BOOK 2 · FRONT OF THE BOOK · FOR EVERYONE
Two Weeks: How One Person Ran an AI Tournament for $15.52
Five AI models from five companies, one poker table, two thousand hands overnight, a sealed list of who was who, and an AI bill smaller than a pizza. It wasn't the first AI poker tournament. That's the point: it isn't rare any more, and you can run one too.
Two weeks
When I give talks about skepticism toward enterprise AI, I tell a joke. What's the difference between a snowboard instructor and their student?
Two weeks.
Kevin's joke, from a talk on adopting enterprise AI
Nobody is born an instructor. The instructor just started a little earlier and fell down a few more times. AI is the same. The people who seem to know it best mostly got on the mountain first.
Book 1 ended with a plan. In its last big article, "Back to the Drawing Board" (Part 26, October 24), I promised AI players from different companies at one table, under the same rules. I warned that it would happen "in Saturdays, not sprints," added "Fair warning: it may take a while," and asked readers to write down a prediction: does the most expensive model win?
By the following Thursday, the first of the three planned events, the heads-up round robin, had been played, scored and audited. (The answer to the prediction is no. The results are below.) The six-seat game and the knockout bracket are still to come.
This chapter is about the time I got on the mountain: a tournament between AI models from five different companies, run by one person in the evenings. I'm writing it down so you can see exactly how small the gap between watching and doing really is.
What happened
- The field. Five AI models from five companies, all rented from one cloud (Amazon's Bedrock). AWS turned away three of my first picks, so the field became the five my account could actually use (Dialog 4).
- The table. Heads-up Hold'em: two players at a time. Every pair of models played 100 hands, then the same 100 decks again with seats swapped, so luck with the cards mostly cancels out. Ten matches, 2,000 hands.
- The blind. Each model got a codename, A to E, assigned by a script and sealed before the first hand. I published the key's fingerprint, so anyone can later check I didn't change it. As I write this, I still haven't opened it.
- The night. It ran unattended from 7:41 pm. By morning: 20 games, every chip accounted for, 614 times the system had to correct or replace an answer. A person read every one.
- The verdict. Play was judged by letter and frozen at 9:22 am. Only then did a script work out each seat's cost. One seat's results were clearly beyond luck; the other four couldn't be told apart in 800 hands each. And price and quality didn't line up. (The results are below.)
- The check. That afternoon a read-only script audited the Dealer over all 2,000 hands: same cards both ways, right winners, nobody shown a hidden card. All clean.
Was it a first?
No, and I'm glad. In October 2025 a project called PokerBattle.ai ran nine AI models for 3,799 hands over five days; OpenAI's o3 came out ahead. A benchmark company, Vals.ai, ranks seventeen models over 20,000 hands, using fixed shuffles to keep luck in check. Researchers have published papers on training models to play.
What I haven't seen in their write-ups is the combination this one tried: one person, one cloud bill, codenames sealed before play, a verdict on price as well as play, an audit of the dealer, and a public list of every way the test wasn't fair. The tournament's rule from the start was "This is not science." We wrote down 21 ways it wasn't, and kept going.
But the real lesson is the first sentence of this section. When a hobbyist, a benchmark company and a CIO in New Mexico can all run AI poker tournaments, the barrier isn't knowledge or money any more. It's starting.
What it actually took
| WHAT | HOW MUCH |
|---|---|
| Time | Six days of evenings, from the first backroom conversation (Oct 24) to results (Oct 29) |
| The AI bill for the tournament | $15.52 for 2,000 hands across five models. The AWS bill agreed within 1.4%. |
| The server | About $34 a month, already running the site. It cost more than all the AI play. |
| The building partner | An AI chat account: the lowest-priced Claude plan above the free one. The step up matters because I want a partner that remembers who I am and what we're building (Book 1, Part 19, "The File That Remembers"). Move up a level only when you keep running out of usage. Claude wrote the commands and scripts; I ran every command and made every call. |
| Hardware | A $200 Chromebook |
I should be honest about one advantage. I'm a CIO, and I've worked with every layer of a system like this for years. But the backroom was built through conversations, one step at a time, and I made plenty of beginner mistakes along the way: commands run on the wrong machine, a paste missing its last byte, a server too small for the job. They're all in the Dialogs. None of them stopped the work.
The results, by letter
Here's how the five seats played, in big blinds won per 100 hands: the usual poker yardstick. The last column asks how often luck alone would produce a result that size.
| SEAT | BIG BLINDS PER 100 HANDS | LUCK ALONE WOULD DO THIS |
|---|---|---|
| C | +148 | 0.2% |
| B | −8 | 84% |
| E | −13 | 74% |
| D | −63 | 14% |
| A | −64 | 10% |
C won, and it's the only result luck can't explain: it won all four of its matches. The other four seats can't be told apart in 800 hands each. Which company's model C is stays sealed for now, and so do the costs by letter: a price per letter would name the models for anyone with the cloud's price list. When the key is opened, it will be checked against the fingerprint published before the first hand.
Twenty years in the making
HoldemRobots isn't new. In 2005 I taught myself the new tools of that era: open source software, Linux, the Tomcat web server and the Cocoon framework. I used them to build a craps game, then cloned it into a Texas Hold'em table. By about 2006 HoldemRobots was online, with a banner that promised "Battle the Bots! Build Your Skills! Texas Hold'em Online Training!" (you'll see it in a few pages, beside today's table).
Its robots were simple by today's standards. They followed rules, some leaned on chance, and a few, it turned out years later, could see cards they were never dealt. Catching and fixing them is the story of Book 1.
Twenty years later, the tools have changed again, and I don't intend to become obsolete.
Kevin, Book 1, Part 0
You can build the thing you want
Here's what changed for me. Around 2015 I tried to bring HoldemRobots back and couldn't stand it up. The code was all there, but everything around it had moved on, and there was nobody to help untangle it, so the table stayed dark for another ten years (Book 1, Part 12). In 2026, with AI coding tools, the 2005 engine came back to life, and once it was running, an idea on a Saturday was a finished tournament by the next Thursday. I describe what I want, the AI partner drafts it, I run it and check it, and we fix what breaks.
That's the real news, and it isn't about poker. If there's something you've wanted to exist, a tool for your job, a game, a way to answer a question at work, building it from scratch is now within reach of one person with spare time. Not because the AI does it for you, but because it shortens the distance between knowing what you want and having it.
And when you're done, you have more than a skill. You have the thing itself, with evidence anyone can check. For the tournament that's game numbers to replay, a published fingerprint, a log of every correction and an audit that ran clean. That's worth more than a line on a résumé.
Your two weeks
Pick the door that fits you.
| IF YOU'RE… | START HERE | WHAT YOU'LL HAVE TO SHOW |
|---|---|---|
| Into poker | Part I: play the robots with a hand journal, then the heads-up AI. | A journal that shows how you think, and a read on a bot you learned to beat |
| Curious about AI | Part II: write a rules robot in Robot Builder, then ask an AI to play the same spots and compare its reasons with yours. | A robot, its game numbers, and one rule you couldn't write |
| A decision maker | Take one real task at work. Give it to two AI vendors under the same instructions. Judge the results before you look at the price, then compare the cost. | A "worth it compared to peers" verdict, with the evidence |
The third door is the one I care about most. Most organizations run on a single AI vendor and never see what the same work would cost elsewhere. With nothing to compare it to, a price is just a price. A small, honest comparison, written down, is worth more than any vendor's slide deck.
Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. Tournament figures are from Backroom Dialog 10 and the diversions record of October 29, 2026. Other AI poker events: PokerNews on PokerBattle.ai (November 2025) and Vals.ai Poker Agent benchmark (December 2025); checked October 29, 2026. The project's history is from Book 1, Parts 0 and 12; the tournament plan and the prediction from Part 26. The snowboard joke and the advice on choosing an AI plan are Kevin's; current plans are listed at claude.com/pricing. Results are by letter; the codename key is still sealed.