Writing
Same AI, same chess app, 100x the bill
2026-10-10
A thought experiment: an experienced coder and a novice build the same app with the same AI. The difference in cost comes from the decisions, not the typing.
X (@g0mu_)
Give an experienced coder and a complete novice the same AI and the same job: build a chess app that runs in the browser.
Thought experiment: my guess is the coder gets there on about 10 million tokens and the novice burns a billion. Tokens are roughly what you pay the AI to read and write. Same model, same app, about 100 times the bill.
Where does the rest go?
The typing costs about the same, because the AI types for both of them. The difference is in the questions the novice can't ask yet.
The coder says: use an existing chess library for the move rules. Keep the board state in one place. Test castling and en passant before anything looks pretty. Stop when it works.
The novice says: make me a chess game. Then: the knight moves wrong. Then: now the pieces disappear. Every fix drags the whole history back in and pays for it again. The AI is doing its best with decisions nobody made.
So with AI, most of the cost of building sits in knowing what to decide up front, what to check, and when you're done. That knowledge used to show up as code you wrote. Now it shows up in the instructions you give.
That's why I don't think AI makes experience worthless. Right now, experience is the thing that cuts the bill.
If you've tracked token spend on a real project, send me your number.
#MakingSandThink
Give an experienced coder and a complete novice the same AI model and the same job: build a chess app that runs in the browser. Thought experiment: my guess is the coder gets there on about 10 million tokens and the novice burns through a billion. Tokens are roughly what you pay the AI to read and write, so that's the same model and the same app at about 100 times the cost.
The typing costs about the same for both, because the AI does it. The gap comes from the questions the novice doesn't know to ask yet.
The coder opens with decisions. Use an existing chess library for the move rules instead of writing your own. Keep the board state in one place. Test the awkward moves, like castling and en passant, before anything looks pretty. Stop when it works.
The novice opens with "make me a chess game", then "the knight moves wrong", then "now the pieces disappear". Every fix drags the whole history back in and pays for it again. The AI is doing its best with decisions nobody made.
So with AI, most of the cost of building sits in knowing what to decide up front, what to check, and when you're done. That knowledge used to show up as code you wrote. Now it shows up in the instructions you give.
That's why I don't think AI makes experience worthless. Right now, experience is the thing that cuts the bill.
If you've tracked token spend on a real project, send me your number.