AI, Machine Learning, Neural Networks, and LLMs: A Beginner's Map

AI, machine learning, neural networks, deep learning, generative AI, and large language models describe different scopes. A product can use several of them at once, so treating the terms as synonyms makes later questions about training, hardware, and model files harder to answer.

The map shows a common modern path, rather than a rigid taxonomy for every historical system. An application can combine learned models with search, rules, tools, and databases.

AI is the broad field

Artificial intelligence covers machine-based systems that perform tasks such as prediction, perception, planning, recommendation, generation, or decision support. The implementation does not have to learn from data.

A chess program can search future positions using hand-written rules and an evaluation function. That is an AI system even if it does not train a modern machine-learning model. This gives the first useful boundary: not all AI is machine learning.

Machine learning changes a model from data

Traditional code expresses behavior directly through instructions. Machine learning uses examples or feedback to adjust parameters so a model improves according to an objective.

A spam classifier might train on messages labeled spam and not spam. The learning procedure changes the model; after training, the resulting model maps a new message to a score or label. Linear regression, decision trees, support-vector machines, nearest-neighbor systems, and neural networks are all machine-learning approaches.

This gives the second boundary: not all machine learning uses neural networks.

Neural networks and deep learning

A neural network applies layers of parameterized numerical transformations. Weights control how input values contribute to later values; biases shift those calculations; activation functions introduce nonlinearity.

Deep learning uses neural networks with multiple learned layers. In an image model, early representations may respond to local visual patterns while later layers combine them into more useful features. The network does not need a human to name each intermediate feature.

Large datasets, accelerator hardware, automatic differentiation, and improved architectures made it practical to train deep networks at scales that earlier systems could not reach.

Generative AI produces new sequences or media

Many models classify or estimate: image to label, transaction to risk score, or house features to price. Generative models produce new content or structured output, including text, images, audio, video, and code.

Generative AI is a product and model category, rather than another name for an LLM. Diffusion-based image generators and autoregressive language models can both be generative while using different architectures and output representations.

An LLM models token sequences

A large language model is a large neural-network model trained on language-like token sequences. Influential modern LLMs use the Transformer architecture introduced in Attention Is All You Need.

During generation, an autoregressive LLM repeatedly estimates scores for the next token, selects one continuation, appends it to the sequence, and runs the next step. Scaling this training objective across large datasets and models can produce broad language, code, and few-shot capabilities.

An LLM is therefore a machine-learning model, a deep neural network, and part of AI. AI also includes many systems that are not LLMs.

Model, runtime, and application are separate

These three pieces are easy to collapse into the word “AI”:

Changing a user interface does not retrain the model. Changing a runtime does not necessarily change the weights. Quantizing an artifact changes its numerical representation while preserving the model family and architecture.

A real application can also combine an LLM with search, retrieval, ranking, safety filters, tools, authentication, and ordinary business logic. Ask which component performs a claimed behavior before attributing the whole product to the model.

Use the map to place later topics

Training changes model parameters. Inference uses trained parameters on new input. Tensors organize the numbers. A forward pass transforms an input into an output. Loss measures a prediction against a target. Gradient descent uses derivatives to update parameters.

Those concepts connect in sequence. What Does It Mean for a Machine to Learn? begins that path with one parameter and one measurable objective.