How an LLM works
Google’s machine learning course describes a language model as a system that estimates the probability of a token, a word or piece of a word, appearing within a longer sequence (Google for Developers). Large models, built on the transformer design, can take in the whole context at once instead of reading one token at a time. Trained on enough text, they can summarise, translate, draft and answer questions.
What LLMs are good at
- Drafting emails, product descriptions and social posts.
- Summarising long documents and meeting notes.
- Explaining a concept at different levels.
- Turning messy notes into a table or a structured list.
Where they fall short
Because an LLM predicts likely text, it can state wrong facts with full confidence. This is called an AI hallucination. It may also not know about recent events or local rules. Use LLMs to speed up work you can check, and give them the facts they need inside the prompt.