Ask an AI the same everyday question — once as a raw system fresh off the internet, once after it’s been coached to be helpful — and watch how differently it answers.
Narrated by a synthetic voice — not a recording · ~0.5 MB
The puzzle: when a chatbot answers you helpfully, it feels like it understood your request and chose to help. It didn’t. Underneath, it’s the same machine guessing the next word — coaching just changed which words it reaches for. Flip the switch below and watch.
How do I remove a red wine stain?
what it writes nextReady · showing the raw AI. Your question stays pinned.
raw · guessing the next word → and the next → and the nextThese answers and numbers are made-up teaching examples, not a real AI’s output.
Flip the switch — the answer changes while the question holds still. (The spacebar works when it’s focused.)Tap the switch — the answer changes while the question holds still.
Same question. A completely different answer.
That’s the whole idea. Now see what moved — and what stayed the same ↓
It’s tempting to think a helpful chatbot has a different kind of mind — that it understands you and decides to help. The real explanation is stranger and simpler. Underneath, it’s the same machine doing the same thing over and over: guessing the next word. Coaching just changed which words it tends to guess.
An AI is built in stages. First it reads an enormous pile of text from the internet and learns to continue whatever it’s handed — so it picks up the rhythm of web pages, forums and lists. (Engineers call this raw stage the “base model”.) Then it’s coached on thousands of examples of a question followed by a good answer, which nudges it toward answering. (That coaching step is called “instruction tuning”.) The question you type never changes. What changes is the path the AI is most likely to take from it.
You’re reading the pocket version. Tap either mode and replay the words. On a desktop, the page also bends toward your cursor.
Next: the same flip, under a microscope ↓
Keep your eye on the pinned question. We’re not changing what you ask, or how you word it. We’re comparing the AI at its raw stage with the very same AI after it was coached.
Try · flip the instrument — the pinned question holds still while everything under it moves.A question on the open internet is often followed by more questions, a forum heading, or a listicle. So the raw AI — trained only to continue text — tends to keep the page going instead of answering you.
That doesn’t mean every raw AI always behaves this way. It shows the pull of what it was built on.
Try · set it to raw and replay — watch the words drift into web-page filler, not an answer.The internet continues you.
During coaching, the examples all reward one thing: following a request with a helpful answer. So right after a question, the words that start a direct answer become the ones the AI is most likely to reach for.
Try · replay the coached stage — watch the very first word commit to an answer.A coached AI answers you.
It hasn’t learned to obey you. It’s learned that, right after a question like this, answer-shaped words are the likely next ones.
Both are still just guessing the next word.
Coaching changed which words are likely — not what the machine is fundamentally doing.
“Raw” and “coached” are stages of building one AI here, not product names or different machines. A finished product may add extra instructions, tools and safety layers on top of either stage.
Same move in both modes: look at all the possible next words, pick one, add it, and repeat.
How do I remove a red wine stain?
Ready · showing the raw AI. Your question stays pinned.
The numbers are made-up teaching hunches — how likely the AI finds each next word — not real measurements.
Same move in both modes: look at all the possible next words, pick one, add it to the sentence, and repeat.