The Best Poker Strategy, Tested on 1 Billion Hands - Dice Lab
November 13th, 2026 at 9:11:18 AM
permalink
The Best Poker Strategy, Tested on 1 Billion Hands - Dice Lab - 261 subscribers
https://www.youtube.com/watch?v=6iL4oswBOS0
Every poker player has a theory. Play tight. Bluff more. Never slow-play aces. Never play seven-two. So I built a Texas Hold'em table in code, sat seven bots at it with seven different styles, and dealt almost two billion hands.
The bots: the Rock (top 8% of hands only), the TAG (tight-aggressive, the "pro"), the LAG (loose-aggressive), the Calling Station, the Maniac, the Shover (all-in every hand) and Random.
What came out:
The pro didn't win. At mixed tables the Calling Station came first with +6.44 big blinds per hand. With only the four serious bots at the table, it came last (−0.82 per hand) and the LAG won.
Heads-up, the Rock lost to a bot that plays at random.
7-2 offsuit is only the 5th-worst hand heads-up (3-2 offsuit is the worst at 32.4%). Against five opponents it really is the worst, at 8.6%.
Pocket aces win 85% against one random hand but less than half the time against five.
Bluffing a calling station lost money at every frequency we tested. Against Rocks it helped, but only a little.
Slow-playing aces helped slightly at a wild table and lost 1.4 big blinds per hand against good players.
How it was tested: 6-max no-limit Hold'em, 100 big blind stacks reset every hand, no rake. 1,970,600,000 hands in total: 819 million full bot-league hands with betting, plus equity runouts and random deals. The hand evaluator was checked against all 133,784,560 possible 7-card hands. These are rule-based bots, not solvers, so the results describe these styles against each other.
https://www.youtube.com/watch?v=6iL4oswBOS0
Every poker player has a theory. Play tight. Bluff more. Never slow-play aces. Never play seven-two. So I built a Texas Hold'em table in code, sat seven bots at it with seven different styles, and dealt almost two billion hands.
The bots: the Rock (top 8% of hands only), the TAG (tight-aggressive, the "pro"), the LAG (loose-aggressive), the Calling Station, the Maniac, the Shover (all-in every hand) and Random.
What came out:
The pro didn't win. At mixed tables the Calling Station came first with +6.44 big blinds per hand. With only the four serious bots at the table, it came last (−0.82 per hand) and the LAG won.
Heads-up, the Rock lost to a bot that plays at random.
7-2 offsuit is only the 5th-worst hand heads-up (3-2 offsuit is the worst at 32.4%). Against five opponents it really is the worst, at 8.6%.
Pocket aces win 85% against one random hand but less than half the time against five.
Bluffing a calling station lost money at every frequency we tested. Against Rocks it helped, but only a little.
Slow-playing aces helped slightly at a wild table and lost 1.4 big blinds per hand against good players.
How it was tested: 6-max no-limit Hold'em, 100 big blind stacks reset every hand, no rake. 1,970,600,000 hands in total: 819 million full bot-league hands with betting, plus equity runouts and random deals. The hand evaluator was checked against all 133,784,560 possible 7-card hands. These are rule-based bots, not solvers, so the results describe these styles against each other.
November 13th, 2026 at 11:56:48 AM
permalink
The Best Poker Tournament Strategy, Tested on 500,000 Tournaments - Dice Lab
https://www.youtube.com/watch?v=JMTZKsIy4bM
You asked for tournaments, so I built a poker tournament room in code: 1,000 players, $100 buy-ins, nine-handed tables, a blind clock, table breaking and real payouts (150 paid, $170 min-cash, $17,000 for 1st). Then I filled it with 10 bot strategies and ran 540,000 tournaments, more than 4 billion hands.
The players: Rock, Tag (tight-aggressive), Lag (loose-aggressive), Station, Maniac, Random, Survivor (tightens up to make the money), Jammer (all-in or fold, every hand), Pro (Tag + a push/fold chart + ICM) and the Pro's Twin (same, without ICM).
What came out:
Lag won in every room it played in: +110% profit in the all-bots league, +136% against a realistic field. Tag +112%, Pro +83%, Twin +86%, Survivor +70%, Jammer +33%, Rock +27%, Station −80%, Maniac and Random about −99%. (My recreational players are very weak, so real-world profits are far smaller; the order is the point.)
Playing to make the money: the Survivor cashed 28% of the time vs Tag's 23%, but won 42% fewer tournaments and made +45% profit vs +96%.
My first run said all-in-or-fold was the best strategy in poker. It was two bugs in how my bots called all-ins. Fixed, everything re-run: the Jammer is fine in fast tournaments (+28%) but loses 62% at a World Series pace and 32–54% against players who know its range.
ICM: 50% of the chips = 38.4% of the money; doubling a short stack adds just 23% to its equity; a bubble coin flip needs 69% to call. Using ICM around the bubble made my Pro cash more (15.5% vs 14.5%) but win less than half as many tournaments, and it lost ~10% to its Twin. ICM at the final table only came within ~1%. Against a normal field, it made no difference.
A chip and a chair: the final-table chip leader wins 23%; the shortest stack wins 3.1% (1 in 32). 95% of winners fell below 10 big blinds at some point; one got down to 0.002 big blinds and still won.
Luck: even with a huge edge, 17% of Lag's 100-tournament careers lost money; it takes ~300 tournaments to be 95% sure it's a winner. The Random bot won 34 tournaments.
Structure: the slower the tournament, the more skill pays (Pro +33% turbo → +176% at a World Series pace). A 10% rake cut the Pro's profit by 19 points.
How it was tested: No-Limit Hold'em, 9-max, 1,000-player freezeouts, 100 big blinds, a new level every 10 hands (+~26%), big-blind ante from level 3, top 15% paid. Card evaluator checked against all 133,784,560 seven-card hands; zero chips created or lost in 4.48 billion hands; identical bots win equal shares. Real-world facts: WSOP Main Event 2024 (10,112 entries, $10M first), Chris Moneymaker 2003 ($86 online satellite → $2.5M), Jack Straus 1982 (a chip and a chair, according to the legend), Jordan Griff (71% of the chips heads-up, 2024).
Chapters:
0:00 540,000 tournaments
1:10 How a tournament works
2:39 My tournament
3:26 The 10 players
4:28 Round one: every strategy
5:38 Round two: a realistic field
6:36 The bug and the Jammer
8:32 The bubble
9:02 Survivor vs Tag
10:27 ICM: chips are not money
11:53 Does ICM pay?
13:25 The push chart
14:09 A chip and a chair
15:24 One full tournament
15:59 Luck vs skill
16:58 Turbo vs deep, field size, rake
18:13 The final ranking
18:39 The best tournament strategy
https://www.youtube.com/watch?v=JMTZKsIy4bM
You asked for tournaments, so I built a poker tournament room in code: 1,000 players, $100 buy-ins, nine-handed tables, a blind clock, table breaking and real payouts (150 paid, $170 min-cash, $17,000 for 1st). Then I filled it with 10 bot strategies and ran 540,000 tournaments, more than 4 billion hands.
The players: Rock, Tag (tight-aggressive), Lag (loose-aggressive), Station, Maniac, Random, Survivor (tightens up to make the money), Jammer (all-in or fold, every hand), Pro (Tag + a push/fold chart + ICM) and the Pro's Twin (same, without ICM).
What came out:
Lag won in every room it played in: +110% profit in the all-bots league, +136% against a realistic field. Tag +112%, Pro +83%, Twin +86%, Survivor +70%, Jammer +33%, Rock +27%, Station −80%, Maniac and Random about −99%. (My recreational players are very weak, so real-world profits are far smaller; the order is the point.)
Playing to make the money: the Survivor cashed 28% of the time vs Tag's 23%, but won 42% fewer tournaments and made +45% profit vs +96%.
My first run said all-in-or-fold was the best strategy in poker. It was two bugs in how my bots called all-ins. Fixed, everything re-run: the Jammer is fine in fast tournaments (+28%) but loses 62% at a World Series pace and 32–54% against players who know its range.
ICM: 50% of the chips = 38.4% of the money; doubling a short stack adds just 23% to its equity; a bubble coin flip needs 69% to call. Using ICM around the bubble made my Pro cash more (15.5% vs 14.5%) but win less than half as many tournaments, and it lost ~10% to its Twin. ICM at the final table only came within ~1%. Against a normal field, it made no difference.
A chip and a chair: the final-table chip leader wins 23%; the shortest stack wins 3.1% (1 in 32). 95% of winners fell below 10 big blinds at some point; one got down to 0.002 big blinds and still won.
Luck: even with a huge edge, 17% of Lag's 100-tournament careers lost money; it takes ~300 tournaments to be 95% sure it's a winner. The Random bot won 34 tournaments.
Structure: the slower the tournament, the more skill pays (Pro +33% turbo → +176% at a World Series pace). A 10% rake cut the Pro's profit by 19 points.
How it was tested: No-Limit Hold'em, 9-max, 1,000-player freezeouts, 100 big blinds, a new level every 10 hands (+~26%), big-blind ante from level 3, top 15% paid. Card evaluator checked against all 133,784,560 seven-card hands; zero chips created or lost in 4.48 billion hands; identical bots win equal shares. Real-world facts: WSOP Main Event 2024 (10,112 entries, $10M first), Chris Moneymaker 2003 ($86 online satellite → $2.5M), Jack Straus 1982 (a chip and a chair, according to the legend), Jordan Griff (71% of the chips heads-up, 2024).
Chapters:
0:00 540,000 tournaments
1:10 How a tournament works
2:39 My tournament
3:26 The 10 players
4:28 Round one: every strategy
5:38 Round two: a realistic field
6:36 The bug and the Jammer
8:32 The bubble
9:02 Survivor vs Tag
10:27 ICM: chips are not money
11:53 Does ICM pay?
13:25 The push chart
14:09 A chip and a chair
15:24 One full tournament
15:59 Luck vs skill
16:58 Turbo vs deep, field size, rake
18:13 The final ranking
18:39 The best tournament strategy
