Artificial Intelligence (AI) is a topic that is currently generating much interest and gathering much attention. The term "AI" means different things to different people. Currently, AI seems to refer to at least four interrelated things.

For many, AI is an interdisciplinary area of study that straddles the interfaces that connect computer science, neuroscience, linguistics, engineering, mathematics, physics, philosophy, psychology, and other fields. Among other things, the field of AI aims to:

• investigate the underpinnings of human intelligence and cognition
• simulate human intelligence in machines, primarily computers
• engineer software and/or hardware systems that can "learn" from their experiences in the world around them, or from data provided to them
• build systems that can perform equally as well or even outperform humans on tasks that are considered hallmarks of intelligence

For some, AI is used to refer to one or more products of research and development in this area. We might say things like "My AI can write sonnets", "Your AI doesn't know what it's talking about", or "I wonder if that AI knows what I am thinking".

In this current era of interest in generative AI applications such as ChatGPT (more to follow), AI is also used specifically to refer to tools that can generate text, images, computer code, or other content based upon natural language prompts.

Finally, the term AI can refer to the use of tools from the fields of machine learning and deep learning, which — until a few years ago — were mostly referred to as "machine learning" and "deep learning", respectively. Given the current surge of interest in "AI", however, much of that work has been rebranded under this more exciting and marketable moniker.

Many scientists and educators have endeavored to describe what AI is, emphasizing different aspects of this large and multifaceted discipline, and of the entities produced by such work. The widely used textbook by Russell and Norvig, "Artificial Intelligence: A Modern Approach", highlights the role of intelligent agents, and defines AI as "the study of agents that receive percepts from the environment and perform actions." Let's break that definition down into its parts to get a better idea of some of the relevant aspects:

• An "agent" refers to some sort of entity that performs actions or produces outputs
• A "percept" is something that is perceived, or perhaps data that is received as input
• The "environment" is world around the agent, providing a context in which it acts

Russell and Norvig state further that "Each such agent implements a function that maps percept sequences to actions", and spend much of the rest of their book describing different ways to represent and implement those functions. Much of AI is about agents that take inputs from the world around them, and produce outputs, which might be in the form of digital decisions, physical actions, or some combination thereof.

But in order for an agent to exhibit some degree of "intelligence", those functions that map inputs to outputs, that produce actions based on percepts, need to somehow be useful, responsive, productive, and informed, such that the action produced is not indiscriminate and disconnected from the input or percept. An agent that always responds in the same manner regardless of input, or always responds in a random fashion regardless of the input, would not be construed as intelligent if it never made any meaningful use of the perceptual inputs that it received.

We will discuss some of this in more detail in the pages that follow, noting here that much of the "learning" that takes place in AI is associated with the process of developing the useful, productive, informed, and "intelligent" actions described above. Much of this will be described in the topic on Machine Learning and Deep Learning, where flexible computational representations of arbitrary mathematical functions can be effectively configured and parameterized through a process of training with data.

As noted, the field of AI has broad and diverse aims. Russell and Norvig identify four overarching classes of approaches to AI: acting humanly, thinking humanly, thinking rationally, and acting rationally. The first two associate AI with a desire for fidelity to human performance, whereas the latter two emphasize that a key goal of AI is finding solutions that are rational or optimal in some sense. Two classes focus on internal processes of thought, while the other two emphasize the external presentation of behavior and action. The famous Turing Test, proposed by Alan Turing, defines AI in the context of acting humanly, in that the test would be passed if a human — engaged in conversation with a computer — could not actually determine whether it was interacting with a computer or with another human. Russell and Norvig emphasize that the focus on acting rationally, or the rational agent approach, has come to dominate much of AI research and development, with an emphasis on the "the study and construction of agents that do the right thing" based on what objectives are provided to or learned by such an agent.

While much current excitement surrounds the use of deep learning and neural networks to implement various forms of generative AI, it is important to step back and appreciate that the field of AI consists of many different subfields and subproblems addressing the broader goal of achieving machine intelligence. Russell and Norvig, for example, structure their book around four main themes:

• Problem-solving
• Knowledge, reasoning, and planning
• Uncertain knowledge and reasoning
• Communicating, perceiving, and acting

Within these broad themes, there are rich and interesting questions and results that can seem far-removed from the current enthusiasm for chatbots and image generators. Stepping back and appreciating the connections between all these topics and approaches will ultimately help propel forward the quest for artificial intelligence.