Topic 2: How large language models work
How Large Language Models work
Large Language Models (LLMs) learn patterns from very large amounts of text. They use these patterns to understand your input and generate a response one piece of language at a time.
For Product Managers and Business Analysts, you don’t need to understand the complex mathematics behind an LLM. What matters is understanding how your input becomes an output and why the quality of that output can vary.
How an LLM processes your request
- You provide an input: You ask a question, give an instruction, share a document or provide context.
- The input is broken into tokens: The LLM processes language as smaller units called tokens, rather than reading sentences exactly as humans do.
- The model uses learned patterns: It considers the tokens, their relationships and the context you provided.
- It predicts what comes next: The LLM generates its response by predicting the most likely next token, then repeating this process to build the answer.
- You review the output: The response may sound confident and well-structured, but it still needs human review.
Why this matters for PMs and BAs
Understanding this process helps explain why an LLM can produce an excellent answer for one request and a weak answer for another.
An LLM does not automatically understand your organisation, product, customers or business rules. The clearer and more relevant the context you provide, the better equipped the model is to produce a useful response.
DSM Tip
Think of an LLM as a powerful language and reasoning assistant, not as a source of guaranteed truth.
Use it to help you explore, analyse, structure and draft your work but you remain responsible for reviewing the output and deciding what should actually be used.
