AI as a System (Part 2): What AI Models Actually Are

This is Part 2 of the AI as a System series.
See the full series here.

The Engines: Why GPT, Claude, and Others Feel Different

In the last post, we established that AI is a stack made up of models, tools, instructions, and memory. This post focuses on the foundation of that stack: models.

What a Model Actually Is

A model is not an app, an interface, or something you interact with directly.

A model is a system trained to predict the next piece of information based on context.

In Part 1, we simplified this as “predicting the next word.” That’s still accurate - but the important detail is that this prediction happens repeatedly, allowing the model to generate full responses.

For language models, this means taking input text, predicting the next word or phrase, and repeating that process many times. At scale, this powers everything from chatbots to code generation.

Large Language Models (LLMs) vs Machine Learning (ML)

Machine learning is a broad field focused on building systems that learn patterns from data to make predictions without being explicitly programmed.

Large language models are a specialized subset of machine learning. They are trained on large amounts of text data and designed to understand and generate language by predicting the next token. This makes them well-suited for tasks like writing, summarization, and conversation.

AI is Advanced Pattern Prediction

Different models are optimized for different strengths. Some are better at reasoning, others at long-form writing, coding, or real-time information. For a high-level comparison of major models, see Part 1 of this series.

Models do not understand meaning, intent, or context in the way humans do. They identify and reproduce patterns.

Because those patterns are learned from massive datasets, the output can appear to involve reasoning or judgment. That appearance is what makes modern AI feel intelligent.

Why Models Feel So Different From Each Other

Although models operate on similar principles, they can behave very differently in practice.

Even though models are built on similar principles, they can behave very differently in practice.

Training data shapes tone and knowledge. Training approach (fine-tuning, reinforcement learning, human feedback) shapes behavior. Context window determines how much information the model can work with at once.

A context window is the maximum amount of text, measured in tokens, that a model can process and reference at one time.

A larger context window allows the model to consider more information at once, improving its ability to reason across longer inputs. When the limit is exceeded, earlier parts of the conversation may be lost.

The Model Is Not the Tool

As a reminder from Part 1, the model is not the tool.

The model is the underlying engine. The tool is the interface that wraps it. This distinction matters because the same model can behave very differently depending on how it is used.

Why Model Choice Matters

Model choice directly affects output quality.

Stronger models tend to require fewer corrections, handle edge cases more effectively, and produce more structured responses. Weaker models often need more guidance and are more likely to miss important details or generate incorrect information.

The Beginning of Multi-Model Thinking

Instead of relying on a single model, it is often more effective to choose models based on the task.

Model routing is the practice of selecting the most appropriate AI model for each task instead of using a single model for everything.

Smaller and faster models are often better for simple or high-volume tasks. More capable models are better suited for complex reasoning. Some models are stronger for coding, while others are better for real-time information.

This reflects a shift toward treating models as components within a system rather than a single solution.

What’s Next

Now that the role of models is clear, the next step is to look at how interfaces shape the experience—how the same model can feel completely different depending on the tool you use.