3 ms·
> His third point is more interesting: at scale, even 99.9% accuracy is going to leave hundreds of thousands of mistakes a day. No one will care about the 99.9%
by klyrs 6y ago
> His third point is more interesting: at scale, even 99.9% accuracy is going to leave hundreds of thousands of mistakes a day. No one will care about the 99.9% you got right, only the things you missed. Reporters will call public attention to these.
This is indeed the crux. The .1% failures are quite inevitable. They represent the most valuable training data, and that deserves attention. The breakdown that I see is that it becomes a popularity contest: if you have enough friends, followers, or you can bend the ear of somebody who does, then and only then do these cases receive human attention. In a sense, this is outsourcing the human-level moderation to the public, a level of bad PR determines the escalation. This will fail to capture a large portion of the false positives, fail to feed back into the training set, and the number of failures will continue to scale with the system. The goal should be to eradicate these mistakes, and I can't see how that can be done without investing in human metamoderation -- even if they catch 1% of the .1%, that could provide an opportunity to continuously improve the system. If your profits scale, then so must your investments, or as you say, public confidence will collapse.
- searchableguy 6y agoYou described the problem well. How would you convince someone to increase budget for moderation without bad PR?
- klyrs 6y agoRepeated bad PR stories erodes public confidence, which can result in regulations? My reasons include altruism and a long term view, which don't always play well with business types.