Play with AI from the inside.
Not theory: five simulations you can touch right now to understand how AI actually works under the hood. They're real pieces of my course. Play with them — and if it sparks your curiosity, keep going.
How AI splits your text
Models don't see letters or words: they see tokens. That's why they sometimes miscount letters. Type something and watch.
Tokenizer: see how AI splits your text
22
Tokens
11
Words
67
Characters
2.0
Tokens / word
Visual approximation: models don't see letters or words, they see tokens (fragments). That's why they sometimes miscount a word's letters — they literally don't see them.
AI predicts, it doesn't know
It doesn't look up the answer: it builds it token by token, choosing by probability. Turn up the temperature and watch it get bolder.
Token-by-token prediction (with temperature)
Next token — probability by temperature (click to pick):
The model doesn't look up the answer: it builds it by choosing the next token by probability. At temperature 0 it always picks the most likely one; turn it up and it ventures into less obvious options.
Why AI 'forgets'
Its memory is a fixed-size desk. When it no longer fits, the oldest thing falls off the edge. Watch it happen here.
The context window: a fixed-size desk
Capacity: 1/5
It fits a lot, but not infinitely. If you keep stacking, the first thing you said falls off the edge — that's why the model 'forgets' your initial instruction. The fix: remind it, or open a new chat.
A good prompt changes everything
The gap between a mediocre answer and a brilliant one is almost always in how you ask. Build a prompt and watch its quality climb.
Prompt lab: build a good one piece by piece
Your prompt
Build me a login form component with email and password fields.
Generic result
The AI guessed your stack and your design. It probably won't fit and you'll redo it.
A good prompt = Context + Task + Format + Constraints. Turn on the pieces and watch the same model go from handing you garbage to handing you gold.
How you ask AI for an image
Generating images with AI is a language: subject, style, light, framing. Build yours piece by piece.
Build an image prompt, ingredient by ingredient
Generic cliché
A clock, a rocket, or people in suits. It doesn't fit your brand — you toss it.
Your image prompt
A mediocre image is almost never the model's fault: it's a vague prompt. Turn on the 6 ingredients and watch yourself go from a cliché to an on-brand cover.
Did it spark your curiosity?
Leave me your email and I'll send you the best of how to build with AI from scratch — no filler, from the surgeon who built 13 products learning with AI.
Or go straight to the good stuff:
See the free course (3 modules) →