THE SHORT ANSWER

A large language model is a model trained at scale to work with language. Many text-generating LLMs predict a sequence one token at a time, using the context available to them. A token can be a word, part of a word or punctuation.

From a prompt to a sequence

A text-generating model assigns probabilities to possible next tokens and selects an output sequence. Training has adjusted its parameters so those predictions reflect patterns in the training material. The 2020 GPT-3 paper is one primary example of this autoregressive approach.

A useful mental picture is a model continually extending the text it has been given. But ‘autocomplete’ alone can understate the complexity of the patterns it has learned. Neither comparison tells you whether a particular answer is correct.

Evidence & context: Brown and colleagues, 2020

Context helps steer an answer; it is not permanent knowledge

Put a short passage in a prompt and ask for a summary. The passage becomes part of the material available for that response. This differs from changing the model's learned parameters. A chat application's memory features, document search and conversation history are additional mechanisms; do not assume they are all the same thing.

For an illustrative test, describe an imaginary library with two membership rules. Ask questions that are answered by those rules, then one that is not. A useful response to the last question should identify missing information rather than invent a third rule.

Where the Transformer fits

The 2017 Transformer paper introduced an architecture built around attention mechanisms. Attention lets representations incorporate relationships among elements of a sequence. The paper's original application was not a general-purpose chat assistant; it evaluated sequence-transduction tasks, including translation.

Architecture is only one part of an LLM system. Training data, objectives, later training, available context and the application around the model also affect its behaviour. ‘It uses a Transformer’ is not a quality assessment.

Evidence & context: Vaswani and colleagues, 2017

A language model is not automatically a browser

Retrieval-augmented generation combines generation with retrieved material. This is one way to supply information beyond what is represented in model parameters. It also creates questions about which sources were retrieved and whether the answer used them correctly.

Likewise, a model connected to tools may form part of an AI agent. Ask what the product can access and execute, not just which model name appears on the screen.

  • Does the answer come from supplied material, retrieved sources or unsupported generation?
  • Can I inspect the evidence for the consequential claims?
  • Does the system recognise when the context is insufficient?

Evidence & context: Lewis and colleagues, 2020

Sources & further reading

  1. Language Models are Few-Shot Learners

    Brown and colleagues, 2020. Primary research on an autoregressive language model and learning from examples in context; not a guarantee about all models.

  2. Attention Is All You Need

    Vaswani and colleagues, 2017. The original Transformer paper. Its experiments concern sequence transduction, particularly translation.

  3. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Lewis and colleagues, 2020. Research combining generation with retrieved material. Retrieval should not be interpreted as proof of factual accuracy.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗