A ten-minute case for three small words
Your AI isn't reading your mind.
It's placing a bet.
The thing that keeps happening
Made him a cup of tea.
Drove him to A&E.
Fed his cat while he's away.
Buried him under the patio.
It didn't misunderstand you. It understood you four ways — and picked one.
What is actually happening in there
That's the whole trick. There is no plan, no draft, no moment where it decides what the finished answer will be. Each piece is chosen with only the pieces before it for company.
Autocomplete that went to university.
Watch it think · one chunk at a time
Illustrative figures. A real model ranks its entire vocabulary — tens of thousands of options — at every step; these are the top four.
A small correction
The pieces are called tokens — roughly three quarters of a word each. It reads and writes in bites, not words. Nothing else about the argument changes.
Remember 34 vs 31
It doesn't hedge. It commits — and then every chunk after that defends the choice.
Why it never just says "I'm not sure"
One confident sentence.
One confident sentence.
Same fluency, same tone, same certainty. The doubt exists for a fraction of a second inside the numbers and then it is gone — spent, not reported. A fluent answer to the wrong question looks exactly like a fluent answer to the right one.
Same prompt · one difference
Cost of the guess: 15 to 20 years, and you find out on Tuesday.
Why three words move the needle
It isn't detecting confusion and deciding to report it. It's doing as it's told — and two separate forces make that land hard.
Following an explicit instruction is the behaviour it was trained hardest to produce. Those three words spend that.
Right after a request, producing the goods is what gets rewarded — asking is rare. The instruction lifts that brake.
It spots "unclear" by pattern, not by reading its own numbers. Useful — not reliable.
When to spend the twelve seconds
Rule of thumb: the more it costs to be wrong, the more the question is worth.
Steal this
Ask if unclear.
Paste it at the end of any request that matters. It costs you one line and it buys you the question the machine would otherwise have answered on your behalf.