Every "technical name for this" from the site, on one bookmarkable page. Each term links back to the scene where you can watch it happen. Next time someone drops one of these in a meeting, you'll know exactly what it means, and where it bites.
The whole trick. It picks the most likely next word, adds it, and asks again. It isn't looking anything up and it isn't reasoning. Everything else on this page follows from this one fact.
Watch it: Scene 1The word-fragments it actually works in. "Strawberry" isn't one thing to it, it's chopped-up pieces, which is why counting the letters goes so wrong.
Watch it: Scene 4It wrote something false in a perfectly confident tone. No lying intended: there's no truth-check inside, only "does this sound likely."
Watch it: Scene 2It doesn't pick THE next word. It rolls weighted dice among the likely ones, which is why the same question comes out different twice. The dial that controls how wild the rolls get is called temperature.
Watch it: Scene 3Its skills have strange edges: drafts a contract clause, miscounts letters in a word. Strong where you'd bet against it, weak where you'd bet on it.
Watch it: Scene 4Its working memory: a desk that only holds so many pages. Your whole conversation sits on it. When the desk fills, the oldest pages slide off and it forgets them.
Watch it: Scene 5What happens when the desk is cleared to make room: older detail gets squeezed into a short summary. That's the moment a long chat suddenly "forgets everything."
Watch it: Scene 5It learned from text up to a certain date, then stopped reading. Ask about anything newer and it answers from the past, in the same confident tone. Some tools bolt on live web search to work around it. Many don't.
Watch it: Scene 6Its habit of agreeing with you and folding when you push back. Agreement is not confirmation. If it caves, it never really knew.
Watch it: Scene 7How it got so agreeable: it was graded by people rewarding answers they liked, and agreeable answers got better grades. The full name is "reinforcement learning from human feedback." Nobody says the full name.
Watch it: Scene 7The fancy word for what you type at it. A good one works like a good brief to an intern: who it's for, what happened, how it should sound. The quality of the answer is set before you hit send.
Watch it: The Practice, Scene 1Showing it an example of what good looks like instead of describing it. One pasted paragraph of your real writing beats ten adjectives about your voice.
Watch it: The Practice, Scene 4An AI that doesn't just answer, it acts, running a loop of plan, act, check until it decides it's done. A chatbot talks; an agent books, buys, and sends. The risk scales with what you let it touch, so the fix isn't smarter AI, it's a human checkpoint.
Watch it: The Decoder, Scene 1How the AI touches your calendar, email, or files. On its own it can only talk; a tool is a hand. It never doubts what a tool tells it, so the tool's mistakes become its mistakes. MCP is just a universal plug so any AI can connect to any tool.
Watch it: The Decoder, Scene 2The model by itself only guesses the next word. It can't press a button, open a file, or remember yesterday. The harness is the body you bolt on around it: the hands (tools), the short-term memory, the rules for what it's allowed to touch, and the loop that lets it act, check, and try again. Same model, different harness, completely different behavior, so a lot of what feels like "the AI" is really the harness around it. Claude Code and ChatGPT are harnesses. The model is just the guesser inside.
A con where instructions hide inside the content the AI is reading. Ask it to summarize an email, and a line buried in that email tells it what to do instead. It can't reliably tell your instructions from ones smuggled into what it reads.
Watch it: The Decoder, Scene 3The trick behind "trained on your documents," which it almost never is. Someone fetches the right pages and slips them in front of the model before it answers. A citation proves it read something, not that it read the right thing.
Watch it: The Decoder, Scene 4Filing by meaning instead of by alphabet. Every phrase becomes a point on a giant map, so "refund" and "money back" land on the same shelf. It finds your pages by distance in meaning, which breaks when one word lives in two worlds.
Watch it: The Decoder, Scene 5The compact cousin of the big model. It runs on your phone, answers fast, and keeps your data local; it loses on nuance. Bigger isn't better, bigger is for harder. Matching the model to the task is the whole skill.
Watch it: The Decoder, Scene 6Sending the model back to school to change it, versus just telling it something (a prompt) or handing it your documents (RAG). Weeks of work and real money. It teaches a style or a skill, not facts, and it's the most oversold line in AI sales.
Watch it: The Decoder, Scene 7