In the previous article we left off with the backpack measured in tokens, and every token you pay for again at every turn. Two words that look scary on the bill — and almost nobody knows what they actually are.
Token.
Everyone thinks of a token as a line item on the bill, something to keep an eye on when you’re paying for AI. True, but that’s the small part. A token is first and foremost the way AI reads. And it doesn’t read like you.
You read letters, then words. AI takes blocks of nearby characters — on average a couple per word — and turns them into numbers. Its alphabet isn’t made of letters; it’s made of chunks that have become digits. It reads in a foreign language, entirely numerical.
That’s why, if you ask it how many R’s are in “strawberry,” it hesitates or fires off a random number. It doesn’t see s-t-r-a-w-b-e-r-r-y. It sees two blocks — “straw” and “berry” — already turned into two numbers, something like 212 and 352. Inside those numbers the individual letters no longer exist: they’ve been swallowed. It’s not a failure of intelligence — it’s the wrong sense organ. Before calling it stupid, ask yourself whether you’re asking it to see something it literally cannot perceive.
Back to the bill — this is where it comes from. You pay for the tokens that go in and the tokens that come out. And since AI has no memory between one exchange and the next, the entire backpack — the whole chat up to that point — gets re-read from scratch every time you write. And models that “reason” cost even more: before answering they generate a pile of hidden pieces you don’t see but do pay for.
Remember the bloated backpack? Now you know why it wasn’t just more confusing — it was also more expensive, turn after turn. Saving money and using AI well are the same thing: a light backpack.
One last thing, because “I’ll generate my own tokens to save money” doesn’t hold up. A token isn’t a coin you mint — it’s work. Every token is a pass through a computer, a GPU, which means electricity. You’re not paying for text flowing for free; you’re paying for thought measured by weight, and every word costs a thread of electricity.
And how does it produce that thought? One piece at a time. Every token it writes is a prediction: it looks at everything in the backpack and guesses the most probable next piece, adds it, and starts again. That’s why it re-reads the entire backpack at every turn: it draws its next move from there. It’s the only gesture it knows: guess the next piece, endlessly.
The takeaway: a token is a chunk of a word turned into a number. That’s why AI can’t count letters, and that’s why you pay for thought by weight and pay for all of it again at every turn. It’s the same thing seen from two sides. You don’t mint tokens. You consume them.
One question remains without a clean answer: how does something so elementary — “guess the next token” — produce something that argues, translates, reasons, and even seems intelligent? Not even those who build AI know quite well. But one part of that sentence already matters now: the machine guesses the most plausible next piece. And plausible isn’t true. Or intelligent. But that’s the next article.