I will bet 500 karma that a funny picture thread will appear on Less Wrong within one year. If anyone is interested in the bet, we can better define terms.
Right now the LW software doesn't support karma transfers. Until it does and we can develop a more robust prediction market, let's just record the karma transfers on the wiki page that already exists for this purpose.
I will also give 100 karma to anyone that donates $10 to the SIAI before the current fundraising campaign is over.
10,000 karma for the first person with a karma transfer source code patch?
All the benefits of betting "even fake currency" that you name: I absolutely agree -- particularly for people who are self-consciously interested in rationality.
But I think you overestimate the marginal utility of Less Wrong karma (in the amounts that frequent contributors to this site tend to possess it) versus money. Fundamentally, absolutely nothing changes for a person who loses most of their Less Wrong karma, except the ability to make more ill-considered wagers. We can't expect top-notch rationality skills from people when those are the stakes. So as I say: better than bullshitting on Usenet, less good than a true betting market.
Also, I have noticed that offers to wager are a subtle applause light on Less Wrong, garnering substantial karma at times for comments that consist of a prediction and an offer to back it up with a bet. If I am right about that, using karma as betting currency obviously somewhat changes the nature of the decision to propose a bet.
These considerations lead me to think that the result would depend too heavily on the good will of the commenters on the site. Again, though, I think you're mostly right.
Expanding somewhat on the problem of not using real money - if people lose all their money they can't bet any more. This combats obstinance. If people are simply too stubborn to update on evidence then they will no longer be able to bet. Their poor judgement is removed from the market. The only way this would come in to play on lesswrong is if the bets started to become seriously large. If the system was karma transfer (which I recommended against) then poorly calibrated predictors would lose their karma reserves. This would benefit the quality of large st... (read more)