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 7 min read
Kitchen table — It has read more than any person ever could — most of the internet, whole libraries — and got astonishingly good at one game: guess the next word. 'The cat sat on the...' — you know it is 'mat'. Play that game billions of times with harder and harder sentences and something surprising happens: to guess well, you end up having to know things. But remember it is still guessing. It sounds just as confident when it is wrong, like a friend who never, ever says 'I don't know'.
Whiteboard — Transformers and attention, pretrained on next-token prediction at a scale that turns a party trick into reasoning-shaped behaviour, then aligned with human feedback so the raw predictor becomes an assistant. The capabilities are real and so are the edges: confident interpolation where training data was thin, no native sense of its own uncertainty, and knowledge frozen at the training cut-off. Treat outputs as drafts from a brilliant, occasionally hallucinating colleague — verification is part of the architecture, not an afterthought.
Boardroom — Extraordinary at language work — drafting, summarising, translating, triaging — and unreliable as a source of fact, a combination that decides where it belongs in your business: anywhere a fluent draft saves an expert an hour, nowhere a fabricated detail costs you a customer or a lawsuit unreviewed. The capability is rented from a handful of providers, so concentration risk and pricing power belong on the risk register next to the productivity upside. It is a brilliant drafting colleague, not a database.
large language model, LLM, transformer, attention, next token prediction, pretraining
Kitchen table — The machine does not read words the way you do. First the words get chopped into pieces — 'unbelievable' might become 'un', 'believ', 'able' — and each piece becomes a long list of numbers. Here is the lovely part: pieces that mean similar things get numbers that are close together, like houses on the same street. 'King' and 'queen' live near each other; 'king' and 'kettle' do not. Meaning becomes a map, and the machine finds its way around by distance.
Whiteboard — Subword tokenisation sets your context budget and your bill — token counts, not word counts, are the unit of everything. Embeddings are the workhorse nobody demos: dense vectors where cosine distance approximates semantic similarity, giving you search, clustering, deduplication and recommendation almost for free. Chunking strategy and embedding-model choice quietly decide your retrieval quality, and mixing embeddings from different models is the classic silent failure — the maps do not share a coordinate system.
Boardroom — Two practical facts hide in this book. Tokens are the unit you are billed in — every price on an AI invoice is per-token, so verbose prompts and long documents are literally expensive, and a cost review of prompts is a real savings line. Embeddings are what make 'search that understands meaning' possible over your own documents — the unassuming technology behind most successful enterprise AI so far, and considerably cheaper than the chatbots that get the headlines.
tokens, tokenisation, embeddings, vector, semantic search, similarity
Kitchen table — How you ask changes what you get — same as with people. 'Help me with my homework' gets you a shrug; 'here is the question, here is what I tried, here is where I got stuck' gets you real help. The machine also has a limited desk: it can only look at so many pages at once, so choosing what to put on the desk is most of the game. Give it the right facts and it is brilliant. Give it nothing and it guesses — confidently.
Whiteboard — Instructions, examples, retrieved documents, tool results and conversation history all compete for one finite context window, and the model attends unevenly across it — order matters, the middle gets lost, and everything you leave out is a fact the model does not have. Compaction, summarisation and retrieval-on-demand are design decisions, not conveniences. 'Prompt engineering' matured into context engineering: less magic words, more information architecture for a reader with no memory of the last request.
Boardroom — Most disappointing AI results are not a model problem — they are a context problem: the system was never shown the policy, the price list or the customer's history it needed. That reframing changes where the money goes: away from waiting for a better model, toward the unglamorous work of connecting your own information so it arrives at the right moment. When a pilot underwhelms, ask 'what did the model actually see?' before asking 'do we need a bigger model?' The first question is usually the answer.
prompt engineering, context engineering, context window, few-shot, system prompt, compaction
Kitchen table — Instead of answering from memory, the machine is allowed to look things up first — like an open-book exam. Ask about your school's rules and it does not guess; it goes to the shelf, finds the right page in the rulebook, reads it, and answers from what it just read — and it can show you the page. Much harder to make things up when the page is right there. The trick, of course, is finding the right page.
Whiteboard — Chunk the documents, embed them, retrieve by hybrid search (dense vectors plus keyword — dense alone misses part numbers and names), rerank, then generate with citations. Retrieval quality sets the ceiling: no prompt polish rescues a bad top-k, so evaluate retrieval separately from generation or you will tune the wrong half. The unfashionable determinants of success are document hygiene, chunk boundaries that respect structure, and freshness pipelines — librarianship, wearing a vector database.
Boardroom — This is how a general model answers from your policies, your contracts, your data — without the cost of training anything, and with a source you can check attached to every answer. That auditability is why it became the default enterprise pattern; when the answer is wrong you can see which document misled it, which is a governance property no raw model offers. The realistic budget note: the hard cost is not the AI, it is getting your document estate clean and current enough to be worth retrieving from.
RAG, retrieval augmented generation, vector database, chunking, reranking, hybrid search, citations
Kitchen table — The machine already speaks — years of reading taught it that. Fine-tuning is like teaching it your family's way of speaking: your jokes, your rules, how you sign your letters. You do not send it back to school; you give it a short course of examples until it picks up the style. But if you want it to know today's facts, a short course is the wrong tool — you would hand it the facts to read instead. Style is taught; facts are looked up.
Whiteboard — LoRA and the PEFT family: small trainable deltas over frozen weights, so you tune millions of parameters instead of billions and swap adapters per task. Excellent for form, tone, format compliance and domain vocabulary; a poor and expensive way to install facts, which go stale and are better retrieved at inference. The classic failure is fine-tuning on data you could have put in context — spending training-money on a retrieval problem. Evaluate against the base model plus good prompting first; the delta is often smaller than the meeting that approved it.
Boardroom — Use it when the model must reliably sound like you or follow your formats — brand voice, document conventions, regulated phrasing. If the goal is 'know our data', retrieval is usually cheaper, faster to update and easier to audit; fine-tuning bakes knowledge in like print, retrieval keeps it loose-leaf. The questions that sort most proposals: how often does this knowledge change, and can we produce the thousands of good examples tuning needs? Style changes rarely; facts change weekly. Match the tool to the shelf-life.
fine-tuning, LoRA, adapters, PEFT, instruction tuning, domain adaptation
Kitchen table — Older computer programs could only sort what already existed — this pile is spam, that photo is a cat. The new ones can make new things: a letter that never existed, a picture nobody drew, a summary of a meeting in ten lines. It is like the difference between a librarian who finds books and a friend who writes you one. Wonderful — as long as somebody checks the work, because making things up is precisely what it is built to do.
Whiteboard — The generation call is the easy tenth of the system. The engineering is the review path (who approves, at what confidence), the fallbacks when output fails validation, the audit trail of what was generated from which inputs by which model version, and the honest list of tasks that should never be generated at all. Structured outputs with schema validation beat free text everywhere the result feeds another system. Design for the failure modes first — the demo already works; the demo always works.
Boardroom — The dependable wins are in drafting, summarising and triage — high-volume language work where a human still signs off and the AI removes the blank page, reliably clawing back thirty to fifty percent of the time. Full automation of customer-facing output is a different risk class entirely: every generated sentence carries your letterhead, and 'the AI wrote it' is not a defence any regulator accepts. Sequence accordingly — assistance first, autonomy only where the error cost is genuinely low and measured.
generative AI, GenAI, content generation, summarisation, drafting, copilots