Large Language Models: Foundations and Internals

About this course

The entry point to the generative-AI track. Learners open up a large language model layer by layer: tokenisation, embeddings, attention and the transformer block; pretraining, scaling laws and post-training; decoding, inference mechanics and long context; reasoning models and the current landscape of hosted and open-weight model families. The aim is a working mental model precise enough to predict how a model will behave, debug it when it misbehaves, and explain its limits honestly to colleagues and clients.

For: software engineer, ml practitioner, data professional, student, architect.

Learning outcomes

On successful completion you will be able to:

1. 1Explain how text becomes tokens and embeddings, and how attention, positional encoding and the transformer block turn them into next-token predictions.
2. 2Describe how LLMs are pretrained and post-trained, and how data, compute and parameter scaling shape model capability and cost.
3. 3Implement and instrument the core components of a small transformer and its decoding loop in Python, and relate each to its production counterpart.
4. 4Analyse how decoding strategy, context length and prompt structure affect output quality, latency and cost.
5. 5Compare hosted and open-weight model families, including reasoning models, against task requirements, context needs, cost and data-residency constraints.
6. 6Judge the limitations and failure modes of LLMs (hallucination, bias, brittleness, knowledge cut-off) and communicate them honestly to stakeholders when proposing an LLM solution.

Modules and lessons

1. Week 1From text to tokens and vectors⌄

   • Tokenisation: how text becomes tokenslab
   • Embeddings and the geometry of meaning
2. Week 2Attention and the transformer⌄

   • Attention: queries, keys and valueslab
   • Anatomy of a transformer block
3. Week 3Pretraining and scale⌄

   • Pretraining: objectives and data
   • Scaling laws and compute budgetslab
4. Week 4Post-training: from base model to assistant⌄

   • From base model to assistant
   • Reasoning models and test-time compute
5. Week 5Generation and inference⌄

   • Decoding: how a model chooses its next tokenlab
   • Inference mechanics: prefill, decode and the KV cache
6. Week 6Context: windows, prompts and caching⌄

   • Context windows and long context
   • Prompting as interface designlab
7. Week 7The model landscape⌄

   • Model families: hosted and open-weight
   • Choosing a model for a task
8. Week 8Limits, failure modes and responsibility⌄

   • Hallucination and other failure modes
   • Looking inside, and speaking honestly about, LLMs

Assessment

• Quiz (25%)

  Eight weekly quizzes drawn from the lesson quiz banks; best seven count.

• Lab (35%)

  Five graded notebook labs, from a byte-pair tokeniser to a working decoding loop over a small transformer.

• Project (25%)

  Model-selection brief for a realistic scenario - shortlist, evaluation on a 30-prompt test set, cost model and an honest limitations section.

• Viva (15%)

  Twenty-minute technical conversation in which the learner explains their transformer components and defends the brief.

Pass mark 60%. Grades: Distinction, Merit, Pass.