hawkes.aiKey ideas
- The one idea — The model predicts the next chunk of text; it does not look answers up.
- How it is shown — "The capital of Australia is ___" — live probability bars; then the same bars forced to answer an unknowable question.
- The trap to avoid — "Are you sure?" doesn't check anything — it re-rolls the guess.
- What it sets up — The truthful sentence and the likely sentence are different horses.
Ask an AI a question it cannot possibly know. It answers anyway, fluently. Inside there's no drawer to come back empty — some guess is always tallest.
The one idea
The model predicts the next chunk of text; it does not look answers up.
Everyone thinks it knows things. It doesn't. It guesses. This is a search engine: your question goes in, a stored answer comes back out. Retrieval. A fact, fetched from a shelf. This is a language model. Nothing is fetched, because nothing is stored as sentences. It plays one game, over and over: given every word so far — what word comes next? "The capital of Australia is…" — the model scores every word it could say next. Canberra: very likely. Sydney: less. Banana: basically zero. It rolls dice weighted by those scores, and commits.
How it works — the demo
"The capital of Australia is ___" — live probability bars; then the same bars forced to answer an unknowable question.
The new word joins the sentence, and it plays again. And again. Every reply you have ever read was built like this — one weighted guess at a time. So why is it so often right? Because predicting text written by people who knew things eventually requires the same patterns. Grammar. Geography. Code. The shape of true sentences gets pressed into the machine. Now the crack. Ask it something it can't know — your sister's name. There's no drawer to come back empty, no "nothing" to return. Every bar is low… but one is still tallest. It answers anyway.
The trap to avoid
"Are you sure?" doesn't check anything — it re-rolls the guess.
Why it matters — and what’s next
The truthful sentence and the likely sentence are different horses.
And here's what trips everyone: the sentence sounds the same whether the winning guess scored ninety-nine percent or nine. Fluency is free. Certainty isn't. The trap: asking "are you sure?" doesn't check a database. It spins the machine again — sometimes onto a different answer. That's not lying, and it's not memory failing. It's a guesser doing exactly what it was built to do. "Just guessing" sounds like an insult. It's the design. It's why this thing writes poems no database contains. Everything in this series — training, hallucination, agents, safety — is engineering wrapped around one question: how do you shape, steer, and check a guesser? Because later you'll meet the darker version of today's idea: sometimes the truthful sentence and the likely sentence are different horses — and this machine always bets on likely. Remember that.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Everyone thinks it knows things. It doesn't. It guesses.
This is a search engine: your question goes in, a stored answer comes back out. Retrieval. A fact, fetched from a shelf.
This is a language model. Nothing is fetched, because nothing is stored as sentences. It plays one game, over and over: given every word so far — what word comes next?
"The capital of Australia is…" — the model scores every word it could say next. Canberra: very likely. Sydney: less. Banana: basically zero. It rolls dice weighted by those scores, and commits.
The new word joins the sentence, and it plays again. And again. Every reply you have ever read was built like this — one weighted guess at a time.
So why is it so often right? Because predicting text written by people who knew things eventually requires the same patterns. Grammar. Geography. Code. The shape of true sentences gets pressed into the machine.
Now the crack. Ask it something it can't know — your sister's name. There's no drawer to come back empty, no "nothing" to return. Every bar is low… but one is still tallest. It answers anyway.
And here's what trips everyone: the sentence sounds the same whether the winning guess scored ninety-nine percent or nine. Fluency is free. Certainty isn't.
The trap: asking "are you sure?" doesn't check a database. It spins the machine again — sometimes onto a different answer. That's not lying, and it's not memory failing. It's a guesser doing exactly what it was built to do.
"Just guessing" sounds like an insult. It's the design. It's why this thing writes poems no database contains. Everything in this series — training, hallucination, agents, safety — is engineering wrapped around one question: how do you shape, steer, and check a guesser?
Because later you'll meet the darker version of today's idea: sometimes the truthful sentence and the likely sentence are different horses — and this machine always bets on likely. Remember that.