The source is here. I'll restate the problem in simpler terms:
You are one of a group of 10 people who care about saving African kids. You will all be put in separate rooms, then I will flip a coin. If the coin comes up heads, a random one of you will be designated as the "decider". If it comes up tails, nine of you will be designated as "deciders". Next, I will tell everyone their status, without telling the status of others. Each decider will be asked to say "yea" or "nay". If the coin came up tails and all nine deciders say "yea", I donate $1000 to VillageReach. If the coin came up heads and the sole decider says "yea", I donate only $100. If all deciders say "nay", I donate $700 regardless of the result of the coin toss. If the deciders disagree, I don't donate anything.
First let's work out what joint strategy you should coordinate on beforehand. If everyone pledges to answer "yea" in case they end up as deciders, you get 0.5*1000 + 0.5*100 = 550 expected donation. Pledging to say "nay" gives 700 for sure, so it's the better strategy.
But consider what happens when you're already in your room, and I tell you that you're a decider, and you don't know how many other deciders there are. This gives you new information you didn't know before - no anthropic funny business, just your regular kind of information - so you should do a Bayesian update: the coin is 90% likely to have come up tails. So saying "yea" gives 0.9*1000 + 0.1*100 = 910 expected donation. This looks more attractive than the 700 for "nay", so you decide to go with "yea" after all.
Only one answer can be correct. Which is it and why?
(No points for saying that UDT or reflective consistency forces the first solution. If that's your answer, you must also find the error in the second one.)
Damn, another one of my old comments and this one has a mistake. If we hold all of the other individuals fixed on the tails plan, then there's a 100% chance that if you choose heads that no money is donated.
But also, UDT can just point out Bayesian updates only work within the scope of problems solvable by CDT. When agents' decisions are linked, you need something like UDT and UDT doesn't do any updates.
(Timeless decision theory may or may not do updates, but can't handle the fact that choosing Yay means that your clones choose Yay when they are the sole decider. If you could make all agents choose Yay when your were a decider, but all choose Nay when you weren't you'd score higher on average, but of course the linkage doesn't work this way as their decision is based on what they see, not on what you see. This is the same issue that it has with Counterfactual Mugging).