LLM

A large language model: a neural network trained on vast text corpora to predict and generate language, the engine behind modern AI features.

A large language model is a neural network with billions of parameters trained on enormous text corpora to do one thing extremely well: predict the next token. Out of that single objective emerges translation, summarization, code generation, reasoning chains and conversation. GPT, Claude, Gemini, Llama, Mistral and DeepSeek families are all LLMs, accessed through APIs or run locally with tools like Ollama.

Why it matters to an indie builder: the LLM is the engine of every AI feature you will ship, and treating it as a component, not magic, is what makes your product buildable and affordable. The model handles language; your product still owns the data, the workflow, the interface and the accountability for results.

A concrete example: a contract-review feature. Sending a full contract, about 15,000 tokens, to a frontier model with a carefully written instruction yields a clause-by-clause risk summary for roughly $0.03 per contract at current prices. The same task with a human reviewer bills minutes at far higher rates. The feature is viable only because you can reason about tokens, prices and latency, which is why the token and inference glossary entries pair with this one.

Common mistakes: picking a model before defining the task (start with the cheapest model that handles the job, upgrade where quality measurably suffers), trusting outputs without validation (models hallucinate confidently; ground them with your data and verify critical claims), ignoring version pinning (provider deprecations have broken production apps overnight), and rebuilding infrastructure the ecosystem already offers: guardrails, tracing and evaluation are solved problems in the llm-ops tooling listed on IndieBoot. The aiaas blocks compare current providers side by side. The llm-ops blocks on IndieBoot exist to make that ownership practical.