Rodney Brooke’s updates on his AI predictions [4] are helpful. Neither of them think we are very close to AGI.

AI vs. AGI vs. Analytics

To me, AGI is a valid line of research working on important questions and making beneficial breakthroughs. However, I believe that the definition of AI (that of AGI) is not the end of the discussion. There seems to be another important definition out there.

In the years leading up to 2016, the term “analytics” was hitting its stride and even the use of the term “machine learning” was seen as somewhat normal. In academia, new analytics programs were springing up everywhere. I co-founded a company in 2013, and we put “analytics” right in the name.

Sometime around 2016-2017, the term AI escaped into the wild. That is, it was no longer confined to the area of general intelligence. Around this time, you could listen to talks from Andrew Ng (one of the key people at the heart of the 2012 breakthroughs) about AI [5]. Ng mentioned the expected topics like image recognition, but also mentioned using AI for logistics. By 2018, many companies were noting their AI efforts in their quarterly reports to investors. Also in 2018, The Economist featured a cover story on how AI could add trillions of dollars of new value to the world in areas such as marketing, supply chain and risk management [6].

During this time, most software vendors and consultants were rebranding themselves to say that they were an AI company. In the summer of 2018, at our company meeting, we had a spirited debate about whether we should rebrand ourselves away from analytics and toward AI. Although some people argued that AI should only mean AGI, we decided to go with what the market was saying and call what we were doing AI.

We found that people were using the term AI to mean more than just general intelligence, which led us to realize that there is a second definition of AI: the use of data, computing power and sophisticated algorithms to solve specific problems. These algorithms were being used to help detect new patterns (e.g., fraud), read documents (e.g., invoices or receipts), make interesting new predictions (e.g., when a shipment will arrive, when a worker will quit or what demand will be next week), or make a complicated decision (e.g., the myriad of optimization applications). In this second definition, there is no assumption of trying to solve for general intelligence. However, there is an assumption that we are using sophisticated algorithms to make a prediction or decision, or do something that would have seemed like a human task (e.g., read a document). This definition also implies that AI is an umbrella term for many different algorithms: deep learning, reinforcement learning, optimization, machine learning, among others. When thought of this way, AI fits nicely with the INFORMS profession.

Many books, articles and conference talks confirmed this second definition of AI – implying that AI is an umbrella term for solving practical problems with data, computing power and sophisticated algorithms. Two books that implicitly back up this definition include “Prediction Machines: The Simple Economics of Artificial Intelligence” [7] and “Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World” [8].

In “Prediction Machines,” the authors argue that many business problems can be reframed as a prediction problem, and since the cost of running prediction algorithms (what they call AI) is dropping fast, more companies should rethink their business models and processes. The authors of “Competing in the Age of AI” stress that you need more than a few applications; you need to have a full AI mindset to change your business. At the risk of oversimplifying, they say the way to do this is to create three things: 1) a data layer that is open to all groups in an organization, 2) an algorithm layer to make the prediction or recommendation, and 3) a software layer to deploy the solution to capture business value. I think of this as a blueprint for implementing the second definition of AI.

Same Thing, New Buzzword?

Taken cynically, it might seem like we just replaced one buzzword for another. Instead of saying analytics, we say AI. There might be some truth to this, but I prefer the term AI over analytics for three main reasons.

1. Despite years of arguing that analytics includes predictive and prescriptive, descriptive analytics still seems to dominate. I would be happy to let reporting, dashboarding and BI systems have the term analytics while other groups lay claim to AI.
2. AI implies you are doing much more than reporting. It implies you are using sophisticated algorithms in an interesting way. I also think it suggests a mindset change. To build a report can mean to just give people information. To lead an AI effort means you want to automate a process, and make a complicated prediction or complex decision.
3. At this point AI leads to less confusion because the term is already popular. (I suspect it caught on because it sounded cool.)

What Does this Mean for the INFORMS Community?

My first recommendation is that the INFORMS community should take the lead in clarifying the two definitions of AI – just like the Institute did with defining the three types of analytics. This would promote clearer thinking and communication around AI.

Fortunately, the first definition, artificial general intelligence, has a good name. Unfortunately, I don’t know what to call the second definition. AI might just be confused with AGI. I’ve seen several names for the second definition, including artificial narrow intelligence (ANI), practical AI and enterprise AI.

My second recommendation is the strongest: INFORMS should embrace the term AI. As others have suggested, people from our community are and can play a big role in AGI. More importantly, no one has a better claim to the second definition of AI. INFORMS and its members are in the best position to understand the algorithms, understand how to pull the data into the algorithms, and how to turn the output into results that people and organizations can use. And, quite often, I’ve noticed that when businesses and organizations create AI groups, they tap into people traditionally trained in OR/MS to lead the groups.

Of course, there are many edge cases to consider: How does the term data science fit into all of this? Will AI as a term be replaced by something else? If you are running one small linear program, are you doing AI? What about the areas of INFORMS that don’t use algorithms or help with more strategic decisions?

We can eventually answer these questions, but in the meantime, INFORMS shouldn’t miss the opportunity to capture the excitement around AI to its benefit. We could be tightly tied to the AGI movement, and we could be the key academic and professional organization for ANI or practical AI.

References

1. https://en.wikipedia.org/wiki/The_Singularity_Is_Near
2. https://openai.com/
3. https://melaniemitchell.me/aibook/
4. https://rodneybrooks.com/predictions-scorecard-2021-january-01/
5. https://www.youtube.com/watch?v=21EiKfQYZXc
6. The Economist, 2018, “The workplace of the future,” March 28, https://www.economist.com/leaders/2018/03/28/the-workplace-of-the-future.
7. https://www.predictionmachines.ai/
8. Iansiti, M. and Lakhani K. R., 2020, “Competing in the Age of AI: Strategy and Leadership When Algorithms and Networks Run the World,” Harvard Business Review Press: Boston.