What is a large language model, and who makes GPT, Claude, and Gemini?

What is a large language model, and who makes GPT, Claude, and Gemini?

A large language model is a computer program trained on enormous amounts of text so it can produce text in response to a request. When someone types a question and a program writes back a fluent answer, a large language model is doing the writing. The abbreviation is LLM, and when a resume or a job description says "AI" in a software context, an LLM is usually the thing being described.

GPT, Claude, and Gemini are the three model names that come up most. They are products, made by OpenAI, Anthropic, and Google respectively. The relationship is the same as Amazon and AWS: one is a company, the other is the thing it sells.

Every name in this lesson has a shelf, the way every technology name does. The sections below unpack the words one at a time: what a model is, what "large" and "language" add, who the labs are, and the three terms that travel with every model everywhere: prompt, token, and context window.

What kind of thing is a model?

A model is software whose behavior was learned from data rather than written line by line. That single sentence separates it from everything covered so far in this course.

An ordinary program is a set of instructions a developer typed out. Every rule it follows exists because someone wrote it, and when it misbehaves, someone can read the code and find the line responsible. A model works differently. Its builders choose the data and run the learning process, and the behavior emerges from the patterns in that data. Nobody wrote a rule anywhere inside GPT that says how to answer a question about Python. The pattern was absorbed from millions of examples.

The word "model" carries its meaning honestly. The program models the patterns in its data, the way a scale model represents a real building. This is also why the builders of a model can be surprised by it: a team that trained a model discovers what it does by testing it.

Language models are one kind of model among many. Models built for images power photo generation and medical scan reading, and recommendation models ran quietly inside products for a decade before anyone said "AI" in a hiring meeting. Each is a model in the same sense: behavior learned from data, built for one kind of material.

What makes a language model large?

"Language" names the material: text in, text out. The model read an enormous share of the written internet during training, and everything it does afterward is reading text and producing text. Answering questions, summarizing documents, drafting emails, and translating are all text tasks, which is why one kind of program handles all of them.

Written human text has a name in this field: natural language. English, Spanish, and Japanese are natural languages. The contrast is with a programming language, which is written for machines and follows strict rules. A large language model handles both, because code is text too. The same model that drafts an email can write working Python, since to the model a Python file and a paragraph of prose are both sequences of text whose patterns it learned.

"Large" describes the machinery. Inside the model are billions of small adjustable settings, and the learning process tunes every one of them. More settings mean the model can hold more patterns: more facts, more writing styles, more programming languages. Size also carries a bill, because a bigger model needs more computing power for every single answer. This is why each lab sells its models in several sizes, and why a developer choosing between a large model and a small one is making a cost and speed decision as much as a quality decision.

Who are the labs, and what do they sell?

Three companies make the models that dominate resumes and job descriptions. The industry calls them labs, a name kept from the years when this work was research rather than product.

OpenAI makes the GPT family of models and ChatGPT, the app built on top of them. ChatGPT is the product that made this whole field public in late 2022, and OpenAI remains the lab most people name first.

Anthropic makes Claude. Anthropic was founded in 2021 by former OpenAI researchers, and Claude is both the model family and the consumer app built on it. Claude models are particularly common in coding tools, so the name shows up often on engineering teams.

Google makes Gemini. Google researchers published the 2017 breakthrough that today's language models are built on, and Gemini is the model family Google ships across its own products and sells to developers.

Each lab sells the same two things. There is a consumer app, the chat window a person types into, and there is paid access for developers through an API, so other companies can build products on top of the models. That second business is the one that matters for engineering hiring, and the full map of who builds on whom is the subject of the next lesson.

What is training?

Training is the one-time process that produces a model from data. The mechanics are a prediction game played at massive scale. The model is shown a stretch of text with the next piece hidden, it guesses, and its billions of internal settings get nudged toward whatever would have made the guess better. Repeated across a huge portion of everything ever written, this process leaves the model holding grammar, facts, reasoning patterns, and style.

At the frontier, training is one of the most expensive activities in software. It runs for months across thousands of specialized machines in data centers, and the budgets reach hundreds of millions of dollars. That cost is why only a handful of companies train large models from scratch, and it explains the shape of the entire industry: a few labs train, and everyone else builds on what they trained.

Once training ends, the model is frozen. Using it afterward changes nothing inside it, which is why a model can hold a long conversation and remember none of it the next day. The gap between the few developers who train models and the many who build on them is the highest-leverage distinction in AI hiring, and it gets its own lesson.

What is a prompt, and what is a context window?

A prompt is the text sent to a model. When a person types a question into ChatGPT, that question is a prompt. When a company's product sends a customer record to Claude with instructions to summarize it, the record and the instructions together are the prompt. Everything a model does starts with one.

A token is the unit a model actually reads and writes. A token is a small chunk of text, roughly three quarters of an English word, so a 1,000-word document is about 1,300 tokens. The word keeps its old meaning of a small standard unit used for counting, like a subway token. Tokens matter beyond trivia because they are the billing unit of the whole industry: the labs charge developers per token, for the text going in and the text coming out.

The context window is the amount of text, counted in tokens, that a model can consider at one time. Everything inside the window is visible to the model when it answers. Everything outside the window might as well never have been written, and has to be supplied again or looked up. The conversation so far, any documents handed over, and the instructions a developer wrote all share the same budget. Windows have grown from a few thousand tokens to a million or more, enough to hold several long books, and deciding what belongs in the window is now a real engineering skill, one the module on how developers work with AI picks up in full.

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

The three terms nest inside each other, widest first.

AI, artificial intelligence, is the broadest label: software performing tasks people associate with intelligence. The term dates to the 1950s and has covered many technologies since, which is why it is the vaguest word on any resume.

Machine learning is the part of AI where behavior is learned from data. Everything in this lesson sits inside it. A resume that says machine learning describes work with models of some kind, though the kind varies widely: a fraud detector, a recommendation system, and a language model are all machine learning.

Deep learning is the part of machine learning built on layered networks loosely inspired by the brain, and it is the branch that produced today's large models. When the field made its leaps over the last decade, deep learning was the engine.

So every large language model is deep learning, all deep learning is machine learning, and all machine learning is AI. The nesting reads in one direction only. A candidate with machine learning experience from 2015 has real, relevant depth, and it may sit in a different branch of the field than the LLM work in this module. The rest of the module's map sorts those branches into the companies, the roles, and the claims a recruiter actually meets.