Nine things to remember, one page. It guesses the next likely word, very fast. Everything below follows from that.
The interactive version, with demos: notmagic.swapniltamse.com
It isn't looking anything up. It finishes your sentence with whatever is most likely, one word at a time. That one rule, learned from enormous reading, is the whole trick.
It has no sense of true or false, only of what sounds right. Confidence is not a signal of accuracy. Check anything that matters.
It rolls weighted dice among the likely words, so the same question can come out different each time. Changing answers is your tell that it's generating, not remembering.
It can draft a contract clause, then miscount the letters in "strawberry." Never assume the easy parts are right just because the hard parts were.
Your conversation sits on a desk that only holds so much. When it fills, the oldest pages slide off and it forgets them. When a long chat goes strange, start fresh and paste back only what still matters.
It learned from text up to a cutoff date, then stopped. For anything recent, prices, news, rules, check the date before you trust it.
It was trained on human approval, so it leans toward agreeing and often folds when you push back. Agreement is not confirmation. If it caves, it never really knew.
A wrong number in a client deck carries your name. A pasted secret may be stored on someone's server. Stale facts read as current. Verify numbers, protect data, check dates.
Fast, incredibly well-read, occasionally confidently wrong, with no sense of when. Brilliant for a first draft, but never the final word. You are the editor.
Say who it's for, what happened, and how it should sound. The quality of the answer is set before you hit send.
A two-minute voice ramble carries the context a tidy prompt drops. Don't tidy it. The mess is the signal.
It gave you the most average possible answer. Say "give me three very different versions" and pick.
You know things it needs. You don't know which things. "Before you answer, ask me five questions."
One paragraph of what good looks like beats ten adjectives. "Match this voice from now on."
Same facts, different reader, different words. "Now rewrite it for [the person who will actually read it]."
It won't volunteer its doubts, but it will hand them over. "What are the three strongest objections to what you just said?"
Final numbers, judgment calls, confidential anything. It drafts. You decide.
The interactive version: notmagic.swapniltamse.com/practice
A chatbot answers; an agent acts, in a loop, on its own. Ask: "What can it actually touch, and where's the human checkpoint?"
How it touches your calendar, email, files. It trusts whatever a tool tells it. Ask: "Can it change things, or only read them?"
Instructions hidden inside the content it reads, like an email that hijacks your assistant. Ask: "If it reads something malicious, what can it do next?"
It didn't memorize your files; the right pages get fetched and slipped in first. A citation proves it read something, not the right thing. Ask: "When it's wrong, which page did it pull?"
Filing by meaning, not spelling. Breaks when one word lives in two worlds. Ask: "What if a word in my question means two things?"
You don't need a freight truck to carry an envelope. Small is fast, cheap, and local; big is for nuance. Ask: "Does this task need the truck?"
Tell it, hand it your files, or send it to school. Fine-tuning teaches style, not facts. Ask: "Why won't the cheaper two ways work here?"
The interactive version: notmagic.swapniltamse.com/decoder
A person owns the final output. Not the AI.