Module 01 · AI Fundamentals Lab
What does a model actually see when you type?
Not words, and definitely not letters. Your text is cut into chunks called tokens before the model sees anything. Type here, watch where the cuts land, and you will understand why models miscount letters, fumble long numbers, and charge you more for Tamil than for English.
Type anything. Watch it get chopped.
Each coloured box below is one token — one chunk the model handles as a single thing. The colours just mark where one chunk ends and the next begins; they carry no other meaning.
These counts are estimates. This page splits text with a small approximation of byte-pair encoding built into the page itself. It has no real vocabulary file and never contacts a model, so exact counts differ from model to model. The behaviour it reproduces — where the cuts tend to land — is what matters here.
1. Counting letters
Ask a model how many rs are in “strawberry” and it often says two. Here is the input it was working from.
So how is it supposed to count?
It cannot look at the letters, because they are not separately there. It has to answer from what it learned about these chunks during training — which is recall, not counting. That is why the answer is confident, fast, and sometimes wrong.
2. Long numbers
Digits get grouped in short runs. Type a number and watch the groups — then notice what happens when the number gets one digit longer.
As the model sees it
The same digits with one more in front
Why arithmetic goes wrong
Nothing tells the model which chunk holds the hundreds and which holds the thousands, and every boundary moves when the number changes length. Adding two long numbers means lining up columns it cannot see. It usually gets the shape of the answer right and a digit in the middle wrong — which is the worst kind of wrong, because it looks right.
3. Cost and language
The same sentence — “Never share your password with anyone” — written four ways. Estimated tokens for each.
API bills and context limits are counted in tokens, not characters. A sentence in a non-Latin script can cost several times more than the same sentence in English — more money per request, and less room left in the context window for anything else.
Estimates from this page's approximate tokenizer. Real tokenizers handle non-Latin scripts better than this one does, but the gap itself is real and this is the direction it runs in.
Every token count in these experiments is an estimate produced in your browser by this page's approximate tokenizer. No model was asked anything.
A model never sees your words or your letters — it sees tokens, and everything it is bad at is downstream of where those cuts land.
Counting letters, long arithmetic, and the price of a sentence in Tamil are all the same fact wearing different clothes.
My observations
Kept in this browser tab only, for this session.