The Workbench

Welcome to The Workbench — field notes from an ongoing experiment in building small, transparent tools for teaching finance and probing how AI language models actually work. Some entries are long essays; others are short notes from work in progress.

The current series, Understanding LLMs to use them better in management and finance, opens up a Transformer language model from three angles: the attention machinery, the embedding space, and the final next-word step.

Transformers and QKV Attention: A Primer

LLMs

Transformers

Attention

Teaching

How a Transformer moves information between words — queries, keys, values, the residual stream, and why attention is the only channel where tokens talk to each other. Part 1 of 3.

May 31, 2026

75 min

Embeddings and the Maps We Draw of Them

LLMs

Embeddings

Visualization

Interpretability

Teaching

Where a language model’s vectors come from, why hand-wiring meaning fails, and how to look at a high-dimensional space — PCA, concept directions, t-SNE — without fooling yourself. Part 2 of 3.

Jun 1, 2026

36 min

The Final Step: The Language Model as a Relentless Seeker of the Best Next Word

LLMs

Transformers

Teaching

The last layer turns all the model’s work into one next word — a similarity search over the vocabulary. Confidence, sampling, hallucination and tools all fall out of one dot product. Part 3 of 3.

Jun 2, 2026

21 min

Categories

Transformers and QKV Attention: A Primer

Embeddings and the Maps We Draw of Them

The Final Step: The Language Model as a Relentless Seeker of the Best Next Word