LLM (Large Language Model)
An LLM is a neural network trained on vast amounts of text that predicts the next fragment of a word. Everything else grows out of that single mechanism: answers, translation, code, reasoning.
How it works
The model stores no text and searches no database. It stores weights, billions of numbers tuned so that a beginning of a phrase predicts its continuation. An answer is built one token at a time, each chosen in light of everything before it.
Training runs in two broad stages. Pre-training on a corpus of text teaches the model language and factual associations. Then fine-tuning on examples of desired behaviour and human feedback teaches it to answer usefully rather than merely continue plausibly.
Hence the key practical consequence: the model does not "know" in the human sense, it reproduces patterns. Confidence in the tone of an answer is unrelated to whether the answer is right.
Where it is strong and weak
- Strong: working with text you provideSummarising, translating, spotting contradictions, rewriting in another register. The material is in front of it and nothing needs inventing.
- Strong: structure and first draftsBreaking down a task, proposing a plan, sketching code: places where shape matters and verification stays with you.
- Weak: precise facts from memoryDates, figures, quotations and links come out plausible rather than reliably correct. All of it needs separate checking.
- Weak: arithmetic and recent eventsStep-by-step calculation is harder for a model than prose, and its knowledge stops at the training date unless it can search.