Large Language Model (LLM)

What Is Large Language Model (LLM)?

A large language model (LLM) is a neural network trained on enormous amounts of text to predict the next token in a sequence. At sufficient scale, that single objective produces broad ability to write, summarize, reason, and generate code, which is why LLMs power chat assistants, coding agents, and most modern AI products. Adoption reflects that reach: in the 2025 Stack Overflow survey, 84% of developers said they use or plan to use AI tools in their work, up from 76% the year before [1].

Key Takeaways

• Everything an LLM does is next-token prediction. Chat, code, and reasoning are all the same operation: pick the most plausible continuation, one token at a time.
• LLMs are stateless. The model remembers nothing between API calls; the entire conversation is resent inside the context window every turn, and you pay for those tokens every time.
• Capability comes in stages: pretraining builds raw knowledge, then post-training with human feedback shapes the model into a usable assistant that follows instructions.
• For engineering work, model choice is a cost-latency-quality triangle. Frontier models handle complex multi-step tasks; smaller models are often ten times cheaper and fine for classification, extraction, and simple edits.

How It Works

Training starts with pretraining: the model, almost always a transformer model, reads trillions of tokens of text and code and adjusts billions of internal weights to get better at guessing what comes next. Nothing is labeled and no task is specified. Grammar, facts, APIs, and idioms all get absorbed as statistical structure because they help predict continuations. The scale keeps climbing: Stanford's 2025 AI Index found that the compute used to train notable models doubles roughly every five months, and training datasets double every eight [2]. A second phase, post-training, uses instruction examples and human preference feedback to turn that raw predictor into an assistant that answers questions, refuses harmful requests, and calls tools.

At inference time the loop is simple. Your prompt is split into tokens through tokenization, the model computes a probability for every token in its vocabulary, one is sampled, appended, and the process repeats until a stop condition. Sampling settings like temperature control how adventurous those picks are, which is why the same prompt can produce different outputs.

The practical consequences for engineers follow directly from this design. Because the model only predicts plausible text, it can produce fluent, confident output that is wrong, the failure known as AI hallucination. Because it is stateless, agent frameworks must manage what the model sees each turn, which is the whole discipline of context engineering. And because cost scales with tokens processed, long transcripts and fat prompts translate directly into API spend.

Example

A team building a coding agent routes work across two LLMs. A frontier model plans the change and writes the diff for a multi-file refactor, at roughly the cost of a coffee per complex task. A small, fast model handles the high-volume chores around it: writing commit messages, labeling failing tests, and summarizing tool output before it enters the main agent's context. The split cuts their monthly inference bill by more than half without touching the quality of the shipped diffs, because the expensive model only sees the work that needs it.

What People Get Wrong

The persistent misconception is that an LLM looks things up in a database of stored documents. It does not contain its training data and cannot quote it reliably. Facts live as diffuse patterns in the weights, which is why a model can explain an algorithm perfectly and then invent a plausible-looking citation in the same answer. When you need verifiable facts, you attach retrieval or tools and practice AI grounding rather than trusting recall.

FAQ

What does LLM stand for, and what makes one "large"? LLM means large language model. "Large" refers to parameter count, the learned weights inside the network, and to the training data volume. GPT-3, the model that kicked off the modern LLM era, was trained with 175 billion parameters in 2020, ten times more than any previous non-sparse language model [3]. Modern models range from a few billion parameters for on-device use to frontier systems estimated in the trillions.

Is ChatGPT an LLM? ChatGPT is a product built on top of LLMs, wrapped with a chat interface, tools, memory features, and safety systems. The distinction matters when debugging behavior: many quirks come from the product layer, not the underlying model.

What is an LLM in simple terms? It is an extremely capable autocomplete. It read a huge portion of the internet and learned to continue any text plausibly. That turns out to cover writing, coding, and answering questions, as long as you remember that plausible and true are different properties.

Sources

1. Stack Overflow. "2025 Developer Survey, AI section: 84% of developers use or plan to use AI tools." https://survey.stackoverflow.co/2025/ai. Accessed August 2026.
2. Stanford HAI. "AI Index Report 2025: training compute doubles roughly every five months, datasets every eight." https://hai.stanford.edu/ai-index/2025-ai-index-report. Accessed August 2026.
3. OpenAI. "Language Models are Few-Shot Learners: GPT-3 trained with 175 billion parameters." https://arxiv.org/abs/2005.14165. Accessed August 2026.