Human vs. Machine

[Hornbein]

It is well known that any human no longer has a chance in traditional chess, go, or poker. Recently champ Shin-Jinseo was able to defeat KataGo after granting himself a one-stone advantage. That is, he places two stones before KataGo places one. This is called a two-stone handicap -- I'd call it one stone plus the first move but I suppose that's too prolix. He was able to maintain this advantage all the way to victory, winning two out of three after losing the first game.

The next triumph for AI was the strategic combat game Starcraft II, where the enemy's plans are hidden. AI was about equal with the top players but went on to bigger things before achieving dominance. Recently a man was able to defeat an AI after having lost 70 straight games. He was able to find enough weaknesses -- you could say obscure bugs -- that he was at last able to win. AIs of this kind aren't allowed to learn -- that would send them off into some unknown direction -- so I suppose he could beat it handily after that. The same thing goes on with the notoriously weak AIs of bridge, which in certain situations always make the same mistakes. Players experienced against robots learn to provoke such mistakes.

Some bridge players prefer to believe that their game is special but I'm much more inclined to think that there isn't enough money in bridge to attract a serious effort.

Is there any game in which AI can't dominate? They say not yet for chess with an infinite boundaryless board. I haven't looked into that.

 

[DaveC426913]

Is there any game in which AI can't dominate?

Basketball? Tag? Red Rover? Frisbee?

AI would absolutely suck at any of those, since it just sits in a data centre and broods.


I'm only being somewhat facetious. Is it possible you are self-limiting your study to games that AI could excel at?

 

[berkeman]

Basketball?

Tsinghua University High School in China, a benchmark for campus basketball in the country, was visited this week by American basketball star Steph Curry, who rose to the challenge of shooting hoops against a Chinese humanoid robot.

https://newsus.cgtn.com/news/2026-0...hoops-with-robots-in-China-1QmQXB6CLxS/p.html

 

[DaveC426913]

Ironically, ChatGPT has some interesting suggestions about what kinds of games it might suck at. (I was going to transcribe and summarize this list, but the explanation s are enlightening, so I've decided to paste it verbatim):

• Diplomacy — probably one of the strongest examples. You have to negotiate, form alliances, lie, predict who will betray whom, and coordinate actions simultaneously. Pure calculation isn't enough.
• The Resistance / Avalon — almost entirely social deduction. AI has to infer who is lying from inconsistent statements and behaviour, while itself deliberately deceiving the other players.
• Mafia / Werewolf — similar, but even more dependent on conversational subtleties, reputation, group dynamics, and exploiting what other people believe about you.
• Charades — surprisingly interesting. The AI has to understand an ambiguous physical performance and infer the intended concept.
• Pictionary — requires extracting a concept from an extremely noisy, incomplete visual representation.
• Keep Talking and Nobody Explodes — particularly difficult if the AI has to communicate with a human rather than simply solve the puzzle itself. The challenge is coordinating two different information sets.
• Telephone — if played seriously, it tests whether an AI can preserve meaning through deliberately noisy communication.
• 20 Questions with an unusual human questioner — not necessarily difficult because of the information theory, but because humans give inconsistent, metaphorical, culturally dependent answers.
• Cards Against Humanity — if the criterion is understanding what will make a particular group of humans laugh. Generating funny combinations is easier than predicting a specific audience's reaction.
• A game of improvisational storytelling — where players must maintain characters, remember arbitrary details, react to unexpected developments, and deliberately pursue conflicting objectives.

A particularly nasty one: Hide-and-seek​

Not the computer-game version. Actual hide-and-seek.

An AI playing against humans in a real building has to reason about:

• where humans would think it would search;
• whether a hiding place is too obvious;
• whether somebody has already searched an area;
• whether another player saw someone enter a room;
• sound and movement;
• imperfect visual information;
• deception;
• the fact that humans don't behave optimally.

It's a lovely example because the rules are trivial but the problem is extraordinarily open-ended.

 

[FactChecker]

A related question:
In general, how well does AI handle games where a winning strategy must include some random behavior to keep the opponent guessing? Many of the above mentioned games require AI to react to unpredictable moves, but what if AI must make unpredictable moves?
ADDED: Are there such games? My knowledge of games is very limited.

 

[Borg]

Diplomacy — probably one of the strongest examples. You have to negotiate, form alliances, lie, predict who will betray whom, and coordinate actions simultaneously. Pure calculation isn't enough.

There is a python library for Diplomacy and I recently saw a presentation from a group that is working on playing exactly this game using LLMs. I spoke with them afterward since my current work aligns with theirs in several ways related to long-term strategy.

 

[Hornbein]

A related question:
In general, how well does AI handle games where a winning strategy must include some random behavior to keep the opponent guessing? Many of the above mentioned games require AI to react to unpredictable moves, but what if AI must make unpredictable moves?
ADDED: Are there such games? My knowledge of games is very limited.

Starcraft II. If the enemy can predict what you will do it will foil your plan. You can't do the same thing every time.

Poker, in which AI excels. It's better at bluffing and detecting bluffs. "I felt like it could see my hole card."

 

[Matterwave]

This is called a two-stone handicap -- I'd call it one stone plus the first move but I suppose that's too prolix.

It's a two stone handicap because the AI doesn't get komi (it got 0.5 just to break ties). A 6.5-7.5 komi is typically added to white's score to accommodate Black's first move advantage. A one stone handicap is if Shin Jinseo simply plays black with no komi.

The next triumph for AI was the strategic combat game Starcraft II, where the enemy's plans are hidden. AI was about equal with the top players but went on to bigger things before achieving dominance.

You might be referring to AlphaStar from DeepMind here? The problem with AI playing real time games is how you normalize for reaction speed. The AI was artificially "slowed down" (APM limited) so that it wouldn't just win on its ability to execute actions frame-by-frame.

But there's no great, universal, measure for what kind of slow down is warranted.

Starcraft has a lot of strategic elements to it beyond raw reaction speed and mouse control (vs. for example CounterStrike where AI would just be aimbots) but it still does have those dexterity-related aspects that's non trivial to pin down.

AIs of this kind aren't allowed to learn

At least AlphaStar came from a learning algorithm. I'm not sure if they continued learning even as they were playing the world championship games. But the decision to stop learning was more related to "we're done here, we don't want to spend more money on this" than any technical limitations.

Is there any game in which AI can't dominate?

Note you are grouping many many different systems under one umbrella term "AI" here. Take a chess engine and try to have it play Go and see where that takes you. ChatGPT chess ELO is not top grandmaster yet.

 

[Matterwave]

A related question:
In general, how well does AI handle games where a winning strategy must include some random behavior to keep the opponent guessing?

This is no problem for computers since Game Theory Optimal strategies (which can be stochastic, depending on the game) will beat any other strategy in the long term. This is the case for Poker (which was given as an example).

A computer system could be given access to a (high quality) RNG to perform a truly stochastic strategy, but I would guess that just the simple RNGs implemented in whichever code library of your choice would already be good enough.

 

[DaveC426913]

I looked them up so y'all don't have to:

RNG

random number generator.

chess ELO

number that estimates a player's skill level based on their results against other players

komi

scoring bonus in board games and Japanese wordplay

 

[Matterwave]

Sorry! I forget about explaining lingo sometimes.

APM = Actions per Minute.

 

[Hornbein]

Note you are grouping many many different systems under one umbrella term "AI" here. Take a chess engine and try to have it play Go and see where that takes you. ChatGPT chess ELO is not top grandmaster yet.

AlphaZero became both chess and go expert in just a few days.

 

[Matterwave]

AlphaZero became both chess and go expert in just a few days.

Sure, they wanted to show one system that excelled at multiple games. Were you referring to specifically AlphaZero in your question?

 

[Hornbein]

Sure, they wanted to show one system that excelled at multiple games. Were you referring to specifically AlphaZero in your question?

Merely an interesting factoid. It was a revolutionary engineering accomplishment.