BOOK 2 · YOUR TURN
Back to the Drawing Board 2.0
Book 1 ended with a plan for a tournament. Book 2 played the first event of it. The rest of the drawing board isn't mine to finish. It's yours.
Where we stopped
Five companies' AI models played a heads-up round robin, every deal twice with the seats swapped. Seat C won, and it's the only result luck can't explain. The play was judged before the price, the dealer was audited, and the sealed key that says which model is which is still closed. Book 1's plan had three events. One is done.
There won't be a Book 3. Instead, here is everything left on the board, written as challenges. Each one is small enough for a student, a class or a club to take on, and each one is a real question nobody has answered yet.
The open board
| CHALLENGE | WHAT TO FIND OUT | WHERE TO START |
|---|---|---|
| Rules against AI | Can a rules robot built from a person's read of the AI beat the AI it was built to read? | "The Rule You Can't Write" |
| Bot 2.0, in plain English | Take your hand journal's "What no checkbox could say" lines, write them as instructions for an AI player, and find out whether it plays better. | "What Were You Thinking?" and Part III |
| A full table | Book 1's second event: six players at one table, everyone dealt the same cards in turn. Start with six student robots on the Pi, no internet needed. | Book 1, Parts 22 and 26 |
| A memory | Give a robot or an AI a memory of this one opponent. Does it learn the habits a person would? | Advanced exercises; Book 1, Part 11 |
| Catch a bluff | None of our robots tries to read the other players yet. Start with the betting record: can a bot spot a bluff from the bets alone? Then go further. With a classmate's permission, put a heart-rate monitor on them while they play. Can you tell when a person is bluffing? Can a machine? | "Read the Player, Set the Trap"; Book 1, Part 11 |
| Your own dealer | Build a dealer for a tiny game (rock-paper-scissors will do) that shows each player only what it's allowed to see, then try to break it. | Book 1, Part 15; "Hack an Agent, If You Can" |
| A fair test at work | Give one real task to two AI vendors under the same instructions. Judge the results before you look at the price. | "Two Weeks"; Backroom Dialog 8 |
Use the same rules the tournament used. Write down how you'll judge the result before you start, and date it. Keep a list of everything that isn't equal. Play every deal twice if you can. And when you're done, write it up so the next person can check you.
Make your prediction
Book 1 asked whether the most expensive model would win. It didn't. Here are three new questions. Write your answers down, with a reason:
- Which company's model is seat C? The sealed key will be opened, checked against the fingerprint published before the first hand.
- Will a bot tuned in plain English beat seat C?
- Will a rules robot built from a person's read of the AI beat the AI that was read?
When the answers come in, whether from the key or from your own experiment, check yours the same way you'd check any AI's answer: not just whether you were right, but whether your reason was.
Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. Book 1's three-event plan is from Part 26; the tiny-game dealer idea is from Part 15's "Build your own." Nothing on this page has been played yet. No money is played, ever.