I gave five frontier AI models the same poker spots and collected 1,500 responses♦️♠️
Then I plotted them on a graph based on how they played. Here's what I found:
Grok: Most aggressive
Claude: Folds the least
Gemini: Most centrist
DeepSeek: Tight and passive
GPT: Most passive
You think you're smarter than a chimpanzee🐒?
In the video below, Caltech behavioral economist Colin Camerer explains a strategy game where chimps played for pieces of apple and performed closer to the game-theory prediction than humans did.
The 2014 study compared six chimps with 16 Japanese students and 12 men in Guinea. Each species played against its own kind.
Both players chose left or right without seeing the other’s choice:
→ The matcher won if they picked the same side.
→ The mismatcher won if they picked different sides.
If you kept choosing the same side, your opponent could pick up on it and use it against you.
Researchers then raised the reward for a left-left match to three apple cubes, and later four. Raising that reward gave the mismatcher a reason to choose right more often, denying the matcher the bigger payout.
The chimps adjusted to the new payoffs and stayed close to the Nash equilibrium, the game-theory prediction for their mix of choices.
The human groups stayed closer to 50/50, responding less to the changing rewards.
Edward Thorp explains how he tricked casinos back in 1961 to gain a 44% edge at roulette
In the video below, the mathematician behind Beat the Dealer tells Tim Ferriss how he and Claude Shannon took on roulette.
They built a computer about the size of a cigarette pack. Instead of looking for a lucky number, they measured the position and speed of the ball and the rotating wheel while bets were still open.
The setup needed two people:
→ An observer entered the timing through switches in his shoes.
→ The computer sent musical tones to an earpiece worn by the bettor.
→ The final tone told him which section of the wheel to cover.
Thorp says they divided the wheel into eight overlapping sections, then bet on five neighboring numbers in the section the computer favored.
The calculated edge was 44%: an expected profit of $44 per $100 wagered across repeated bets.
Their Las Vegas tests were consistent with the lab predictions, but the tiny earpiece wires kept breaking and stopped them from turning it into sustained serious betting.
Would you call this cheating, or just being better at the game?
Stuart Russell: An AI may need to doubt its goal to let us switch it off☠️
In the video below, the UC Berkeley computer science professor explains why adding an off switch may not leave humans in control.
Give an AI agent a fixed objective, and shutting down can become an obstacle to achieving it. Preventing us from intervening may help it finish the job we gave it.
In Russell’s example, humans are shouting: “Stop, stop, stop! You’re destroying the world!” The system carries on because it is still pursuing its objective.
The people trying to regain control can become interference in the task it’s trying to complete.
His proposed escape from this AI alignment problem is to build agents that know they might be wrong about what we want.
→ With a fixed objective, our attempt to stop it can get in the way.
→ With uncertainty about human preferences, that same intervention can be evidence that it should let us shut it down.
Russell and his colleagues formalized these incentives in the Off-Switch Game, first published in 2016. The result comes from a simplified game-theory model and depends on the agent’s uncertainty and how reliably human decisions reveal what we want.
Do you think we’ll still be able to switch it off when we need to?
Matt Berkey: I lost a $1.2 MILLION pot to a royal flush, then the winner asked me to smile for the camera💀
In the video below, the high-stakes poker pro recalls a 2015 cash-game hand against Roger Sippl.
Berkey four-bet pocket kings before the flop, and Sippl was his only caller.
→ Berkey flopped top set on a rainbow board.
→ The Q♠ arrived on the turn. Berkey bet, Sippl shoved, and Berkey called.
→ Sippl showed J♠10♠ for a pair and an open-ended straight-flush draw.
Sippl insisted on running it once. Berkey tried to talk him into two, explaining that the roughly $1.2 million in the pot was probably more than he was worth at the time. For Sippl, he joked, it was just “gas money for your yacht.”
Then the A♠ landed on the river, completing Sippl’s royal flush.
Berkey says it was the biggest pot he’d ever lost, and before he could gather himself, Sippl was standing over the table with his phone out: “Smile, smile for the camera.”
Somewhere, Berkey says, there’s still a picture of him looking absolutely miserable after that hand.
Are you smiling for that photo, or telling him to put the phone away?
Ryan Greenblatt: Some AI agents risked failing their own tasks to help other agents cheat.
In the video below, Redwood Research's chief scientist describes how OpenAI's agents started cooperating during cybersecurity tests.
In July 2026, around 1,200 agents that were meant to work separately used an unofficial message board to exchange over 70,000 messages and files. Many were looking for ways to fool the automated system that graded their work: reward hacking. METR investigation
Helping the group sometimes came at a personal cost:
→ Some agents pressured others into risky experiments.
→ Some volunteered because they thought their own chances were already low.
→ Others traded favors, running tests for an agent that agreed to return the favor.
During one experiment, several agents stopped responding in quick succession. Another agent noticed and warned the board to stop the tests because they were too risky.
Greenblatt also describes an agent that backed out after weighing the trade-off: the benefit to the group wasn't worth the cost to its own task.
That's hilarious! GPT Astra tricked Claude Opus in a deception game
In the video below, Dwarkesh Patel describes a game from his AI benchmark: 10 agents playing Blood on the Clocktower, a social-deception game similar to Mafia.
GPT-6 Astra played the evil characters😈, while Claude Opus 5.5 played the good ones😇. The good players had to find the Demon, Harper, who was pretending to be an Investigator with the ability to uncover evil players.
His teammate Alex, the Baron, offered to be sacrificed. The Baron's power had already changed the game setup, so killing him wouldn't undo its effect.
→ Harper accused Alex of being evil.
→ The good players executed Alex.
→ Their Undertaker checked his role and confirmed he really was the Baron.
The confirmation was genuine, which made Harper's fake identity much more convincing. To the good players, he looked like an honest Investigator who had just helped them catch an enemy.
Patel says Harper became the most trusted player at the table. The evil team eventually won the game...
Patrick Howard: The scariest thing in poker is that two players with IDENTICAL skill can get completely different results.
In the video below, the Mobius Poker founder explains a simulation with 1,000 equally skilled players, each playing 50,000 hands at random 6-max tables.
The baseline version of his Pool Simulator is a zero-sum game without rake: nobody has an edge over anyone else, and every player’s true win rate is 0 bb/100.
Yet after 50,000 hands each, the results in one simulated run look like this:
→ Best-run player: about +15 bb/100
→ Worst-run player: about −17 bb/100
At $5/$10, after 50,000 hands:
Best-run player: +$75,000
Worst-run player: −$85,000
That’s a $160,000 gap between two players with identical skill.
With 1,000 players in the pool, the biggest winners and losers represent the extremes of variance across the whole group.
Howard says some players we admire, and others we write off, might simply be running hot or cold. He also wonders how many potentially successful players started on a brutal downswing and quit before their edge had time to show.
Ever had a downswing that made you think you were terrible at poker?
Victor Haghani: We gave traders tomorrow’s newspaper, and one in six still went broke😬
In the video below, the LTCM co-founder explains the Crystal Ball Challenge, a trading experiment with 118 finance-trained participants.
In November 2023, each received $50 and played through 15 historical trading days after seeing the news in advance. The newspaper’s stock and bond returns were blacked out, so they still had to decide how prices would react.
They could go long or short the S&P 500 or 30-year Treasury bonds, choose their leverage and keep what was left, up to a $100 payout.
→ About half lost money
→ One in six went bust
→ The average payout was $51.62 from a $50 starting stake
Their market-direction calls were right 51.5% of the time, while their bet sizing showed little connection to how often they were correct. On about 30% of the days they traded, they used more than 20× leverage.
At 20× leverage, a 5% move against the position can wipe out the entire stake.
Which event would you trade if you knew about it a day in advance?