Knowledge, Context and Agentic Engineering for Knowledge Work

System understanding

Module 01 · “Understanding” Large Language Models

A functional account of what language models represent and generate, including uneven capability and the distinction between model, product, harness and infrastructure.

Scope and learning question

Learning question: Which properties of an observed result follow from the model, its context and the surrounding system?

Language models generate continuations from token sequences. Their observable behaviour also depends on training, runtime context, product design, harness functions and deployment infrastructure.

Learning objectives

  • Explain next-token prediction, tokens, contextual representations, training and inference at a functional level.
  • Analyse model output as a representation that requires source-sensitive checking.
  • Locate an observed capability or failure within the model, current context or surrounding system.

Concept map and starter sequence

  1. Research materialis represented as text, images, data or metadata
  2. Tokens and contextform the input available for the current inference
  3. Trained modelestimates and generates a next-token sequence
  4. Product and harnessadd runtime instructions, tools, state and interface behaviour
  5. Observed resultinherits capabilities and limits from the compound system

Read in order: a scholarly object is represented for computation; that representation enters a bounded context; a trained model generates a continuation; product, harness and infrastructure shape the result that can be observed.

Research data are constructed representations

A facsimile, transcription and TEI file expose different features of one source. Each representation carries scholarly decisions, affordances and uncertainty, so generated transformations must remain connected to the material and rules that justify them.

Next-token prediction drives generation

An autoregressive language model estimates a probability distribution over possible next tokens. Transformer layers and self-attention build context-dependent representations that let information at different sequence positions influence the continuation.

Tokens become contextual representations

Tokenisation converts character sequences into discrete units. Embeddings initialise numerical representations, which repeated model layers transform according to surrounding tokens and the current task.

Training and inference play different roles

Pretraining and posttraining change model parameters and shape available capabilities and assistant behaviour. Ordinary inference uses those trained parameters with current instructions and examples; it does not normally update the weights.

Capability and failure are uneven

A model can solve a demanding task and fail on a nearby task that appears simple. Plausible unsupported claims, bias, sycophancy and incomplete interpretability make external evidence and task-specific evaluation necessary.

Model, product and deployment form distinct layers

Parametric knowledge, retrieved resources and current context are different information layers. Runtime instructions, tools, permissions and infrastructure extend the model into a compound system whose behaviour cannot be attributed to model weights alone.

Learning routes

  • Available workshop Web Module

    Where language models can assist

    Relates model-supported tasks in an estate-material workflow to the sources and rules required for checking.

  • Available teaching route

    Module 01 Video Library

    Extends the starter sequence through the maintained primary-module selection.

Teaching material

Primary videos

Open the complete Module 01 group in the Video Library · Open the ordered Video Playlist

Representative workshops

All seven registered workshops include Module 01. These three profiles show different audiences and research-data settings.

Browse all workshop profiles