The part everyone has heard of, and the part most often misunderstood. Titles are on the spines; the notes underneath are the three tellings, in short — for a ten-year-old, for the architect, and for the boardroom.
6 concepts on this shelf · Beginner → Advanced · about 2 min read
Kitchen table — It has read an enormous amount and got very good at guessing the next word. Do that again and again and you get sentences.
Whiteboard — Transformers, attention, pretraining then alignment. Next-token prediction at scale turns out to buy reasoning-shaped behaviour — and confident nonsense at the edges.
Boardroom — Extraordinary at language work, unreliable as a source of fact. Treat it as a brilliant drafting colleague, not a database.
large language model, LLM, transformer, attention, next token prediction, pretraining
Kitchen table — Words get chopped into pieces, and each piece becomes a list of numbers. Pieces that mean similar things end up close together.
Whiteboard — Subword tokenisation drives your context budget and your bill; embeddings give you similarity search, clustering and retrieval for almost free.
Boardroom — Tokens are the unit you are billed in. Embeddings are what make 'search that understands meaning' possible over your own documents.
tokens, tokenisation, embeddings, vector, semantic search, similarity
Kitchen table — How you ask changes what you get. Give it the right facts to look at and it does far better.
Whiteboard — Instructions, examples, retrieval, tool results and history all compete for one finite window. Ordering, compaction and what you leave out are design decisions.
Boardroom — Most disappointing AI results are not a model problem. They are a context problem — the system was never shown what it needed to know.
prompt engineering, context engineering, context window, few-shot, system prompt, compaction
Kitchen table — Before it answers, it goes and finds the right page in your own books, and answers from that.
Whiteboard — Chunking, embedding, hybrid search, reranking and citation. Retrieval quality sets the ceiling; no amount of prompt polish rescues a bad top-k.
Boardroom — How you make a general model answer from your policies, your contracts, your data — with a source you can check.
RAG, retrieval augmented generation, vector database, chunking, reranking, hybrid search, citations
Kitchen table — The model already knows how to talk. This teaches it to talk the way your company talks.
Whiteboard — LoRA and friends: small trainable deltas over frozen weights. Excellent for form, tone and format; a poor and expensive way to install facts.
Boardroom — Use it for style and consistency. If the goal is 'know our data', retrieval is usually cheaper, faster and easier to keep current.
fine-tuning, LoRA, adapters, PEFT, instruction tuning, domain adaptation
Kitchen table — Older computers sorted things that already existed. These ones make new things — letters, pictures, summaries.
Whiteboard — The interesting engineering is rarely the generation. It is the review path, the fallbacks, the audit trail and knowing which tasks should never be generated at all.
Boardroom — Biggest wins are in drafting, summarising and triage — high-volume work where a human still signs off. Pick those first.
generative AI, GenAI, content generation, summarisation, drafting, copilots