hawkes.aiKey ideas
- The one idea — Tokenization is why models miscount letters — they can't see inside a token.
- How it is shown — Spelling "s t r a w b e r r y" with spaces gives every letter its own token — and the count is suddenly perfect.
- The trap to avoid — "It can't even count letters" reads blindness as stupidity — judge a system by what it can see.
- What it sets up — Every token is an ID number.
The strawberry fail isn't stupidity. The model never receives letters at all — 'strawberry' arrives as sealed capsules, and an ID number contains no R.
The one idea
Tokenization is why models miscount letters — they can't see inside a token.
Ask it to count the R's in "strawberry." Watch a machine that writes symphonies fail a kindergarten quiz. You see ten letters. S-T-R-A-W-B-E-R-R-Y. Three R's. Easy — because you can see them. The model never sees letters. Before anything reaches it, your text is chopped into tokens — chunks of characters — and each chunk is swapped for its ID number. "Strawberry" stops being letters and becomes two or three sealed capsules. What the model actually receives is just the numbers. Not S. Not R. The letters are sealed inside — and the capsule never opens. A token's ID no more contains an R than your passport number contains your face. Why chop at all?
How it works — the demo
Spelling "s t r a w b e r r y" with spaces gives every letter its own token — and the count is suddenly perfect.
Compression. The tokenizer was built by scanning mountains of text and gluing the most frequent character-pairs together, over and over. Common words fuse into single chunks. Rare ones shatter. A brilliant trick for reading fast — and blind by design. So "count the R's" asks about objects the model never received. It can't look. It can only do what it always does — guess the likeliest answer. And for years, the internet's likeliest answer to trick questions was wrong. Same blindness, different costumes: reversing a word. Rhyming precisely. Arithmetic on long numbers — which get chopped into odd digit-chunks. If the task needs to see inside the capsule, expect stumbles. Want proof it's the capsules, not stupidity?
The trap to avoid
"It can't even count letters" reads blindness as stupidity — judge a system by what it can see.
Why it matters — and what’s next
Every token is an ID number.
Put spaces between the letters. Each letter becomes its own token — the model can finally see them — and suddenly it counts perfectly. Same model. Same question. Different eyes. The trap: "it can't even count letters, it's dumb." Wrong lesson. It's a blindfolded reader quizzed on font colors. Newer models often pass — some by tools, some by brute training — but the blindness underneath hasn't gone anywhere. Judge a system by what it can see, or you'll misjudge everything it does. One more thing. Every capsule has an ID — a number in a fixed vocabulary. Which means somewhere in that vocabulary, there are numbers that almost never appeared in training. Empty rooms. Three episodes on: the haunted tokens that break models — entirely by accident.
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
Ask it to count the R's in "strawberry." Watch a machine that writes symphonies fail a kindergarten quiz.
You see ten letters. S-T-R-A-W-B-E-R-R-Y. Three R's. Easy — because you can see them.
The model never sees letters. Before anything reaches it, your text is chopped into tokens — chunks of characters — and each chunk is swapped for its ID number. "Strawberry" stops being letters and becomes two or three sealed capsules.
What the model actually receives is just the numbers. Not S. Not R. The letters are sealed inside — and the capsule never opens. A token's ID no more contains an R than your passport number contains your face.
Why chop at all? Compression. The tokenizer was built by scanning mountains of text and gluing the most frequent character-pairs together, over and over. Common words fuse into single chunks. Rare ones shatter. A brilliant trick for reading fast — and blind by design.
So "count the R's" asks about objects the model never received. It can't look. It can only do what it always does — guess the likeliest answer. And for years, the internet's likeliest answer to trick questions was wrong.
Same blindness, different costumes: reversing a word. Rhyming precisely. Arithmetic on long numbers — which get chopped into odd digit-chunks. If the task needs to see inside the capsule, expect stumbles.
Want proof it's the capsules, not stupidity? Put spaces between the letters. Each letter becomes its own token — the model can finally see them — and suddenly it counts perfectly. Same model. Same question. Different eyes.
The trap: "it can't even count letters, it's dumb." Wrong lesson. It's a blindfolded reader quizzed on font colors. Newer models often pass — some by tools, some by brute training — but the blindness underneath hasn't gone anywhere. Judge a system by what it can see, or you'll misjudge everything it does.
One more thing. Every capsule has an ID — a number in a fixed vocabulary. Which means somewhere in that vocabulary, there are numbers that almost never appeared in training. Empty rooms. Three episodes on: the haunted tokens that break models — entirely by accident.