Wiki · AI & Machine Learning · concept
Large language model
A transformer trained on huge amounts of text to predict the next token, which yields broad language and reasoning ability.
A large language model is a transformer trained on enormous amounts of text to do one deceptively simple thing: guess the next token. Do that well enough, over a large enough slice of everything humans have written, and the model has to internalize grammar, facts, arguments, and the shape of reasoning itself. The “large” is literal — billions of parameters, trained on a big fraction of the public web — and it’s what turns a next-word predictor into a general-purpose language engine.
How it works
The training task never changes: read some text, predict what comes next, adjust, repeat. What changes everything is scale. Feed a big enough model enough text and new abilities appear that no one hand-coded — summarizing, translating, writing code, following instructions. The most striking is in-context learning: you can teach the model a new task inside the prompt, just by showing a couple of examples, without retraining anything. The model isn’t looking anything up; it’s continuing the pattern you started, one token at a time.
Why it matters (for bio × AI)
In agentic science and drug-discovery agents, the LLM is usually the dispatcher, not the expert. It reads the goal, decides which tool to call — a docking program, a protein structure predictor, a database — reads the result, and picks the next move. That’s orchestration, not intelligence: the reasoning is real, but the hard science lives in the specialist tools it coordinates. Separately, the same recipe transfers beyond English: DNA, RNA, and protein “language” models treat sequences as text and predict the next residue, learning biology’s grammar the way an LLM learns ours.
Related
See transformer for the architecture underneath, foundation model for the broader class an LLM belongs to, fine-tuning for how a general model gets adapted to a specific job, and protein language model for the same recipe pointed at biological sequences instead of text.