Machine Learning vs AI: What's the Difference?

With AI everywhere these days, there’s one key question that keeps popping up: what exactly is Artificial Intelligence (AI)?

If you ask this question at an AI conference, you’ll get as many definitions as attendees. But let’s start with a definition from IBM that encapsulates the core ideas:

“Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity, and autonomy.”

This is a solid start, but the conversation starts getting tangled when we bring up Machine Learning (ML), a close cousin of AI. For now, I’ll save myself the effort of finding more sources (it’s why engineers invent things, i.e. to remove all effort) and stick with IBM's definition:

"Machine learning involves creating models by training an algorithm to make predictions or decisions based on data. It encompasses a broad range of techniques that enable computers to learn from and make inferences based on data without being explicitly programmed for specific tasks."

A useful analogy here is to think of AI and ML as different layers in a network cake. At the base, ML handles data processing and classification; (think of recognising whether an image is of a cat or a dog). AI, perched on top, manages more complex, higher-level tasks that often require multiple ML algorithms working in concert.

A simpler definition I have adopted to separate Machine Learning from Artificial Intelligence is:

If a computational system has multiple defined goals, then it is AI.

Else, it is Machine Learning.

Why do I focus on goals? Because goals drive complexity. In this context, a goal is simply a user-specified task that a computer needs to perform. AI becomes necessary when a system must manage multiple goals. For example, take a robotic vacuum cleaner. It relies on lower-level ML algorithms to process sensor data, detect battery state, and classify the dirt level on the floor. These are all straightforward, single-task functions. However, the real complexity—and the AI element—comes into play when the vacuum robot must balance multiple goals: clean the floor effectively, but don’t let the battery drop too low or it won’t make it back to the charging station.

The set of goals doesn’t always need to be in conflict; they may simply be sequential, such as: achieve goal A, then goal B, and so on. But even when the tasks are sequential, it's still not a simple data classification problem. The machine must plan, reason, and make inferences to successfully accomplish the full set of tasks.

Take ChatGPT as another example. Here, a complex neural network and vector inferencing system processes user prompts to translate them into multiple AI-driven goals. It’s not just regurgitating text but instead weighing up different objectives based on the user's intent.

Agents Everywhere

This idea of multiple, goal-oriented systems is where agent-based AI comes into the picture, where the notion of an agent is inherently goal-oriented.

I remember working on Software Agents in the late 1990s after completing a PhD in Robotics. At the time, the hardware just couldn’t keep up with the complexity of the software.

But now, with breakthroughs like Transformers and LLMs, the field is coming back to life. With this new Promethean power, the potential of software agents can be fully realised, hence expect it to become the dominant software paradigm in the near future.

Summary

In short, AI is any computational system that operates with a set of task-driven goals.

As the field of AI continues to evolve, we’ll see it implemented across a range of physical substrates—from biological to quantum systems (prototypes already exist…). Yet one thing will remain the same: the core concept of having and managing goals.

Unlike the England men’s football team.