Also worth noting, since I've been learning a lot about how transformers and LLMs work the past few days and am seeing everything in relation to that (https://www.reddit.com/r/PhilosophyMemes/s/zsZjEkcg2B): at the end of a run, the transformer gets a probability distribution over all possible tokens of which comes next. Using a "greedy" selection algorithm, where it always picks the absolute highest most likely word to go next, actually doesn't give as good results in practice as using some other algorithm to pick from among the top choices. The "temperature" parameter modulates how far from the highest probability token it will go
And another thing, I would prefer sortition over elections in politics. The person who is best at winning elections is not necessarily the best at governing or decision making or picking good advisors. I trust neither myself nor my fellow citizens to figure out the best policies, figure out the best candidates for said policies, and actually vote for them. The wisdom of the masses is great for counting jellybeans, I'm not so convinced for extremely complex unsolved (maybe unsolvable) problems.
This take is hotter than I expected when I started writing, note that I hold this tentatively ("I trust neither myself nor...")
"hiring one fewer grantmaker usually means those millions will just sit in an account for another year rather than being deployed to useful ends. And when a strong candidate turns down a CG offer, the result is often not “a slightly-less-good grantmaker," it’s just one fewer grantmaker. We routinely close rounds with fewer hires than we'd planned for."
Just an example of this dynamic, imo. See also the comments on that thread.
Agreed. So many of the problems we're seeing today arise from that runaway optimization curve. That's going to be the second target for UTW to address, if we're successful in stabilizing the US government.
"set a reasonable bar that gets you what you want, and do not discriminate above that bar. Take the first guy that clears it,"
I think this is how most hiring works, in practice. I used to think it was a diligent search for "the best person for the job". But from the hiring side it's more like "oh, you meet the criteria? great, welcome aboard." There's a competing force against optimizing for lowest pay or most number of skills: if you wait to hire someone, that's time you don't have that position filled.
I'd distinguish optimizing value *capture* from optimizing value *creation*.
If the $7/hr guy stacks shelves just as well as the $9/hr guy, you've captured more of the same pie. If they stack less well, you've captured more of a smaller pie. I'd quibble with "immoral", but like, yeah.
(First approximations!)
Finding the best candidate for a position feels different to me, because (in theory) the best candidate will create more value. But adding hoops also destroys value, not just for the person who gets the job but for everyone who interviews. Is it worth it, from a gods-eye view? My default assumption is "eh, if this was net-negative value then one of the parties to the transaction would quit".
...but there's lots of ways that might not be true, including
* Some jobs are about "who's the best at capturing value", not about creating it
* "Getting the position" is less valuable for the candidate than they think, and/or their chances are lower
* The person offering the position is mistaken about how good the filters are
* The hoops might just exist to make the position more prestigious
* Maybe there's a story where it's worth it for you to set up hoops find the best candidate, and worth it for the candidates to jump through your hoops, but you're imposing costs on people who don't even interview through signalling or something? I'm default-skeptical of this kind of thing being large compared to first-order effects, but not ruling it out.
Also worth noting, since I've been learning a lot about how transformers and LLMs work the past few days and am seeing everything in relation to that (https://www.reddit.com/r/PhilosophyMemes/s/zsZjEkcg2B): at the end of a run, the transformer gets a probability distribution over all possible tokens of which comes next. Using a "greedy" selection algorithm, where it always picks the absolute highest most likely word to go next, actually doesn't give as good results in practice as using some other algorithm to pick from among the top choices. The "temperature" parameter modulates how far from the highest probability token it will go
And another thing, I would prefer sortition over elections in politics. The person who is best at winning elections is not necessarily the best at governing or decision making or picking good advisors. I trust neither myself nor my fellow citizens to figure out the best policies, figure out the best candidates for said policies, and actually vote for them. The wisdom of the masses is great for counting jellybeans, I'm not so convinced for extremely complex unsolved (maybe unsolvable) problems.
This take is hotter than I expected when I started writing, note that I hold this tentatively ("I trust neither myself nor...")
https://forum.effectivealtruism.org/posts/B6d8Wzk4gNzHsXvdi/ai-safety-is-extremely-bottlenecked-on-grantmakers
"hiring one fewer grantmaker usually means those millions will just sit in an account for another year rather than being deployed to useful ends. And when a strong candidate turns down a CG offer, the result is often not “a slightly-less-good grantmaker," it’s just one fewer grantmaker. We routinely close rounds with fewer hires than we'd planned for."
Just an example of this dynamic, imo. See also the comments on that thread.
Agreed. So many of the problems we're seeing today arise from that runaway optimization curve. That's going to be the second target for UTW to address, if we're successful in stabilizing the US government.
"set a reasonable bar that gets you what you want, and do not discriminate above that bar. Take the first guy that clears it,"
I think this is how most hiring works, in practice. I used to think it was a diligent search for "the best person for the job". But from the hiring side it's more like "oh, you meet the criteria? great, welcome aboard." There's a competing force against optimizing for lowest pay or most number of skills: if you wait to hire someone, that's time you don't have that position filled.
I'd distinguish optimizing value *capture* from optimizing value *creation*.
If the $7/hr guy stacks shelves just as well as the $9/hr guy, you've captured more of the same pie. If they stack less well, you've captured more of a smaller pie. I'd quibble with "immoral", but like, yeah.
(First approximations!)
Finding the best candidate for a position feels different to me, because (in theory) the best candidate will create more value. But adding hoops also destroys value, not just for the person who gets the job but for everyone who interviews. Is it worth it, from a gods-eye view? My default assumption is "eh, if this was net-negative value then one of the parties to the transaction would quit".
...but there's lots of ways that might not be true, including
* Some jobs are about "who's the best at capturing value", not about creating it
* "Getting the position" is less valuable for the candidate than they think, and/or their chances are lower
* The person offering the position is mistaken about how good the filters are
* The hoops might just exist to make the position more prestigious
* Maybe there's a story where it's worth it for you to set up hoops find the best candidate, and worth it for the candidates to jump through your hoops, but you're imposing costs on people who don't even interview through signalling or something? I'm default-skeptical of this kind of thing being large compared to first-order effects, but not ruling it out.