I haven’t thought for more than a minute or so about each of these, so they are pretty shitty in both resolution conditions and calibration/accuracy, but I’m putting them up anyway since it’s better than nothing (and also because that way maybe someone else will come along and improve on them):
1. OpenAI total annual revenue from large neural nets either remains unknown or is credibly estimated to be less than $100M: 70%
3. Some sort of Diplomacy AI (with press) milestone is achieved: 50%
4. Trillion-parameter dense models trained on at least 200b tokens exist: 70%
5. At least a million people are involved in some sort of political or ideological AI persuasion learning process (as opposed to e.g. advertising or engagement-maximizing) such as those described here and here: 30%
6. At least a million Americans are involved in… etc.: 15%
What does it mean to be “involved in”? @ike suggested that a deepfake could go viral. Suppose there was a deepfake that had a political purpose, like a fake video involving a politician. If a million people view it, is that “involved”?
Good question. I meant involved in some sort of automated learning process, such as RL. Like, data is automatically collected about the million humans’ behavior and then automatically used to adjust the weights of some neural network that feeds those humans information, repeat.
Yes. For example, recommendation algorithms count as this I think, except that they aren’t currently being optimized to persuade towards anything ideological or political; they merely persuade you to be more engaged, buy more products, report higher user satisfaction, etc.
I haven’t thought for more than a minute or so about each of these, so they are pretty shitty in both resolution conditions and calibration/accuracy, but I’m putting them up anyway since it’s better than nothing (and also because that way maybe someone else will come along and improve on them):
1. OpenAI total annual revenue from large neural nets either remains unknown or is credibly estimated to be less than $100M: 70%
2. MMLU SOTA above 70% like Brundage says: 50%
3. Some sort of Diplomacy AI (with press) milestone is achieved: 50%
4. Trillion-parameter dense models trained on at least 200b tokens exist: 70%
5. At least a million people are involved in some sort of political or ideological AI persuasion learning process (as opposed to e.g. advertising or engagement-maximizing) such as those described here and here: 30%
6. At least a million Americans are involved in… etc.: 15%
What does it mean to be “involved in”? @ike suggested that a deepfake could go viral. Suppose there was a deepfake that had a political purpose, like a fake video involving a politician. If a million people view it, is that “involved”?
Good question. I meant involved in some sort of automated learning process, such as RL. Like, data is automatically collected about the million humans’ behavior and then automatically used to adjust the weights of some neural network that feeds those humans information, repeat.
So there needs to be some two way interaction between the AI system and real humans?
Yes. For example, recommendation algorithms count as this I think, except that they aren’t currently being optimized to persuade towards anything ideological or political; they merely persuade you to be more engaged, buy more products, report higher user satisfaction, etc.