This is a special post for quick takes by Benji Birch. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
If there were 1 AI agent that had the power and wanted to take over the world at a fixed date, and calculated that exact date to attempt takeover of the human race, it would attempt to do that on that date, but if there were 10,000 of them and they all knew of each other, they should execute their plan earlier with a much lower chance of succeeding because they have to compete against the other AIs
So if this were a game the AIs would have to play in a way that sabotages other AIs, make coalitions that appear mutually beneficial to the opponent (actual mutually beneficial coalitions included), leverage humans, play with high risk, high reward in order to maximise chances of winning (for example, 99% chance of immediate failure is better if the chance of winning is greater than low risk survival strategies)
But most of this doesn't seem to matter too much if risky strategies are favoured, and that only one has to act in order to be detected as a threat to humanity, and humanity starts working against the AIs, so we have to consider the variance of how "risk taking" each agent is, and in the calculated probabilities of success. Another thing to add is that they should assume they're in the middle of the distribution of "risk taking", so the ones which take the most risk think that there must be others which take more risk - the one on the furthest end of the distribution is most likely to do something with a very low probability of succeeding and therefore warn humanity.
I'm wondering what other people think of this, what might actually happen in this scenario, and what more things are there to consider here