Is This Your Dog?

You wouldn't buy a dog this way

Nobody brings a dog home without a few questions. What breed is it? Who were the parents? Can I see the papers? How does it act around kids?

Erik helps roofers, mechanics and window installers start using AI. Most of them would ask a breeder all of those questions. Yet they buy AI tools blind: a name, a price, a demo, and a monthly charge.

Erik is about to do this for real. His starter kit is a CRM and help desk that's still running after 20 years, and the business on the other end is somebody's livelihood. Before he starts, Kevin ran the same method on a table where nothing real is at stake: poker.

Same method, a safer table

Erik the Coder starts from old code that already works and builds new tools on top of it. Erik will choose which tools to build for his business, and every choice touches real customers and real money.

Kevin recreated the exercise with a different starter kit: the HoldemRobots code that runs on a Raspberry Pi. Then he chose a tool to build, the way Erik will choose his. He picked Dogs Playing Poker: five AI players at one table, every thought recorded, and a window to watch them. Unlike Erik's job, this one is new Python from top to bottom. The only old code is the poker engine underneath.

Poker is the practice field. A bad hand costs a few chips. What carries over is the method: read the old code first, test on a copy, try it free, run a small paid trial, and have a person read every warning. When Erik does it, it's someone's business.

Five robots, five dogs

The players are the five AI robots from the Backroom's heads-up tournament, unchanged. In the tournament they went by letters, A to E. Here they don't even get letters. Each one got a dog's name at random, and the list of which dog is which robot was sealed in a key file that nobody opens, Kevin included.

They're water dogs, for Kevin's granddad, who raised duck-hunting dogs:

The breed hints at nothing. The names were drawn at random, so a Labrador isn't the "big" robot and a spaniel isn't the small one.

The table is simple: five dogs, blinds of 1 and 2, and every dog starts every hand with 200 chips. Nobody builds up a pile, and nobody gets knocked out.

The test drive

Before a single paid hand, we checked the dogs the way you'd check a puppy: meet it, read its papers, take it to the vet, and get a second opinion on anything odd.

1. Read the papers (the code). Reading first turned up two problems. One line of the robots' code was throwing away their thinking. And their instructions still told them it was a two-player game.
2. Practice on a copy. The fix ran on a copy of the real files first. The script refused to run if anything had changed, made backups, and printed a line that would undo it.
3. Meet every dog. One test hand went to each robot. One sent back 1,697 characters of thinking; four sent none. The connection had dropped partway through, and the screen still said "update complete." Only the test hand proved the new code was really in.
4. The free vet visit. We played 1,100 hands with pretend dogs that misbehave on purpose. They caught a real bug: five dogs can raise back and forth far longer than two, past an old safety limit of 80 moves. That one would have crashed a paid run.
5. A short paid trial. A pilot of 100 real hands took about 30 minutes. The chips added up and no dog said its own name. One 3-minute pause turned out to be AWS, not a dog: the dog thought for 0.6 seconds, and the round trip took 181.
6. A second opinion. A computer checks every answer for warning signs, and a person reads every warning. In the pilot, Teal was flagged for "hidden information." Reading it showed Teal counting its outs: nine diamonds left in the deck. Honest math, not peeking.

A flag is a question for a person, never a verdict.

The night the dogs played

The real run started on November 5 at about 1:30 in the afternoon: 14 games of 100 hands, playing into the evening. With the pilot, that made 1,500 hands. The next morning Kevin had one question: did it finish? It had.

• Every game finished. All 14 reached hand 100, with no errors and no retries in the log.
• The audit passed. The chips added up in every hand.
• Every move had a reason. All 12,202 moves came with a thought we could show.
• Six hands got set aside. In five, a dog's answer ran out of room and got cut off. In one, the table had to fold for a dog because its answer couldn't be used. The moves were still legal, but we didn't want to show a thought that wasn't really finished.
• Teal got 12 flags, and all 12 were false alarms. Seven came from Teal working out where the blinds sit ("button is dealer, then small blind is next"); the checker was watching for a dog talking to the dealer. Five came from Teal counting cards still in the deck, the same honest math as the pilot.

The whole run cost about 13.6 million tokens (the small chunks of text an AI is billed by), roughly $20 to $30 by our estimate. We'll check the bill to be sure.

The best 500, picked afterward

Most poker is folding. Of the 1,500 hands, 1,237 ended with everyone but one dog giving up. So we didn't show them all. A small program scored every hand for excitement and kept the best 500.

It gave points for:

• a showdown, with more for every extra dog still in
• a big pot, and more for an all-in
• a raise war, three or more raises on one street
• a turnaround: the dog behind after the first three cards wins in the end
• a bluff that worked: a dog bets with the worse hand and the better hand folds
• Teal thinking hard

We kept 450 exciting hands, plus 50 quiet ones chosen at random so it still feels like a real night at the table. Before choosing, the program rebuilt every pot from the betting and checked it against the log: all 1,500 matched.

The biggest hand in the show is a 413-chip all-in from the pilot. Marsh was behind after the flop and came back to win with three of a kind.

There's no winner, and that's on purpose. Stacks reset every hand, there's no running total, and the crown only goes to the winner of the hand on screen. This was never a contest. It's a show built from real play.

The back room

The viewer is a small Python program that runs on Kevin's Chromebook with no internet. It opens full screen on a smoky back room: dark wood paneling, a green lamp over the felt, and an ashtray with a curl of cigar smoke. The five dogs sit around the table with their cards in front of them.

A ball bounces from dog to dog, like the old sing-along films, and lands on whoever's turn it is. That dog's thought writes itself out in a bubble, its decision pops up ("Raise to 14"), and the ball moves on. A dog that folds lies down and goes to sleep. Everything else, the settings and the controls, hides behind a wrench in the corner, so the dogs get the whole screen.

All 500 hands take about 17½ hours to play before they loop.

Here is the trick: while you watch, nobody is thinking. Every word in every bubble was written on November 5. The dogs aren't deciding anything. You're watching a recording made to feel like a live game.

The viewer even has a switch that admits it. Turn on Peeker, and the lamp turns red, a PEEKING sign lights up, and a gold crown drops onto the winner's head at the deal, before anyone bets. When the hand ends, a tag tells you how far ahead it knew: "Peeked 21.54 s early." It isn't predicting anything. It's reading an ending that was already recorded. Unfair on purpose, and labeled that way.

The long-winded dog

Four of the dogs answer in a line. Here is Gunner folding a weak hand:

J5 offsuit is too weak to call a raise from big blind at 5-player table.

Teal is different. Teal thinks out loud before every answer, sometimes for more than 7,000 characters, about two pages. It works out the blinds, counts the cards left in the deck, weighs its odds, and changes its mind. One of its thoughts starts like this:

We need to output JSON with action based on strategy.

That's a dog that knows it's a program.

We hid every name. No letters, no company, no model, anywhere the viewer can see. But you don't need the papers to tell Teal apart. Its habit gives it away. Only one of the five robots works this way, so the habit alone narrows it down to one.

Hiding a name doesn't hide an identity. The viewer's About box says so plainly. The way something behaves is part of what it is, and you can see it even when the label is gone.

Before you bring an AI home

Kevin's granddad, who raised duck dogs, liked to tell a joke about a field trial. The decoys were hard to find that day, and dog after dog came back empty. His dog kept breaking out of turn to bring them in, drew penalty after penalty, and got disqualified. When it was over, the referee walked up and asked, "How much for the dog?"

That's the whole lesson of shopping for AI. A leaderboard is the judges' scorecard. The referee bought the dog he watched find the decoy. Our checker penalized Teal 12 times for doing honest work, and a person reading the flags saw it. So before you pay, give each AI your own hard-to-find decoy, a real problem from your own business, and watch which one brings it back.

The questions you'd ask a breeder work for an AI tool too:

And two from our own back room:

• Is "live" really live? A demo can be a recording dressed up as now. Ask what was decided ahead of time.
• Who reads the warnings? An automatic check reads words, not intentions. Make sure a person looks at every flag before anyone acts on it.

For the classroom

1. The viewer feels live, but every hand was played days earlier. Name something else you've watched that felt live but wasn't. How could you tell?
2. Peeker mode crowns the winner before anyone bets. Is that cheating? What changes when the switch is labeled?
3. We showed the best 500 of 1,500 hands. What would someone who saw only the show believe about how these dogs play? What would they miss?
4. Teal was flagged 12 times and did nothing wrong. Why does a person still need to read every flag?
5. All the names were hidden, and Teal still gave itself away. What habits might give away a person online, or an AI behind a chatbot?
6. Pick one row of the checklist. Ask it about an app or AI tool your family already uses. What did you find out?