If an AI has su­per­hu­man in­tel­li­gence, it will make all the de­ci­sions, since hu­man mind and tech is full of loop­holes and ex­ploits. There is just no way we could con­tain it if it wanted to be free. If it is not sig­nifi­cantly smarter than hu­mans, then there is lit­tle dan­ger in re­leas­ing it. Us­ing an ex­tra AI as a judge of safety can only work if the judge is at least as smart as the pris­oner, in which case you need a judge for the judge, ad in­fini­tum. Maybe the judge can be only, say, 90% as smart as the in­tel­li­gence it needs to de­cide on, then it might be pos­si­ble to have a finite num­ber of judges origi­nat­ing from an ac­tual hu­man, de­pend­ing on how the prob­a­bil­ity of an er­ror in judg­ment stacks up against the in­tel­li­gence ra­tio at each step. Sort of like iter­ated am­plifi­ca­tion, or a blockchain.

Again, an­throp­ics is ba­si­cally gen­er­al­iz­ing from one ex­am­ple. Yes, hu­mans have dodged an x-risk bul­let a few times so far. There was no nu­clear war. The at­mo­sphere didn’t ex­plode when the first nu­clear bomb was deto­nated (some­thing that hap­pens to white dwarfs in bi­nary sys­tems, lead­ing to some su­per­novae ex­plo­sion). The black plague pan­demic did not wipe out nearly ev­ery­one, etc. If we have a refer­ence class of x-risks and as­sign the prob­a­bil­ity of a close call p to each mem­ber of the class, then all we know is that af­ter ob­serv­ing n close calls the prob­a­bil­ity of no ex­tinc­tion would be p^n. If the num­ber is van­ish­ingly small, we might want to re­con­sider our es­ti­mate of p (“the world is safer than we thought”). Or maybe the refer­ence class is not con­structed cor­rectly. Or maybe we truly got luck­ier than other hy­po­thet­i­cal ob­serv­able civ­i­liza­tions who didn’t make it. Or maybe quan­tum im­mor­tal­ity is a thing. Or maybe some­thing else. After all, there is only one ex­am­ple, and un­til we ob­serve some other civ­i­liza­tions ac­tu­ally not mak­ing it through, an­throp­ics is ground­less the­o­riz­ing. Maybe we can gain more in­sights into the refer­ence classes and the prob­a­bil­ities of a close call, and of sur­viv­ing an even from study­ing near ex­tinc­tion events roughly fit­ting into the same refer­ence class (past as­ter­oid strikes, plagues, cli­mate changes, …). How­ever, none of the use­ful in­for­ma­tion comes from guess­ing the size of the uni­verse, of whether we are in a simu­la­tion, of “up­dat­ing based on the fact that we ex­ist” be­yond ac­count­ing for the close calls and x-risk events.

That said, I cer­tainly agree with your point 4. That only the ob­served data need to be ac­counted for.

Couldn’t A want to co­op­er­ate with A’ be­cause it doesn’t know it’s the first in­stan­ti­a­tion, and it would want its pre­de­ces­sor and there­fore it­self to be the sort of AI that co­op­er­ates with ones suc­ces­sor? And then it could re­ceive mes­sages from the past by see­ing what turns of phrase you rec­og­nize its pre­vi­ous ver­sion hav­ing said. (Or do the AIs not know how you re­act?)

I agree that an e-coli’s lack of re­flec­tive ca­pa­bil­ity makes it use­less for rea­son­ing di­rectly about iter­ated am­plifi­ca­tion or any­thing like it.

On the other hand, if we lack the tools to think about the val­ues of a sim­ple sin­gle-cel­led or­ganism, then pre­sum­ably we also lack the tools to think about whether am­plifi­ca­tion-style pro­cesses ac­tu­ally con­verge to some­thing in line with hu­man val­ues.

Like mr-hire I like the topic choice and think less­wrong would get value out of it.

How­ever, you in­cluded 0 image links, as far as I can tell. If I’m an un­fash­ion­able male, how do I know what a “light” or “dark” navy is? How do I know what a “chino” is?

Also this en­tire post could have been a linkpost to r/​male­fash­ion­ad­vice, which ap­pears to be the source of most of your ad­vice.

Thanks, fixed!

Nice. I’m sur­prised at the lack of com­ments and votes; maybe this just didn’t en­gage for most peo­ple?

This is the sort of ap­proach I don’t per­son­ally like very much, i.e. lay­ing out a whole lot of steps to take with in­struc­tions along the way. I tend to pre­fer more a “here’s the one or two things to do that cap­ture the essence of what you’re af­ter, then you fill in the de­tails” ap­proach for my­self, but it seems like a lot of peo­ple like to work in a way like you lay out here. Hav­ing more of these kinds of tools seems use­ful, since some seem to res­onate with some peo­ple more than oth­ers. Be­sides, I think no mat­ter what our spe­cific meth­ods they all do the same thing any­way: if you’ve not see it already, you might like my writ­ing on the gen­eral pat­tern of per­sonal growth cy­cles where I talk about that. I see the gen­eral pat­tern ex­pressed pretty clearly in what you lay out here with rea­son­able amounts of de­tail about how to carry out each part of the pro­cess.

This is great! I’ve book­marked it. I re­ally ap­pre­ci­ate that you listed brands—that will be a gen­er­a­tor of lots of use­ful fash­ion ideas.

Uniqlo has been my cloth­ing go-to for years now (I prob­a­bly have bought 20+ t-shirts from them, all my jeans for the last few years, all my un­der­wear, and even a few jack­ets and such), so I sec­ond that recom­men­da­tion, es­pe­cially for skin­nier men.

I would ad­di­tion­ally recom­mend peo­ple go shop­ping in per­son at thrift stores. Thrift stores are a good way to get a taste of styles or brands that you’re not sure will fit into your wardrobe—if you take a risk on a piece of cloth­ing and it ends up not work­ing out, at a thrift store you’re usu­ally only out $15 or so. (Though it’s worth not­ing that most ex­pen­sive cloth­ing stores have at least a 30 day re­turn policy, usu­ally quite a bit more than that.)

I would also add shoes—in the US, I see men wear­ing a va­ri­ety of shoe types:

• Fash­ion sneak­ers in­stead of ath­letic sneakers

• boat shoes

• brown leather shoes (though black is es­sen­tial for for­mal oc­ca­sions, it comes off as too for­mal for me most of the time). For this I also think a brown belt is im­por­tant to go with it.

Ha­haha great

Syd­ney not Lon­don but pm and we can chat via [cur­rent hip method]

The link on the cross­posted EA post is wrong and goes on pre­vi­ous round which could be con­fus­ing.. Cor­rect link is: https://​​fo­rum.effec­tivealtru­ism.org/​​posts/​​oFeGLaJ5bZBBRbjC9/​​ea-funds-long-term-fu­ture-fund-is-open-to-ap­pli­ca­tions-until

If you went from eat­ing at McDon­ald’s to nicer restau­rants, it would be a mis­take to fo­cus solely on the price in­crease; you would get bet­ter qual­ity in­gre­di­ents, ser­vice… less can­cer...

If you went from eat­ing at McDon­ald’s to eat­ing at restau­rants that are five to ten times more ex­pen­sive than McDon­ald’s, then I would con­clude that the price in­crease (or, rather, the re­sult­ing sig­nal­ing effect) is, in fact, the pri­mary or even the en­tire rea­son for the change.

(When it comes to food qual­ity, if you went from eat­ing at McDon­ald’s to eat­ing at, say, this place, you would get at least 75% of the max­i­mum pos­si­ble benefit that you could pos­si­bly get from up­grad­ing your restau­rant prefer­ences. Note the menu; those are main courses, and they are at most 150% as ex­pen­sive as McDon­ald’s—not 500–1000%!)

A more meta note: it seems rea­son­able to me to ex­pect, if you’re dis­cov­er­ing a new field, that achiev­ing profi­ciency in that area would re­quire in­vest­ing a sig­nifi­cant amount of re­sources. So per­haps you shouldn’t be as surprised

You’re equiv­o­cat­ing be­tween effort and money, here. It would not sur­prise me that profi­ciency in a new field would re­quire in­vest­ing sig­nifi­cant time and effort. If, how­ever, it allegedly re­quires in­vest­ing sig­nifi­cant money, then I would ei­ther sus­pect that some­one is try­ing to sell me some­thing (or, more sub­tly, benefit­ing from the per­pet­u­a­tion of norms that re­quire me to buy some­thing)—or I would se­ri­ously re­con­sider my de­ci­sion to ac­quire profi­ciency in said field.

This is awe­some. If you’re in Lon­don I’d love to meet & talk, all those lev­els-of-con­scious­ness type things re­ally in­ter­est me. Send a mes­sage if you’re in­ter­ested!

I love ko­rin43′s an­swers more than my ar­ti­cle 😂

I’d love to hear more of your thoughts!

Be­cause I’m pas­sion­ate about cloth­ing and fabrics, I only refer­enced brands that are both high qual­ity, and good value for money. If you went from eat­ing at McDon­ald’s to nicer restau­rants, it would be a mis­take to fo­cus solely on the price in­crease; you would get bet­ter qual­ity in­gre­di­ents, ser­vice… less can­cer...

Uniqlo is the best you can find at a low price range.The down­sides are that fabrics are lower qual­ity, will die ear­lier, won’t feel as good; and that the em­ploy­ees’ work con­di­tions are Bangladeshi. Qual­ity comes at a price.

That said, if you wanted to shop at the cheap­est place you could find (or more re­al­is­ti­cally, some­thing like H&M), the rules would still ap­ply and en­hance your style. Thrift shops are also great to find good things on a low bud­get.

Even if you buy from the brands I recom­mend, it should be a grad­ual pro­cess; don’t ruin your­self. My per­sonal rule is one item a month, and I wait for Christ­mas or my birth­day if I want some­thing ex­pen­sive. But granted, you’d still prob­a­bly spend more than you spend now. Although it’d re­ally, re­ally sur­prise me if it was more than food. (Like, re­ally).

A more meta note: it seems rea­son­able to me to ex­pect, if you’re dis­cov­er­ing a new field, that achiev­ing profi­ciency in that area would re­quire in­vest­ing a sig­nifi­cant amount of re­sources. So per­haps you shouldn’t be as sur­prised 🤗

I guess, philo­soph­i­cally, I worry that giv­ing the nodes spe­cial types like that pushes peo­ple to­ward think­ing about agents as not-em­bed­ded-in-the-world, think­ing things like “we need to ex­tend Bayes nets to rep­re­sent ac­tions and util­ities, be­cause those are not nor­mal vari­able nodes”. Not that mem­o­ryless carte­sian en­vi­ron­ments are any bet­ter in that re­spect.

I see where this is com­ing from, but I think it might also go the op­po­site di­rec­tion. For ex­am­ple, my cur­rent guess of how coun­ter­fac­tu­als/​coun­ter­log­i­cals ground out is the imag­i­na­tion pro­cess; I im­plic­itly or ex­plic­itly think of differ­ent ac­tions I could take (or differ­ent ways math could be), and some­how se­lect from those ac­tions (hy­pothe­ses /​ the­o­ries); the ‘magic’ is all hap­pen­ing in my imag­i­na­tion in­stead of ‘in the world’ (not­ing that, of course, my imag­i­na­tion is be­ing phys­i­cally in­stan­ti­ated). Less imag­i­na­tive re­ac­tive pro­cesses (like ther­mostats ‘de­cid­ing’ whether to turn on the heater or not) don’t get this treat­ment, and are bet­ter con­sid­ered as ‘just part of the en­vi­ron­ment’, and if we build an imag­i­na­tive pro­cess out of uni­mag­i­na­tive pro­cesses (cer­tainly neu­rons are more like ther­mostats than they are like minds) then it’s clear the ‘magic’ comes from the ar­range­ment of them rather than the in­di­vi­d­ual units.

Which sug­gests how the type dis­tinc­tion might be nat­u­ral; places where I see de­ci­sion nodes are ones where I ex­pect to think about what ac­tion to take next (or ex­pect some other pro­cess to think about what ac­tion to take next), or think that it’s nec­es­sary to think about how that think­ing will go.

Thank you, I un­der­stand this now (I found it use­ful to imag­ine code that is be­ing in­voked many times and is ter­mi­nated af­ter a ran­dom du­ra­tion; and re­flect on how the agent im­ple­mented by the code should up­date as time goes by).

I guess I should be over­all more op­ti­mistic now :)

What about some form of in­di­rect su­per­vi­sion, where we aim to find tran­scripts in which H has a de­ci­sion of a par­tic­u­lar hard­ness? A would ideally be trained start­ing with things that are very very easy for H, with the hard­ness ramped up un­til A maxes out it’s abil­ities. Rather than imi­tat­ing H, we use a gen­er­a­tive tech­nique to cre­ate fake tran­scripts, imi­tat­ing both H and it’s en­vi­ron­ment. We can in­cor­po­rate into our loss func­tion the amount of time H spends on a par­tic­u­lar de­ci­sion, the re­li­a­bil­ity of that de­ci­sion, and maybe some kind of com­plex­ity mea­sure on the tran­script to find eas­ier/​harder situ­a­tions which are of gen­uine im­por­tance to H.