Lesson 02 · 4 min · 6 things to do
Everything becomes numbers
Predict what a model can and cannot see in your words.
A model has no letters or words inside it — only numbers. What has to happen to your sentence first?
- Yes.Those chunks are called tokens. Common words are one token; unusual ones get split into pieces, which is why odd names and long words behave strangely.
- Not quite.That was tried and it is wasteful — sequences get long and patterns get harder to spot. Chunks sit between letters and words.
- Not quite.No grammar is parsed. The chopping is statistical: whatever chunks appeared often in the training text.
The model does not see your words. It sees a list of chunk numbers — which is why counting letters is a strange, indirect task for it.
Which of these are likely to be ONE chunk, and which get split into several?
the
antidisestablishmentarianism
water
Kyzylorda
a rare 14-digit account number
Yes.Common words earned their own chunk because they appear constantly. Rare names, long compounds and digit strings are assembled from fragments — and a model works with fragments less reliably than with whole units.Why do language models so often miscount the letters in a word?
- Yes.It is asking a question about a level below the one the model operates on. The model can often reason its way there, which is not the same as seeing it.
- Not quite.Counting to seven is not the hard part. Knowing what is there to count is.
- Not quite.It had a great many. The difficulty is representational, not a gap in the material.
Words with related meanings end up with nearby numbers. Slide to see how far apart some pairs sit.
king / queenking / castleking / cabbageHow far apart in meaning-space0.2The furthest anything gets1king / queenVery close. They appear in almost identical sentences, so they learn almost identical positions.
How far apart in meaning-space0.55The furthest anything gets1king / castleRelated but not interchangeable. They appear near each other, rarely in the same slot.
How far apart in meaning-space0.95The furthest anything gets1king / cabbageFar apart. Almost nothing in the training text puts them in the same role.
Nobody told the model that a king and a queen are similar. Where did that come from?
- Yes.Meaning here is built entirely out of company kept. It is a genuinely powerful trick, and it is also why the model absorbs whatever associations the text carried.
- Not quite.Dictionaries are in there, and they are a vanishing fraction. The positions come from ordinary usage across everything.
- Not quite.Labelling is used for other stages of training. This structure appears on its own, from raw text.
Move the control to see what changes.
A model's context window holds 8,000 chunks, and one page of prose is about 500. Roughly how many pages can it hold at once?
pagesYes.8,000 ÷ 500. Everything outside that window is gone — not forgotten in a human way, simply never in front of it.You paste a 300-page document into a model with a 16-page window. What happens?
- Yes.Tools work around this by splitting the document and fetching the relevant pieces. The model itself never sees all 300 pages together, and an answer that implies it did is worth doubting.
- Not quite.There is no place to keep them. Anything outside the window has no effect on the next word.
- Not quite.Often it does not — it answers using whatever fits, which is exactly what makes this failure quiet.
Lesson complete
It never sees your words — only chunks, turned into numbers.
Next: Learning is error, made smaller →