Nib asked for
What Else Tokens Explain
Because a model works in chunks rather than letters or digits, the same blind spot that hides the r's in strawberry also shows up in reversing words, rhyming, long arithmetic, and how usage is priced.
4 steps, drawn by Nib. Read it here, or watch him draw it. Updated .
1reversing words
Reversing a word is the strawberry problem in disguise. To flip 'strawberry' backward, the model would need its letters in order, but it received a couple of chunk-numbers, not letters. Without spelling it out first, it has to guess a reversal from weak memory, and that is where garbled results come from.
Reversing a word needs the letters, and the letters are hidden inside chunks.
a common mix-upThe model fails at reversing because it is careless or lazy; in fact it never saw the letters it is being asked to rearrange.
Which trick from page 5 would likely help a model reverse a word more reliably?
Show the answer
Have it write the word out letter by letter first, so each letter becomes a visible token it can then reorder.
2rhyming
Rhyming depends on sound, and sound lives in the ending of a word, often a few letters. A chunk is just a number with no built-in spelling or pronunciation. The model learns rhymes only indirectly, from poems and lyrics in its training text, so it can rhyme common words well but stumble on unusual ones.
Rhyme is about word endings, which chunk-numbers don't directly expose.
a common mix-upThe model hears words the way we do; it only ever handles numbered text chunks, not sounds.
3long arithmetic
Long numbers get chopped into chunks too, and not tidily. A number like 48213 might split into '482' and '13', while a similar number splits differently. Arithmetic by hand relies on lining up digits in columns, and the model can't count on the digits being lined up when the chunk boundaries keep shifting.
Uneven chunking of digits makes column-style arithmetic hard.
a common mix-upA language model calculates like a calculator; it predicts likely text, which is why long numbers trip it up.
Why might writing out a long sum step by step help a model?
Show the answer
Each small step uses short numbers and visible digits, so it checks pieces instead of recalling a whole answer.
4pricing
Tokens also explain the bill. Services charge by tokens because tokens are the unit the model actually processes, and each one costs computing effort. A word is not a fixed size: a short common word may be one token, while a long or rare one may be several. So the same page of text can cost different amounts in different languages.
Usage is priced in tokens because tokens are what the model actually processes.
a common mix-upServices charge per word; words vary too much in size, so they charge per token.
Once you picture the chunks, the strawberry puzzle stops being a one-off: reversing, rhyming, long sums, and the price tag all come from the same hidden pieces.