A large language model is a neural network trained to predict the next token of text, then tuned to follow instructions and hold a conversation. You do not need to train models for this job, but you do need an accurate mental model of what they can and cannot reliably do, because every agent control rests on it.
Read in this order
- VideoIntro to Large Language Models (Andrej Karpathy)Start here: a one-hour tour of what LLMs are, how they are built, and where they break.
- ArticleCatching up on the weird world of LLMs (Simon Willison)A practitioner's view of the odd behaviors and limits you will meet in practice.
- VideoDeep Dive into LLMs like ChatGPT (Andrej Karpathy)A longer, deeper walkthrough of how a chat model is trained and why it behaves as it does.