Goals of AI: An Overview

The Cycle of Intelligence: The AI Agent

A helpful way to think about AI is through the concept of an AI agent. An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators. You, as a human, are a biological agent. A self-driving car is an artificial agent.

These agents often operate in a continuous loop, which provides a great framework for understanding the four main goals of AI.

As you can see, these four components are not just a list; they are interconnected parts of a dynamic process. An agent perceives the world, reasons about what it perceives, takes an action, and then learns from the outcome of that action, which in turn influences future perception and reasoning.

Let's start with a foundational reading that introduces each of these components.

Now that you have a general overview, let's explore each of these four goals in more detail, using different examples to bring them to life.

1. Perception: Sensing the World

Perception is the agent's ability to gather and interpret information from its environment. It's the AI's equivalent of our human senses. Without a way to take in data about the world, an agent is blind and deaf; it has no information to act upon.

• For a human: This involves sight, sound, touch, smell, and taste.
• For an AI agent: This involves:
  • Sensors: Cameras, microphones, LiDAR (Light Detection and Ranging), and temperature sensors in physical robots or self-driving cars.
  • Digital Data: Text from a user's message, data from a financial market API, or pixels in an image file for software agents.

Let's watch a short clip that illustrates the most basic form of perception and action.

The thermostat is a simple but perfect example. Its "perception" is the measurement of the room's temperature via a sensor. This raw data is the necessary first step before any decision can be made. For a more complex agent like a self-driving car, perception involves fusing data from dozens of sensors to build a complete picture of its surroundings.

Given your interest in neuroscience, you might find it interesting to see how this engineering concept parallels the biological brain.

2. Reasoning: Thinking and Planning

Once an agent perceives the world, it needs to reason about that information to make a decision. Reasoning is the "brain" of the operation, where raw data is turned into a plan or a conclusion.

Reasoning in AI can take many forms, from simple rule-following to complex strategic planning.

• Simple Reasoning: If the perceived temperature is below 19°C, then turn on the heat. This is a rule-based approach.
• Complex Reasoning: Figuring out the sequence of moves required to solve a Rubik's cube, or planning a route to a destination while avoiding traffic.

A classic way to visualize AI reasoning is through goal trees. An agent has a main goal, which it breaks down into smaller and smaller sub-goals until it has a series of simple, executable steps.

Let's watch a fascinating demonstration of this in action.

As you saw, by keeping a record of this goal tree, the program can answer "why" questions by moving up the tree to a higher-level goal, and "how" questions by moving down the tree to the specific sub-goals it executed. This is a fundamental form of AI reasoning: structured problem-solving.

More advanced agents use more sophisticated reasoning. Let's briefly see how this can evolve.

These examples show that reasoning isn't just one thing; it's a spectrum of capabilities, from reacting to stimuli to simulating future outcomes to achieve a goal.

3. Action: Affecting the World

Perception and reasoning are internal processes. For an agent to be useful, it must be able to take action to affect its environment. An action is the output of the reasoning process.

• For a human: This could be speaking, walking, or picking up an object.
• For an AI agent: This is accomplished through actuators:
  • Physical: Electric motors in a robot's arm, the steering and braking systems in a self-driving car.
  • Digital: Sending a message in a chat application, executing a trade on the stock market, or displaying a recommended movie on a streaming service.

In the block-stacking video, the 'actions' were grasping, moving, and ungrasping blocks. In the thermostat example, the 'action' is turning the heat on or off. Action is the step that makes the AI an active participant in its environment, rather than a passive observer.

4. Learning: Getting Smarter Over Time

This brings us back to the core concept from our last lesson. Learning is the ability to improve performance through experience. It's what makes an AI system truly "intelligent" rather than just a complex, static program. Learning allows an agent to adapt to new situations and become more effective at achieving its goals.

There are three main ways AI agents learn, which we will cover in much more detail later in the course:

• Supervised Learning: Learning from labeled examples (like we saw with the even/odd numbers).
• Unsupervised Learning: Finding hidden patterns in unlabeled data on its own.
• Reinforcement Learning: Learning through trial and error, guided by a system of rewards and punishments. This is very similar to how many animals (including humans) learn new behaviors.

Let's watch a final clip that explains the components of a learning agent.

The AI chess bot is a fantastic example. It plays a game (Perception, Reasoning, Action), the critic sees that it lost, and the learning element adjusts the bot's internal strategy so it's less likely to make the same mistake in the future. After thousands of games, its performance drastically improves.

The Big Picture: Moravec's Paradox

It's tempting to think that the goals of AI are to perfectly replicate human intelligence. However, one of the most interesting ideas in AI is that artificial and biological intelligence have very different strengths and weaknesses. This is known as Moravec's Paradox.

The paradox, summarized by AI researcher Hans Moravec, states: "It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."

In other words:

• Tasks that are hard for humans (e.g., complex logical reasoning, arithmetic, memorizing huge amounts of data) are often easy for AI.
• Tasks that are easy for humans (e.g., perception like recognizing a face, or action like walking and picking up a cup) are often extremely difficult for AI.

Our brains have evolved over millions of years to be incredibly efficient at perception and movement. The "higher" cognitive functions are a more recent addition. For computers, the opposite is true. Calculation is their native language, while understanding the messy, unpredictable physical world is a huge challenge. This is a key insight for understanding why AI excels in certain areas and struggles in others.

Conclusion

In this lesson, we deconstructed the complex idea of "intelligence" into four concrete goals for AI.

Here are the key takeaways:

• AI systems can be conceptualized as agents that operate within an environment.
• The four primary goals of AI are Perception, Reasoning, Action, and Learning.
• Perception is about sensing the environment (e.g., through cameras or data feeds).
• Reasoning is about processing that information to make a decision (e.g., through rule-following or planning).
• Action is about affecting the environment based on that decision (e.g., moving a robot arm or sending a message).
• Learning is about improving performance over time based on feedback from actions.
• Moravec's Paradox highlights that AI's strengths (e.g., in reasoning and calculation) are often humanity's weaknesses, and vice-versa (e.g., in perception and action).

These four goals are the building blocks of virtually every AI system you will encounter, from a simple smart thermostat to the most advanced language models.

Preview of the next lesson: We now know what AI is and the goals it strives for. But not all AI is created equal. In our next lesson, we will discuss the different levels of AI that are often mentioned in media and research: narrow AI, general AI, and superintelligence. We'll define each one and understand where today's technology currently stands.