Before 2024, will Manifold come up with a way to play chess (and games in general) that I won't consider exploitable?
resolved Jan 6

Resolves to Yes if Manifold users coordinate to play games and win using a prediction market structure that I will consider non-exploitable

Get Ṁ1,000 play money

🏅 Top traders

#NameTotal profit
Sort by:

Off topic.

If you like looking for exploits, could you please check this game and tell me how these 2 players could exploit?

I am sure they won't, but would like to get a critic

@KongoLandwalker Off topic, but eg cooperating to make one of the sides seem to lose while secretly betting on the other side; then winning by the previously seemingly losing side

sold Ṁ19 of NO

Your verdict for Manifold Plays Chess 3? Also pinging @ms

Its a new prediction market structure involving conditional prediction markets. And we managed to win.

predicted NO

@harfe My viewpoint: It was quite exploitable, as I demonstrated in some of the early markets (e.g. move 7 where I used the exact strategies I predicted would be potentially problematic - build a big stake, manipulate the move market probability, and manipulate the move probabilities of subsequent move markets to get payouts on the previous ones.

I chose to force a win, mostly because it was more fun that way, but I'm reasonably confident that I could have forced a loss if I had chosen. Additionally, I also could and did manipulate to force specific moves that I preferred for profit or fun reasons, even if they weren't the best moves.

sold Ṁ26 of YES

I plan to run another round of Manifold Plays Chess:

I do not claim that this market structure is entirely unexploitable. But I believe it would work well if Manifold markets were efficient (which they are not, as far as I can tell). After a couple of moves have been played, I would be curious for your opinion.

bought Ṁ50 of YES

I could probably come up with something, but I still need to figure out details. I might come back to this in 2 months or so.

bought Ṁ25 of YES

My YES votes are not contingent on this, but I think you should clarify how much oracle assistance (i.e. stockfish) is allowed in the evaluation criteria.

For example: it would be really easy to win with "What's the best chess move? (Resolves to stockfish)". But I wouldn't find that interesting, because most of the information is coming from the computer, not the market.

@jack's proposal uses a more grey-area proposal, where "weak stockfish" generates plausible moves and evaluates the position five moves later. The idea is that simple heuristics could be generalized to games where we don't have a strong evaluator. The easy problem is easy to solve, and the market can leverage that to solve a much harder problem.

I think we can solve the hard case where there is no oracle, and only available signal is the "Will X win?" markets, and derivatives. I do want to keep tinkering with these structures, and I'll keep doing so as much as I can without causing the idea of "Manifold Plays Chess" to grow stale.

predicted NO

@citrinitas I think it would be great to try some other game where we don't have super strong engines (i.e. oracles). In chess, the predictors will make use of them regardless of whether the market design makes use of them.

A variant of my proposal that doesn't use any oracles to resolve the markets is to have a main market "Will manifold win the game?", have people propose moves and each proposed move gets a market that resolves to the win probability on the main "Will manifold win?" market 5 moves later, for example. There is definitely some exploitability here, but I think it's less bad compared to the past 2 attempts.

@citrinitas an AGI evaluating the chances of winning at any game (and things like predicting the chances it will output) don't count as market structures that play games and aren't exploitable (because you don't play games, you use the output of something externally powerful at games). An AGI that scans the brains of all the market participants and bans those who manipulate the market so that, e.g., everyone honestly evaluates the probabilities of winning conditional on different moves counts as a non-exploitable way for Manifold to play games.

(a solution that relies on the chess being solved by stockfish doesn't count.

btw, a couple of solutions that involve predicting the chances stockfish will give in 5 moves are exploitable.)

I hope this clarifies it.

(a good enough oracle that uses the market data to produce the probabilities that a market with honest and aligned participants would produce would make me resolve this question to Yes)

bought Ṁ30 of NO

Here's my proposal for conditional markets:

But I think it is next-to-impossible to make it completely non-exploitable.

predicted NO

If someone can figure out a way to make it highly robust to exploits, I would be very impressed.

One man's exploit is another's clever meta-strategy . . .

More related questions