You Say "AI" Like It Means Something

You keep using that word. I do not think it means what you think it means.

The term "artificial intelligence", or "AI" is ubiquitous today. AI is used to recommend movies, approve loans, referee sporting events, make health care decisions, perform military operations, translate languages, and drive cars. This is a broad range of applications - how can the same technology be doing all of them? The reason is that the term “AI,” as it's commonly used today, is broad and poorly defined. In fact, it's so broad that I consider the term almost meaningless. But there are real and impactful technologies going on behind all the buzz, and it's important to understand how they work. In this post, I'm going to talk about what AI is and suggest a heuristic to help people think more clearly about it.

Let's start by asking: What is AI? A lot of definitions have been provided over the years. My favorite is:

AI is “the art of creating machines that perform functions that require intelligence when performed by people.” - Ray Kurzweil, The Age of Intelligent Machines, 1990

As good as any abstract definitions might be, I think the best way to understand comes from thinking about specific examples. There are many forms of AI, and their capabilities range from very simple to very complex. At the simple end are machines that play simple games like tic-tac-toe. AI doesn't have to be associated with machines or computers; even simple mathematical models - which were once simply called "statistics" - are now often considered "AI".

It's important to understand that some AI applications can be quite simple. For example, let’s say you want to predict how much people are willing to pay for lemonade based on the temperature outside. To do this, you need to create a model. For ease of use, you might want to assume that there's a straight-line (or "linear") relationship between the heat and someone's willingness to pay. Now you gather some data - you ask people how much they would be willing to pay and take careful note of the temperature at the time. To turn this into a model, you need to find the line that best fits the data. This is known as linear regression. The details aren't important, but it's a simple algorithm that lets you keep making guesses about the best line and get closer and closer over time. In the image below, you can see an example of linear regression being used to find the best fit line. There are many examples of AI used in science and industry that are no more sophisticated than this.

Linear regression - the process of learning the equation from a line - is an example of AI. (Source)

At the more complex end are robots that can run and jump, computers that do nuanced language translation, and self-driving cars. These are some of the most advanced AI applications in the world today. At the far end are futuristic, Terminator-like robots that can think, talk, and interact like they are human. This last group is known as artificial general intelligence (AGI). There is nothing like AGI happening today and, in my opinion, there is no convincing evidence that it will happen in the near future.

Given the broad range of AI, how can we focus the conversation? Well, despite all the various subfields of AI, the dominant one in the last decade has been machine learning. Machine learning is a mathematical technique of creating algorithms that can learn directly from data. It is a mix of different disciplines, including calculus, linear algebra, statistics, and computer programming. These disciplines are used in combination to rapidly guess, check, and update a model to better fit data. This new field, together with the decreased costs of computation and data storage, has resulted in the explosion of AI applications.

Being able to learn from data is incredibly important. Many tasks, especially in computer vision and language understanding, are so difficult that no human can directly program a computer to do it. This is where machine learning comes in. Instead of trying to figure out exactly how to solve a task, you give a machine learning algorithm lots of examples and say "figure it out for me." The incredible thing about machine learning is that this actually works.

Machine learning is a subfield of artificial intelligence. I've also added neural networks, as those are an example of a machine learning model and an increasingly common term. That's just a subset of machine learning that is generally best for complex tasks like reading and vision. More divisions within neural networks could be included, but that's for another time.

Focusing on machine learning helps us narrow the scope because it is well-defined. Even though this still presents a broad range of capabilities - from linear regression to self-driving car technology - it is much easier to talk about.

Note that not everything happening in artificial intelligence uses machine learning. For example, Deep Blue and the Boston Dynamics robots are primarily not machine learning systems, but they are still AI.

So, given all of that, what's the best way to think about AI? How can we have better and more methodical conversations about it? I propose a heuristic for thinking about AI: When you see the term "AI", substitute it with the phrase, "A combination of linear algebra and statistics to learn from data." That might be a surprising substitution, but I think it's a good one.

This heuristic provides a system for judging whether the term "AI" is being used in a meaningful way. If you replace the word AI with "a combination of linear algebra and statistics to learn from data" and it still makes sense, it's possibly a good sentence. If it can't, it's probably not a very meaningful usage.

People ask questions like "Is the use of AI by police a good thing?" Hopefully, you realize how meaningless and unanswerable this question is. Is the usage of linear algebra and statistics to learn rules from data good for policing? Well, it can be; it depends on how it was done.

Learning from data certainly could be good for policing. But what's the data? What are they learning from it? How are the results being used? How are they measuring success? Are they correcting for any changes in policing that the model caused? Think of it like a spreadsheet - are spreadsheets good for policing? Maybe! It really depends on what you put in those spreadsheets. But without these details, it's a meaningless question.

Without the shroud of mystery around AI, these questions look not only meaningless but actually quite silly. It's as if someone has asked, "Is it good to use a hammer to make a house?" Well, using a hammer in constructing a house makes a lot of sense, but can I have more details on what they did with it?

By now I hope you're starting to see that AI is a lot less magical and sci-fi-like than the media makes it out to be - it's just good ole math and data. Like math, it's just a tool that you can use. And, like every tool, AI can be used productively, destructively, or anywhere in between.