AI (Artificial Intelligence)

Artificial intelligence (AI) is the field concerned with building computational systems that perform tasks involving perception, reasoning, learning, planning, or decision-making. It includes methods for representing knowledge and searching for solutions as well as methods that learn from data. Poole and Mackworth frame AI around computational agents acting in an environment: the system receives information, selects behavior, and is judged against the task and its constraints.

How rules, search, and learning contribute

A symbolic approach represents facts and relationships explicitly. For an illustrative delivery problem, the system may know that road A connects two locations, road B is closed, and a vehicle cannot use a low bridge. Search compares possible routes; a planner considers actions and their consequences. The resulting route can be computed without learning the road constraints from examples. Its quality still depends on whether the map, constraints, and travel-cost estimates match reality.

A learned component can estimate travel times from previous trips. Machine learning adjusts a model using data and an objective, such as minimizing prediction error on observed trips. During inference, that model predicts a new trip’s duration. Learning a predictor and selecting the route are distinct operations: the planner can use predicted durations as inputs while enforcing explicit constraints. Combining these approaches can be more useful than choosing one label for the entire system.

Deep learning is a family of machine-learning methods using neural networks with multiple learned transformations. For example, a vision model can transform camera pixels into representations useful for detecting an obstacle. The designer still chooses the training data, objective, architecture, and evaluation procedure. Learned representations reduce the need to hand-design every feature; they do not remove those design choices.

The relationship AI → machine learning → deep learning describes overlapping scope, not a performance ranking. A route search, a linear duration predictor, and a deep vision model solve different parts of this example. Comparing them as interchangeable contestants would miss their roles. Even on the same prediction task, performance depends on data, representation, training, and the error being measured.

Generative models and systems that act

Generative AI produces content such as text, images, or audio by using a learned model of patterns in data. A predictive model might assign a category to a customer message; a generative model might draft a reply. Many current generative methods use deep learning, but generating an answer is a task capability, whereas deep learning describes a family of methods. The terms are not synonyms.

An AI model is also different from an application that can act. A delivery assistant might use a language model to interpret a request, a routing service to calculate options, and a booking API to reserve a vehicle. The application determines which tools may run and what happens when an operation fails. A plausible sentence saying that a booking succeeded is not evidence that the booking system accepted it.

In Poole and Mackworth’s agent design space, planning horizon, representation, and computational limits are separate dimensions. That distinction helps describe systems more precisely than calling each one simply more intelligent. A system can plan well within a narrow, known environment and fail when its assumptions about the environment are violated.

What a comparison needs to establish

Start with the task. For the delivery example, useful criteria include whether a route obeys constraints, arrives before a deadline, can be computed in time, and handles changed road conditions. A more accurate travel-time predictor may still produce worse deliveries if the route objective ignores reliability or the application uses stale information. Evaluate the complete path from input to action as well as the individual model.

Use a simple baseline when it answers the same task, such as the existing routing procedure or a fixed travel-time estimate. Compare on representative cases that were not used to choose the model. Include unusual but important conditions, because a high average score can conceal failures on a specific road type or operating period. More parameters, more training data, or more computation do not by themselves establish a better result.

Computational cost also needs a comparable unit. A tree traversal, a matrix multiplication, and generating a multi-token answer involve different work. Parameter count alone does not determine latency, memory, energy, or price. Measure the actual input sizes and hardware, including preprocessing and any additional tool calls.

Explanation and reliability have limits

Explicit rules can make a decision easier to inspect, but a long rule set can still contain conflicts or omissions. A learned model can be examined using examples, feature effects, or a simpler approximation of its behavior. An approximation must be checked on cases beyond those used to fit it: explaining its own predictions well does not prove it explains the original model everywhere.

Predictive accuracy, explanation fidelity, and causal understanding answer different questions. A system can reproduce a pattern without knowing why it occurs, and a fluent generated explanation can be inconsistent with the calculation or external evidence. For consequential actions, connect the result to observed inputs and system state, define how uncertain or unsupported cases are handled, and verify the action’s outcome.

The AI label therefore describes a broad field and a set of capabilities. To understand a particular system, identify what it observes, what is programmed or learned, what it produces or changes, and which evidence supports using it for that task.

References

• David L. Poole and Alan K. Mackworth, Artificial Intelligence: Foundations of Computational Agents, 3rd ed., §1.1 What is Artificial Intelligence?
• Poole and Mackworth, §1.5 Agent Design Space