Models are learned functions
At its core, a machine-learning model is a parameterized mathematical function. It receives numbers, transforms them through many operations, and produces numbers interpreted as classifications, predictions, actions, or probabilities. Training does not insert a database of correct answers. It adjusts parameters so outputs become less wrong according to an objective measured across examples.
A neural network organizes those parameters into layers. Each layer transforms a representation into another representation. Early layers may identify local or simple patterns; later layers combine them into more abstract ones. Backpropagation efficiently estimates how each parameter contributed to error, and an optimizer makes small updates over many training steps.
Language becomes geometry
Language models divide text into tokens—words, fragments, punctuation, or other units. Each token is mapped to a vector, a list of numbers representing its position in a learned geometric space. Similar patterns can occupy related regions, but these vectors are not dictionary definitions. They are useful statistical representations shaped by the training objective.
The model predicts a probability distribution over the next token. Repeating that process produces paragraphs, software, or structured data. Apparent planning can arise because training data contains patterns of explanation and because newer systems use additional computation, tools, verification, or intermediate reasoning. Fluency remains different from guaranteed truth.
Attention and transformers
Self-attention allows each token representation to incorporate information from other tokens. The model computes query, key, and value vectors. Similarity between queries and keys determines how strongly information from values should influence the next representation. Multiple attention heads learn different relationships in parallel.
Transformer blocks combine attention with feed-forward networks, normalization, and residual connections. Stacking many blocks creates a system able to model complex relationships across a context window. Transformers train efficiently on modern hardware because sequence positions can be processed in parallel, one reason the architecture scaled so successfully.
Pretraining, adaptation, and inference
During pretraining, a model processes enormous collections of examples and learns broadly reusable structure. Instruction tuning then uses curated examples to make the model follow requests. Preference optimization and safety training can encourage helpful behavior and discourage harmful behavior. Retrieval can supply selected external documents at use time, and tool calling can let the model request calculations, searches, or approved actions.
Inference is the use of the trained system. Temperature and sampling affect variation. Context affects what information is available. System instructions and permissions bound behavior. The final product is therefore more than model weights: it is a sociotechnical system containing interfaces, data pipelines, evaluators, access controls, monitoring, and people.
Agents are systems, not magic
An AI agent combines a model with a goal and software tools that allow it to perform more than one step. The word does not imply consciousness, independence, or unlimited ability. Performance depends on the model, the task, the surrounding software, and the conditions in which it is used.
Because tool-using systems can affect information or services beyond a conversation, their real-world behavior should be evaluated in context. A fluent interaction alone does not establish reliability.
Why intelligence remains uneven
Modern systems can excel on benchmarks while failing simple variations. They may memorize surface regularities, exploit unintended hints, or break when a problem requires longer chains of exact reasoning. Stanford’s AI Index reports dramatic progress on coding and multimodal benchmarks, while noting persistent weaknesses in complex reasoning and long-horizon agent tasks.
The correct mental model is neither calculator nor person. A foundation model is a broad probabilistic engine whose abilities emerge from data, scale, architecture, and adaptation. It can be astonishingly capable and still require verification. Technical literacy begins by holding both facts at once.
Primary research and institutions
Sources are linked directly so readers can examine the underlying evidence. Numerical statements identify the reporting organization and year. Projections are presented as estimates, not established future facts.