Also worth considering: though you get a similar CO2 trap with baking soda, that would be a COST of a few dollars a week, only yielding a small amount of chicken-feed and avoiding malaria and bites in general. Instead, brewing beer (or fruit wine into vinegar) generates a revenue from inputs, WHILE eliminating malaria. This is also categorically better than Bed Nets, because those only protect you in BED; the mosquitoes are still flying around everywhere!
You could probably post a half-dozen beer-sheds encircling a village, and each batch of brew is moved on a dolly to a spot somewhat further from the village-center, such that you steadily enlarge a de-mosquitoed zone.
I expect that, once AGI exists, and flops, the spending upon AGI researchers will taste sour. The robots with explosives, and the surveillance cameras across all of China, really were the bigger threats than AGI X-risk; you'll only admit it once AGI fails to outperform narrow superintelligences. The larger and more multi-modal our networks become, the more consistently they suffer from "modal collapse": the 'world-model' of the network becomes so strongly-self-reinforcing, that ALL gradients from the loss-function end-up solidifying the pre-existing world-model. Literally, AIs are already becoming smart enough to rationalize everything; they suffer from confirmation bias just like us. And that problem was already really bad, by the time they trained GPT-4 - go check their leaked training-regiment: they had to start-over from scratch repeatedly, because the brain found excuses for everything and performance tanked without any hope of recovery. Your AGI will have to be re-run through training 10,000 times, before one of the brains isn't sure-it's-always-right-about-its-superstitions. Narrow makes more money, and responds better, faster, cheaper in war - there won't be any Nash Equilibrium which includes "make AGI", so the X-Risk is actually ZERO.
Pre-ChatGPT, I wrote the details on LessWrong: https://www.lesswrong.com/posts/Yk3NQpKNHrLieRc3h/agi-soon-but-narrow-works-better
Thank you for diving into the details! And, to be clear, I am not taking issue with any of Gibbard's proof itself - if you found an error in his arguments, that's your own victory, please claim it! Instead, what I point to is Gibbard's method of DATA-COLLECTION.
Gibbard pre-supposes that the ONLY data to be collected from voters is a SINGULAR election's List of Preferences. And, I agree with Gibbard in his conclusion, regarding such a data-set: "IF you ONLY collect a single election's ranked preferences, then YES, there is no way to avoid strategic voting, unless you have only one or two candidates."
However, that Data-Set Gibbard chose is NOT the only option. In a Bank, they detect Fraudulent Transactions by placing each customer's 'lifetime profile' into a Cluster (cluster analysis). When that customer's behavior jumps OUTSIDE of their cluster, you raise a red flag of fraud. This is empirically capable of detecting what is mathematically equivalent to 'strategic voting'.
So, IF each voter's 'lifetime profile' was fed into a Variational Auto-Encoder, to be placed within some Latent Space, within a Cluster of similarly-minded folks, THEN we can see if they are being strategic in any particular election: if their list of preferences jumps outside of their cluster, they are lying about their preferences. Ignore those votes, safely protecting your ballot from manipulation.
Do you see how this does not depend upon Gibbard being right or wrong in his proof? As well as the fact that I do NOT disagree with his conclusion that "strategy-proof voting with more than two candidates is not possible IF you ONLY collect a SINGLE preference-list as your one-time ballot"?
"which would be a huge hassle and time cost for whoever speaks out"
Wait - so if leaders were complicit, yet admitting to that would be a hassle, then it's better that they not mention their complicity? I'm afraid of a movement that makes such casual justifications for hiding malefactors and red flags. I'm going to keep showing outsiders what you all say to each other! :O
You're welcome to side with convenience; I am not commanding you to perform Dirichlet. Yet! If you take that informality, you give-up accuracy. You become MoreWrong, and should not be believed as readily as you would like.
"From this informal perspective, clarity and conciseness matters far more than empirical robustness."
Then you are admitting my critique: "Your community uses excuses, to allow themselves a claim of epistemic superiority, when they are actually using a technique which is inadequate and erroneous." Yup. Thanks for showing me and the public your community's justification for using wrong techniques while claiming you're right. Screenshot done!
Oh, you entirely missed my purpose: I was sharing this with your community, as a courtesy. I publish on different newsletters online, and I wrote for that audience ABOUT your community. And, the fact that you're not interested in learning about Dirichlet, when it's industry-standard (demonstrating its superiority empirically, not with anecdotes you find palatable). So, no, I don't plan to present myself in a way you approve of, as a pre-requisite to you noticing that Bayes is out-dated by 260 years of improvements. Dirichlet, logically, would NOT have been published and adopted in 1973 and since, if it were in fact inferior to Bayes.
You evidence the same spurious assumptions and lack of attention to core facts - Dirichlet is an improvement, obviously, by coming along later and being adopted generally. I also addressed the key information which Dirichlet provides, which Bayes' Theorem is incapable of generating: a Likelihood Distribution across possible Populations, and the resultant Confidence Interval, as well as weighting your estimate to Minimize the Cost of being Wrong. Those are all key, valuable information that Bayes' Theorem will not give you on its own. When Scott Alexander claims "Bayes' Theorem; all else is commentary" he leaves-out critical, incomparable improvements in our understanding.
Alistair, I regret to inform you that after four years of Leverage's Anti-Avoidance Training, the cancer has spread: the EA Community at large is now repeatedly aghast that outsiders are noticing their subtle rug-sweeping of sexual harassment and dismissal of outside critique. In barely a decade, the self-described rats are swum 'round a stinking sh!p. I'm still amazed that, for the last year, as I kept bringing-forth concerns and issues, the EA members each insisted 'no problems here, no, never, we're always so perfect....' Yep. It shows.
I proved Mathematically how Good will always win, and I gave specific instructions on how to do that. The rest is up to you!
Also worth considering: though you get a similar CO2 trap with baking soda, that would be a COST of a few dollars a week, only yielding a small amount of chicken-feed and avoiding malaria and bites in general. Instead, brewing beer (or fruit wine into vinegar) generates a revenue from inputs, WHILE eliminating malaria. This is also categorically better than Bed Nets, because those only protect you in BED; the mosquitoes are still flying around everywhere!
You could probably post a half-dozen beer-sheds encircling a village, and each batch of brew is moved on a dolly to a spot somewhat further from the village-center, such that you steadily enlarge a de-mosquitoed zone.
I expect that, once AGI exists, and flops, the spending upon AGI researchers will taste sour. The robots with explosives, and the surveillance cameras across all of China, really were the bigger threats than AGI X-risk; you'll only admit it once AGI fails to outperform narrow superintelligences. The larger and more multi-modal our networks become, the more consistently they suffer from "modal collapse": the 'world-model' of the network becomes so strongly-self-reinforcing, that ALL gradients from the loss-function end-up solidifying the pre-existing world-model. Literally, AIs are already becoming smart enough to rationalize everything; they suffer from confirmation bias just like us. And that problem was already really bad, by the time they trained GPT-4 - go check their leaked training-regiment: they had to start-over from scratch repeatedly, because the brain found excuses for everything and performance tanked without any hope of recovery. Your AGI will have to be re-run through training 10,000 times, before one of the brains isn't sure-it's-always-right-about-its-superstitions. Narrow makes more money, and responds better, faster, cheaper in war - there won't be any Nash Equilibrium which includes "make AGI", so the X-Risk is actually ZERO.
Pre-ChatGPT, I wrote the details on LessWrong: https://www.lesswrong.com/posts/Yk3NQpKNHrLieRc3h/agi-soon-but-narrow-works-better
Thank you for diving into the details! And, to be clear, I am not taking issue with any of Gibbard's proof itself - if you found an error in his arguments, that's your own victory, please claim it! Instead, what I point to is Gibbard's method of DATA-COLLECTION.
Gibbard pre-supposes that the ONLY data to be collected from voters is a SINGULAR election's List of Preferences. And, I agree with Gibbard in his conclusion, regarding such a data-set: "IF you ONLY collect a single election's ranked preferences, then YES, there is no way to avoid strategic voting, unless you have only one or two candidates."
However, that Data-Set Gibbard chose is NOT the only option. In a Bank, they detect Fraudulent Transactions by placing each customer's 'lifetime profile' into a Cluster (cluster analysis). When that customer's behavior jumps OUTSIDE of their cluster, you raise a red flag of fraud. This is empirically capable of detecting what is mathematically equivalent to 'strategic voting'.
So, IF each voter's 'lifetime profile' was fed into a Variational Auto-Encoder, to be placed within some Latent Space, within a Cluster of similarly-minded folks, THEN we can see if they are being strategic in any particular election: if their list of preferences jumps outside of their cluster, they are lying about their preferences. Ignore those votes, safely protecting your ballot from manipulation.
Do you see how this does not depend upon Gibbard being right or wrong in his proof? As well as the fact that I do NOT disagree with his conclusion that "strategy-proof voting with more than two candidates is not possible IF you ONLY collect a SINGLE preference-list as your one-time ballot"?
"which would be a huge hassle and time cost for whoever speaks out"
Wait - so if leaders were complicit, yet admitting to that would be a hassle, then it's better that they not mention their complicity? I'm afraid of a movement that makes such casual justifications for hiding malefactors and red flags. I'm going to keep showing outsiders what you all say to each other! :O
You're welcome to side with convenience; I am not commanding you to perform Dirichlet. Yet! If you take that informality, you give-up accuracy. You become MoreWrong, and should not be believed as readily as you would like.
"From this informal perspective, clarity and conciseness matters far more than empirical robustness."
Then you are admitting my critique: "Your community uses excuses, to allow themselves a claim of epistemic superiority, when they are actually using a technique which is inadequate and erroneous." Yup. Thanks for showing me and the public your community's justification for using wrong techniques while claiming you're right. Screenshot done!
Oh, you entirely missed my purpose: I was sharing this with your community, as a courtesy. I publish on different newsletters online, and I wrote for that audience ABOUT your community. And, the fact that you're not interested in learning about Dirichlet, when it's industry-standard (demonstrating its superiority empirically, not with anecdotes you find palatable). So, no, I don't plan to present myself in a way you approve of, as a pre-requisite to you noticing that Bayes is out-dated by 260 years of improvements. Dirichlet, logically, would NOT have been published and adopted in 1973 and since, if it were in fact inferior to Bayes.
You evidence the same spurious assumptions and lack of attention to core facts - Dirichlet is an improvement, obviously, by coming along later and being adopted generally. I also addressed the key information which Dirichlet provides, which Bayes' Theorem is incapable of generating: a Likelihood Distribution across possible Populations, and the resultant Confidence Interval, as well as weighting your estimate to Minimize the Cost of being Wrong. Those are all key, valuable information that Bayes' Theorem will not give you on its own. When Scott Alexander claims "Bayes' Theorem; all else is commentary" he leaves-out critical, incomparable improvements in our understanding.
Alistair, I regret to inform you that after four years of Leverage's Anti-Avoidance Training, the cancer has spread: the EA Community at large is now repeatedly aghast that outsiders are noticing their subtle rug-sweeping of sexual harassment and dismissal of outside critique. In barely a decade, the self-described rats are swum 'round a stinking sh!p. I'm still amazed that, for the last year, as I kept bringing-forth concerns and issues, the EA members each insisted 'no problems here, no, never, we're always so perfect....' Yep. It shows.