A language model can’t read letters.
It never sees them. Here is what it actually gets — free, and this is lesson one.
Every piece becomes a number. The model only ever sees the numbers — which is why it can write you a sonnet but can’t reliably count the letters in “strawberry”. It never saw the letters.
Real tokenisers are trained on text and split differently — this is a faithful demonstration of the idea, not a copy of GPT’s vocabulary. The lesson in the app uses the real thing.
The rest of it
Thirty-odd more of these, and counting.
Cat.AI teaches how AI is actually built — one idea per screen, about seven minutes a lesson. Tokenisation, embeddings, attention, why models are confidently wrong. No maths degree assumed and none required.
It’s being written now, in the open. Leave an email and I’ll send you the first build — and nothing else.
One email when it’s ready. No list, no newsletter, unsubscribe is replying “no”.