What a model is
The core mental model for AI — tokens go in, tokens come out — plus what a token actually is.
We've covered the hardware. Now let's meet the thing that runs on it: the model. If you remember only one idea from this entire module, make it this one:
A model takes input, and produces output. For a large language model, that means tokens go in, and tokens come out.
Everything else is detail. Let's unpack it.
The model is a box in the middle
Picture a box. On the left, something goes in — your question, an email, a customer record. On the right, something comes out — an answer, a summary, a draft reply. The model is that box. You don't need to know what's inside it to use it well, the same way you don't need to understand an engine to drive a car.
What's actually inside (for the AI behind tools like ChatGPT or Agentforce) is something called a transformer. We're going to treat it as a black box on purpose, but here's the one-sentence version so it isn't total magic: a transformer is a prediction engine that, having studied an enormous amount of text, predicts the most likely next piece of text given what came before. That's it. It's not looking anything up and it doesn't "know" facts the way you do — it predicts.
So what's a "token"?
We keep saying "tokens go in, tokens come out." A token is just a chunk of text — usually a word, part of a word, or a bit of punctuation. Models don't read whole sentences the way we do; they break everything into these chunks first.
Roughly, one token is about ¾ of a word in English. "Salesforce" might be one token; "Agentforce" might split into two ("Agent" + "force"). It's a little unintuitive, which is exactly why it's worth seeing for yourself.
See it yourself
OpenAI has a free tool that shows any text broken into tokens, with each token highlighted: platform.openai.com/tokenizer. Paste in a sentence and watch it split apart.
Why tokens matter (a lot, later)
Tokens aren't trivia — they're the unit AI is measured and billed in. Two things they'll keep coming up for:
- Cost. Using a model is priced per token, in and out. More text = more tokens = more cost.
- Context limits. A model can only "hold" so many tokens at once (its context window). That's why very long documents sometimes have to be broken up.
We'll lean on this repeatedly, so it's worth getting comfortable now.
📝 Practice / Homework
Open the OpenAI tokenizer and paste in one of your own emails or a paragraph from a document you wrote. Notice: How many tokens was it? Did any words split in surprising ways? Bring one surprising split you found.