Neural network
A computational model loosely inspired by biological neurons — layers of weighted nodes that transform inputs to outputs. LLMs are large neural networks; so are image classifiers, recommendation systems, and most modern AI.
A neural network is a stack of mathematical functions that learn by adjusting weights to minimize prediction errors. The "neural" in the name is metaphor; in practice, neural networks are matrix multiplications followed by nonlinear activations, repeated across layers.
For LLMs specifically, the relevant architecture is the [transformer](/glossary/transformer) — a neural network designed for sequence data with attention as the core mechanism. Modern LLMs have tens to hundreds of billions of weights.
For agent builders, you rarely need to think about neural networks below the API. The exception: if you self-host or fine-tune, neural network properties (depth, attention heads, vocabulary size) affect cost, capability, and behavior.
Frequently asked
What is the difference between a neural network and an LLM?+
An LLM is a specific kind of neural network — large, transformer-based, trained on text. "Neural network" is the broader category that includes image models, audio models, recommendation systems, and many other AI applications.
Are neural networks always better than other ML?+
No. For tabular data with fewer than 10K rows, tree-based models (XGBoost, LightGBM) often beat neural networks. For language, vision, and audio: neural networks dominate.