How Worlds Become Problems?

Seeing the illusion is the beginning of intelligence. — Jiddu Krishnamurti

Intelligence Begins Where Perception Ends

An intelligent agent does not act on the world directly. It acts on a belief about the world.

Sensors do not deliver reality. They deliver fragments. Images, voltages, tokens, signals. These fragments arrive in time, disappear immediately, and might never repeat in exactly the same form. If an agent were forced to decide purely from raw percepts, intelligence would collapse into reflex.

The first act of intelligence, therefore, is not action. It is stabilization.

An agent must transform a stream of transient percepts into something persistent enough to reason with. Something that endures across time. Something that ca be updated rather than replaced. That object is the state.

State is not an optional design choice. It is the minimal requirement for deliberation.

State is Born Inside the Agent Program

Chapter 2 of Russel and Norvig introduces agent programs not as black boxes, but as structured systems. This structure is essential to understanding why state exists at all.

Consider the performance element. It chooses actions. But actions must depend on more than the most recent percept. A single camera frame cannot encode velocity. A single sensor reading cannot encode intent. Time matters, and percepts do not remember time.

So the agent introduces memory. But memory alone is useless unless it is organized. The internal state of the agent is that organization. Formally, every agent program implements a recurrence

\[x_t = u(x_{t-1}, p_t)\]

This equation is the quiet center of Artificial Intelligence. It tells us that intelligence is not a function from percepts to actions, but from histories to actions, mediated by state.

The learning element modifies the update function. The critic evaluates how well the state supports decisions. The problem generator perturbs the state to explore alternatives.

State is the shared medium through which all components communicate. Without it, the agent fragments into disconnected reflexes.

From Memory to World

Up to this point, state is internal. It is the agent’s private construct. But problem solving requires something stronger.

To plan, the agent must imagine futures. To imagine futures, it must assume that the world itself has a structure of independent of observation. This assumption introduces the world state.

The world state is not observed. It is posited.

It represents everything about the environment that is relevant to predicting what will happen next. In early AI, a decisive simplification is made. The agent is assumed to know the world state exactly.

This assumption is false in reality. But it is methodologically powerful. By equating internal state with world state, we temporarily remove uncertainty from the picture. What remains is the pure logic of action and consequence.

This move is what makes problem solving possible.

Representation is the First Model of the World

Once the idea of a world state exists, representation becomes unavoidable. A state must be described somehow. Representation answers a deceptively simple question.

What distinctions matter?

If two physical situations differ, but the agent treats them as the same state, that difference is declared irrelevant. If they are treated as different, that difference becomes actionable. Thus, representation is not passive encoding. It is an assertion about reality.

In the simplest case, each state is treated as an indivisible symbol. This atomic view supports generic search but offers no insight. It hides structure completely.

As soon as we introduce components, structure emerges. A state becomes a vector, a set of variables, a relational description. Now actions affect some components but not others. Constraints appear, Causality becomes explicit.

This is the moment where intelligence stops being blind search and becomes reasoning.

Transitions are Hypotheses about Change

States alone describe nothing dynamic. Intelligence requires change.

A transition model defines how the world evolves when the agent acts. It is the agent’s theory of causation. In deterministic domains, this theory takes the form

\[s_{t+1} = T(s_t, a_t)\]

This equation is not a law of nature, but it is a belief.

When the agent plans, it chains this belief forward, imagining sequences of states that do not yet exist. Planning is simulated. The critical insight from Chapter 3 of Russel and Norvig is that planning is impossible without a transition model. Search algorithms do not discover actions. They explore consequences.

If the transition model is wrong, planning still proceeds, but reality will eventually disagree. This gap between model and world is where intelligence must adapt or fail.

Expressiveness Determines What Can Be Thought

Every representation language has limits. Some distinctions cannot be expressed. Some relationships cannot be stated. A representation is expressive if it can capture the relevant structure of the environment. It is useful if it allows efficient reasoning.

These two goals conflict. Highly expressive representations make reasoning hard. Weak representations make reasoning fast but shallow. This is not an engineering inconvenience. It is a theoretical boundary. Expressiveness increases computational complexity.

Thus, intelligence is not about choosing the most powerful representation, but the right one. A good representation is one in which problems becomes easy.

When Worlds Become Problems

At this point, we have states, actions, and transitions. But intelligence does not operate on worlds. It operates on problems.

A problem is a question posed to the world.

• Can I reach this configuration?

• Can I satisfy these constraints?

• Can I transform this state into that one?

Problem formulation is the act of carving a task out of the world. Formally, a problem begins with an initial state. It defines how those actions change states. And crucially, it defines a goal state.

The goal state is what turns motion into purpose. Without a goal, there is no solution. There is only behavior.

Goals as Sets of States

A goal is not a command. It is a condition. It defines a subset of state space. Any state in this subset is acceptable. The agent does not care how it got there.

This abstraction is profound. It separates achievements from process. By defining goals as states rather than sequences, AI makes planning flexible. Multiple paths can lead to success. The agent is free to choose. This freedom is what search algorithms exploit.

The Emergence of Solution Space

Once problem is defined, something remarkable happens. A space appears. This the solution space. It is not explicitly constructed. It emerges from the interaction of states and transitions. Every node is a state. Every edge is an action. Every path is a candidate solution.

Search algorithms do not solve problems. They explore spaces.

The structure of this space is determined entirely by representation. Branching factor, depth, symmetry, redundancy. These properties are consequences of how states were defined. Algorithms merely react.

Why Canonical Problems Matter?

Classical AI relies on standardized problems not because they are realistic, but because they isolate representational choices.

The vacuum world forces the agent to remember what it cannot see. The 8 puzzle exposes combinatorial explosion. Route finding reveals the importance of cost structure.

Each problem teaches a lesson about state. These problems persist because the lessons persist.

State Comes Before Intelligence

Modern AI often begins with algorithms. Classical AI begins with representation. This ordering is not historical accident. It is conceptual necessity.

Before an agent can decide, it must decide what matters. Before it can optimize, it must define success. Before it can search, it must define space. All of these decisions are encoded in the state.

Closing Reflection

This blog has argued that state is not a technical detail, but the substance of intelligence. It is the medium through which perception becomes memory, memory becomes prediction, and prediction becomes action.

States define worlds. Representations define problems. Problems define solution spaces. Algorithms merely traverse what representation has already shaped.

As we move forward in this series, we will study search, planning, and optimization. These methods will appear powerful. Some will appear elegant. Some will appear crude. But none of them will escape the foundation laid here.

Intelligence does not begin with algorithms. It begins with the choice of state.