Large language model
Predicts the next piece of text, which lets it chat, write, summarize, translate and follow instructions.
What it is
A large language model (LLM) is a neural network trained on a huge amount of text to predict what comes next. Scaled up and then tuned to follow instructions, that one skill turns into a general assistant: it can answer questions, draft and edit writing, summarize documents, write code and call tools.
How it works
Text is split into tokens. A transformer reads the tokens so far and outputs a probability for every possible next token; one is picked, appended and the process repeats. Pre-training on web-scale text teaches language and facts, then instruction tuning and reinforcement learning from feedback make it helpful and safer to use.
What it is good for
- Chat assistants and customer support
- Writing, editing and summarizing
- Extracting structured data from text
- Powering agents that plan and call tools
Limits to know
- Can state wrong things confidently (hallucination)
- Knowledge stops at a training cutoff unless given search or documents
- Large models are costly to run and slower than small ones
