AI summary by TwinMind
Oliver Roeder | Seven Games: A Human History | Talks at Google
Talks at Google
58m 58s
Introduction and Background
- Matt Bon Jovi introduces Oliver Roeder, journalist/author, former 538 senior writer, Riddler puzzle column editor
- Roeder holds PhD in economics (game theory focus), studied AI as Neiman Fellow at Harvard
- Discussing new book "Seven Games: A Human History" - biography of enduring games (chess, poker, go, others)
- Book examines why humans play games, rule arcana, interplay between elite champions and emerging AI players
- AI dominance in several gaming realms
Why Humans Play Games - Philosophical Framework
- Three core reasons humans play games:
- Fun/Entertainment: Games occupy time enjoyably
- Practice for Real World: Prehistoric hunting analogy - throwing rocks at trees vs. constant dangerous hunting
- Chess: practice strategy/planning ahead
- Backgammon: grapple with uncertainty via dice rolls
- Poker: confronted with hidden information/deception
- Art Form/Agency: Games capture human agency - deciding, deliberating, acting
- Painting captures visual world, music captures heard world
- Games allow exercise of agency unavailable in everyday life (leading armies in chess)
- Border between games and real life is porous - legitimate learning transfer
- Personal example: Scrabble taught intertemporal trade-offs (scoring today vs. saving tiles for future turns)
- Games provide fast feedback loops - lessons come "fast and furious"
Early Computing and Games Marriage
- Practical Appeal:
- Small, manageable domains
- Easily measurable outcomes (winners/losers generate data)
- Can play many iterations quickly
- Romantic/Philosophical Appeal:
- Games tied to concept of intelligence itself
- 1950s Dartmouth AI symposium listed highest human callings: writing music, proving mathematical theorems, playing chess
- Chess viewed as pinnacle of human intellectual achievement
Checkers - First Real AI Gaming Battleground
- Arthur Samuel on early IBM mainframes testing with checkers programs
- Checkers unfairly dismissed as "child's game" vs. chess
- Actually first real test of machines, site of seminal computer science developments
- Key programming challenges: efficiently traverse game tree, accurately evaluate positions
The Tinsley vs. Chinook Saga
- Marion Tinsley: Best checkers player in world, possibly best competitor in any pursuit in history
- 40-year career spanning ~1,000 tournament games
- Lost exactly 3 times total
- Two great loves: checkers and Christian faith (evangelical minister)
- Jonathan Schaefer: University of Alberta computer scientist
- Developed "Chinook" program (named after Canadian wind, plays on "drafts")
- Life's work for almost 2 decades
- Cooks: Checkers jargon for good moves not in established literature
- From "your goose is cooked"
- Checkers has extensive literature of accumulated human wisdom
- Many games drawn at highest levels, single wins hugely important
- Schaefer set Chinook to hunt down these undiscovered moves
- Top players look 20-30-40 moves ahead, much of tree well-known via memorization
- Race against time: Schaefer trying to beat aging Tinsley before natural human mortality
Human Adaptation to AI Gaming
- Predictable Pattern Across Games:
- Initial utter disdain: "No machine could do what I do"
- Rapid transition after losses: "Hey, can I have that machine to learn from?"
- AI Democratization Effects:
- Democratization of skill - anyone can get good quickly
- No longer need good chess club, teacher, or ability to afford one
- Acceleration of knowledge dissemination
- Broad-based increase in tournament/competitive play levels
Chess and AI Integration
- Spectator Enhancement: AI-aided strength indicators in tournaments
- Visual "thermometer" showing position evaluation
- Real-time assessment of move quality for non-expert viewers
- Double-Edged Sword: Risk of over-reliance on black box analysis without understanding reasoning
- Journalistic Applications:
- Timeline visualizations of game evaluations move-by-move
- Some games "flatlined" vs. others resembling "cardiac arrest"
- One tool among many - still need player interviews, expert commentary
Recent World Chess Championship Analysis
- Three Most Accurate Games in Championship History:
- Out of ~1,000 world championship games ever played
- Measured by "average centipawn loss" - hundredths of pawns lost vs. optimal computer moves
- Average losses only 2-3 centipawns (tiny slivers)
- Players essentially mimicking computer-driven preparation
- Preparation Dominance:
- Players "still in prep" - following pre-analyzed computer lines
- Jan Nepomniachtchi saw entire game on computer during preparation before playing it
- Humans transformed from performers to "hard drives" memorizing computer analysis
- Draw Problem:
- Perfect play likely leads to draws (like tic-tac-toe)
- Many high-level games drawn in "rote fashion"
- Difficult to maintain robust viewership/fandom
- Player frustration with format
- Magnus Carlsen Response: Announced might not defend title due to format frustration
Proposed Solutions to Chess Stagnation
- Speed Chess:
- Forces reliance on intuition vs. memory/prep
- Chess having "blossoming moment online"
- Classical games can last 7-8 hours, allowing extensive memory searches
- Chess960/Fischer Random:
- Randomize piece starting positions (960 possible arrangements)
- Pawns stay same, major pieces randomized within back rank
- Maintains symmetry but eliminates memorization value
- Time Pressure Creates Excitement:
- Game 6 of recent championship was "all-time classic" due to time trouble
- 3-4 blunders in 4-5 move window from time pressure
- Mistakes are human and relatable - "we like to watch that"
Backgammon - Introducing Randomness
- Roeder's Personal Obsession: Never played much before book research, now "utterly obsessed"
- Randomness Impact on Different Players:
- Beginners: Democratizing - can beat world's best ~30% of time in single game
- Elite Players: Don't think about luck - "dice fall how they fall, make best move available"
- Mochy (World's Best): Like "house at casino" - law of large numbers applies
- AI Approach Difference:
- Enormous branching factor (multiple piece movements + dice combinations)
- Traditional search/evaluate techniques ineffective
- Neural networks excel at pattern recognition (like strong human players)
- One of first successful neural network applications in games
- Training Differences:
- Chess: extensive opening theory, rote memorization possible
- Backgammon: tree becomes too wide after few moves, must rely on intuition/experience
- Historical training via "rollouts" - starting same position, playing out repeatedly
- Cheating Suspicions:
- Human response to losing at chance games: "it's rigged"
- XG Extreme Gammon and other programs include disclaimers
- Some offer "roll your own dice at home" settings
- Natural delusional response to machine losses
Bridge - Partnership and Cheating Challenges
- Communication Restrictions: Only allowed to communicate via structured auction before card play
- Cheating Epidemic:
- Easy to cheat via coughing, kicking under table, code words
- Bridge world "rocked and decimated" by ongoing scandals
- Revelation after revelation of top pairs allegedly cheating
- Online Gaming Problems:
- Partners can call each other, share hand information
- Impossible to check for this type of cheating
- AI as Solution:
- Popular format: humans partner with robots instead of other humans
- Can't call robot on phone to cheat
- Ensures "above board" contract bridge games
- AI Fate: Strongest bridge AI now lives on websites as available partner for humans
Audience Q&A Highlights
- Favorite Games:
- Scrabble (personal best skill, graduate school obsession)
- Backgammon (recent love affair, perfect luck/skill blend)
- Chess (nostalgic, played with grandpa since childhood)
- Magnus Carlsen Encounter:
- Simultaneous exhibition at UN headquarters
- Lasted 29 moves, put Carlsen in check once
- Game was over much earlier, "king almost ended up on entirely opposite side"
- Elite Player Qualities:
- Must "love the game" (beyond just liking it)
- Requires "uncountably large number of hours" devoted to study
- Every game demands extensive memorization/preparation
- Gradations between world's best players remain "inexplicable"
- Ancient Game Origins:
- Backgammon set found in ancient Egyptian burial tomb
- Chess has 1,500-year history from Indian military exercises
- Go and checkers have millennia-old ancestors
- Cultural/technological longevity of games "unparalleled"
- Ancient games now played by most advanced supercomputers
- Modern Board Games:
- Appreciates current "golden age" of board game design
- Settlers of Catan, Wingspan, Terraforming Mars mentioned
- Migration from Europe to US design centers
- "Anything that increases people playing games is to the good without limit"
Future Implications
- Human vs. Machine Perspective:
- "No such thing as human versus machine" - humans built these machines
- Like Usain Bolt not caring about race cars being faster
- Magnus Carlsen accepts AI superiority as technological fact
- Technological Achievement:
- Games-playing machines represent "extraordinary technological feat"
- Should reflect on and be proud of human collective achievement
- Real World Applications:
- Key unanswered question: what will these systems do in real world?
- Deep Blue disassembled after beating Kasparov, did "literally nothing else"
- Modern neural network techniques more adaptable/extendable
- Games as "test bed and launching pad" for AI since 1950s
- Previous AI promises failed to materialize, but "that's changing"
- Role of games in foundation for future AI will be "fascinating to watch"
