How do AI, machine learning, deep learning, and LLMs fit together?

A useful nesting, not a product specification

These terms answer different questions:

• Artificial intelligence asks what behaviour the complete system can produce: classify an image, recommend an item, answer a question, or act through tools.
• Machine learning asks how part of that behaviour is acquired: parameters are adjusted from examples or feedback instead of every rule being written by hand.
• Deep learning names machine-learning methods built from many layers of learned numerical operations.
• A large language model is a deep neural network trained at large scale to model sequences of tokens. Most current LLMs use the Transformer architecture.

flowchart LR
  AI[AI systems] --> ML[Machine learning]
  ML --> DL[Deep learning]
  DL --> LM[Large language models]
  LM --> APP[Chat, code, search, or agent product]
  RULES[Rules, retrieval, tools, UI, policy] --> APP

The final product is not simply the smallest box. A chat application may combine an LLM with ordinary code, search, databases, moderation, identity, permissions, and a user interface. A coding agent adds a harness and tools. Calling the entire product “the model” hides the boundaries engineers need to test.

Learning means changing parameters

A traditional program contains rules chosen directly by people. A learned model contains parameters—numbers that training adjusts so examples receive better predictions under a loss function. Deep networks stack simple operations so later layers can work with representations learned by earlier ones.

For a language model, text is tokenized into IDs. The model receives a sequence and produces scores for possible next tokens. During pretraining, an optimizer repeatedly changes the weights to reduce next-token prediction error across batches of data. After training, inference holds the weights fixed while the system calculates outputs for a new input.

An LLM therefore does not store a clean database of sentences and rules. Its behaviour emerges from learned weights, the current context, decoding, and everything the surrounding application does with the result.

Why the labels blur

The categories overlap rather than forming a perfect taxonomy. Some AI systems use planning or rules without machine learning. Some machine-learning models are small decision trees rather than deep networks. Transformers can process images, audio, or mixed inputs, so “language model” may describe only part of a multimodal system.

“Large” has no timeless parameter threshold. “Foundation model,” “generative AI,” and “frontier model” emphasize other properties and are used differently across research, policy, and marketing. When a distinction matters, name the concrete model, training objective, input and output types, system components, and version.

Do not collapse the system into one label

AI is not a synonym for an LLM, and an LLM is not an agent. A model generates outputs; an agent system adds a loop, tools, state, policy, and stopping conditions.

Deep learning is not a guarantee of understanding, correctness, or general intelligence. A fluent response is observable behaviour, not evidence that the system represents knowledge in the same way a person does.

Use the vocabulary to locate failures

Begin by drawing the boundary of the system under test. List every learned model, deterministic component, data source, tool, policy gate, and human decision. Mark which versions can change independently.

Test a learned model on held-out examples and meaningful slices, but test the product at its own interfaces and outcomes. A model-quality improvement can still cause a product regression through different tokenization, latency, cost, formatting, tool proposals, or safety behaviour.

Ask three questions during incident analysis: did the model produce an unsuitable output, did the harness make a bad decision, or did the application use a valid output incorrectly? The vocabulary is useful when it helps locate ownership and evidence.