I haven't read the book so I don't know the context of the lottery thing, but I think that it might be worth thinking about in the context of multi-stage selection processes. If a later stage in the process (whether AI, human, or whatever) only ever sees pre-screened inputs then I think there's likely a greater chance of idiosyncratic biases being a factor at that stage. There might be something good about supplementing shortlists with some unscreened random options (could also signal problems with the shortlist generators if an unscreened option turns out to be better than the screened ones).
Yeah, I think a multi-stage selection process, with a lottery/randomness at a later stage, could work if we really have little predictive ability beyond an initial screening process to get rid of those cases that simply don't belong. But if we still have predictive power after low-level screening, then let's predict the best that we can!
The analogy I'm thinking of is circuit design, where the reason for incorporating feedback or feedforward elements isn't exactly about worrying that the internal elements don't work at all but more about making the overall system more robust and less sensitive to the tuning or minor variations of the individual components than it would be in a strictly serial design.
The reason I won't ever buy a book on AI is that it will be at least 3 editions out of date from the moment it the manuscript starts to be edited until final release. It's moving that fast.I find podcasts are the only way to keep up.
New reasoning models vastly improve accuracy but it won't stop the problem of garbage in-garbage out. And we should humbly admit that most of that garbage is human generated!!
I haven't read the book so I don't know the context of the lottery thing, but I think that it might be worth thinking about in the context of multi-stage selection processes. If a later stage in the process (whether AI, human, or whatever) only ever sees pre-screened inputs then I think there's likely a greater chance of idiosyncratic biases being a factor at that stage. There might be something good about supplementing shortlists with some unscreened random options (could also signal problems with the shortlist generators if an unscreened option turns out to be better than the screened ones).
Yeah, I think a multi-stage selection process, with a lottery/randomness at a later stage, could work if we really have little predictive ability beyond an initial screening process to get rid of those cases that simply don't belong. But if we still have predictive power after low-level screening, then let's predict the best that we can!
The analogy I'm thinking of is circuit design, where the reason for incorporating feedback or feedforward elements isn't exactly about worrying that the internal elements don't work at all but more about making the overall system more robust and less sensitive to the tuning or minor variations of the individual components than it would be in a strictly serial design.
The reason I won't ever buy a book on AI is that it will be at least 3 editions out of date from the moment it the manuscript starts to be edited until final release. It's moving that fast.I find podcasts are the only way to keep up.
New reasoning models vastly improve accuracy but it won't stop the problem of garbage in-garbage out. And we should humbly admit that most of that garbage is human generated!!