romeostevensit

• romeostevensit 15 Mar 2019 0:30 UTC

  9 points

  on: Ac­tive Cu­ri­os­ity vs Open Curiosity

  Con­sider the men­tal move you’re perform­ing if you’re do­ing a word search and you switch which word you’re cur­rently ac­tively hunt­ing for. Or search­ing for a word on the tip of your tongue. Gendlin called this a sharp blank or a blank that knows what it is look­ing for. This seems re­lated to ac­tive cu­ri­os­ity vs the more ex­plo­ra­tion based open cu­ri­os­ity.

  Also re­lated: for­ward vs backchain­ing, open vs closed mode, ex­plo­ra­tion vs exploitation

• romeostevensit 14 Mar 2019 2:03 UTC

  8 points

  in reply to: ryan_b's comment on: You Have About Five Words

  Agree, and I’d roll in the in­cen­tives more closely. It feels more like:

  you have at most space for a few feed­back loops

  you can im­prove this by mak­ing one of the feed­back loops a check­list that makes calls out to other feed­back loops

  the tighter and more di­rectly in­cen­tivized the feed­back loop, the more you can pack in

  ev­ery em­ployer/​or­ga­ni­za­tion is try­ing to hire/​re­cruit peo­ple who can hold more feed­back loops at once and do some un­su­per­vised load bal­anc­ing be­tween them

  you can make some of peo­ple’s feed­back loops man­ag­ing an­other per­son’s feed­back loops

  Now jump to this post https://​slat­estar­codex.com/​2017/​11/​09/​ars-longa-vita-bre­vis/​

  an­other frame is that in­stead of think­ing about how many bits you can suc­cess­fully trans­mit, think about whether the be­hav­iors im­plied by the bits you trans­mit can run in loops, whether the loops are su­per­vised or un­su­per­vised and what range of noise they re­main sta­ble un­der.

• romeostevensit 9 Mar 2019 20:53 UTC

  15 points

  AF

  on: Align­ment Re­search Field Guide

  This is fan­tas­tic stuff. Nice to see oth­ers in­de­pen­dently com­ing up with the trans­mit­ters and re­ceivers model. Also, the struc­ture men­tioned in 3a res­onates strongly for me with the peo­ple grop­ing to­wards some sense that Cir­cling type skills seem to be use­ful for ra­tio­nal­ity but couldn’t quite put their finger on why. My ex­pe­rience is that Cir­cling with good fa­cil­i­ta­tors en­ables ex­actly the kinds of things seen in 3a.

  Two things that we’ve found use­ful at QRI that may ap­ply:

  1. A slack or slack like thing (key­base is nice for the ad­di­tional se­cu­rity) for track­ing the ex­plo­sion of refer­ences and con­ver­sa­tional threads that oc­cur when you find a gen­er­a­tive frame/​ques­tion/​method set is way way more use­ful than things like shared gdocs. It al­lows more of ‘get­ting the lay of the land’ to re­ori­ent your­self when you’ve been away and de­vel­op­ments have hap­pened in the mean­time. Stor­ing links in this for­mat also gives them a juicy sense of dis­cov­ery where other for­mats can make them feel more like home­work needed to par­ti­ci­pate in the convo.

  2. Main­tain­ing mo­men­tum in the load bal­anc­ing of con­nec­tions in the graph of one-on-one meet­ings. That is to say, groups seem to func­tion bet­ter when there is roughly equal com­mu­ni­ca­tion be­tween all the par­ti­ci­pants. Prob­a­bly for a va­ri­ety of rea­sons but one ma­jor one is that it seems to al­low bet­ter boot­strap­ping of blindspots. Cross­ing all the pos­si­ble one-on-ones gives the chance for mi­s­un­der­stand­ings to get worked out so that peo­ple can re­turn to much more flow-like com­mu­ni­ca­tion pat­terns. This is ac­com­plished by reg­u­larly schedul­ing the var­i­ous one-on-ones, pri­ori­tiz­ing them, and mak­ing it so that peo­ple can re­quest them with­out feel­ing like it is a big ask. Gen­er­ally ac­com­plished via video chat when in per­son would be la­bo­ri­ous.

• romeostevensit 2 Mar 2019 2:33 UTC

  2 points

  on: Policy-Based vs Willpower-Based Intentions

  This feels like how I think about TDT as a ca­sual heuris­tic.

• romeostevensit 1 Mar 2019 2:23 UTC

  6 points

  on: Mo­ti­va­tion: You Have to Win in the Mo­ment

  It isn’t all the mo­ments you’re do­ing the ac­tion but the mo­ments of an­other can­di­date win­ner be­ing sur­faced in con­scious­ness giv­ing you the af­for­dance to do some­thing else in­stead. In that mo­ment you can reaf­firm your com­mit­ment to the thing or choose the dis­trac­tion. Poor mod­el­ing of the self who has to make that microde­ci­sion is be­hind much con­fu­sion.

• romeostevensit 26 Feb 2019 4:47 UTC

  3 points

  on: Na­tive men­tal rep­re­sen­ta­tions that give huge speedups on prob­lems?

  fac­tor anal­y­sis, lev­els of anal­y­sis.

• romeostevensit 18 Feb 2019 22:59 UTC

  13 points

  on: Epistemic Tenure

  an­other frame might be Kuh­nian mo­men­tum.

• romeostevensit 16 Feb 2019 4:05 UTC

  5 points

  on: Ped­a­gogy as Struggle

  I’ve seen this referred to as dis­cur­sive learn­ing. The idea that things stick much bet­ter when you’re forced to ex­pend some cog­ni­tion in or­der to figure out how it fits with what you already know.

• romeostevensit 13 Feb 2019 23:52 UTC

  4 points

  on: Three Kinds of Re­search Doc­u­ments: Clar­ifi­ca­tion, Ex­plana­tory, Academic

  I’ve been work­ing on map­ping ex­pla­na­tion-space and have iden­ti­fied 3 ar­eas of de­com­po­si­tion, the how-why-what axis cor­re­spond­ing to lev­els of anal­y­sis (Marr, et al), the tem­po­ral axis delineat­ing past, fu­ture, and now as a spe­cial mo­ment of causal in­ter­ven­tion, and the var­i­ance-in­var­i­ance axis.

• romeostevensit 11 Feb 2019 18:41 UTC

  2 points

  in reply to: cousin_it's comment on: Some Thoughts on Metaphilosophy

  I think it is driven by a gen­eral heuris­tic of find­ing com­press­ibil­ity. If a dis­tri­bu­tion seems com­plex we as­sume we’re ac­ci­den­tally con­flat­ing two vari­ables and seek the de­com­po­si­tion that makes the two re­sul­tant dis­tri­bu­tions ap­prox­i­mate-able by sim­pler func­tions.

• romeostevensit 9 Feb 2019 10:40 UTC

  6 points

  on: Why we need a *the­ory* of hu­man values

  Imag­ine you have 10 sen­sors mea­sur­ing differ­ent things. The en­vi­ron­ment (in­clud­ing things you can af­fect) per­mutes their val­ues over time. Let’s say they op­er­ate like ther­mome­ters and you’re try­ing to keep them within speci­fied ranges. If the sen­sors are tri­nary (too low, too high, just right) you already have 59,049 states to nav­i­gate in your trade­off space. The higher the re­s­olu­tion the sen­sors the faster the com­bi­na­toric ex­plo­sion. So a small num­ber of pa­ram­e­ters leads to a very com­plex seem­ing situ­a­tion.

• romeostevensit 7 Feb 2019 4:38 UTC

  2 points

  in reply to: Raemon's comment on: Rae­mon’s Short­form Feed

  would likely be solved if slack had a ro­bust 3 level on­tol­ogy rather than two level. Threaded con­ver­sa­tions don’t work very well.

• romeostevensit 6 Feb 2019 0:38 UTC

  7 points

  in reply to: Raemon's comment on: Rae­mon’s Short­form Feed

  I agree that slack is a bet­ter in­ter­ac­tion modal­ity for mul­ti­ple peo­ple try­ing to make progress on prob­lems. The main draw­back is chaotic chan­nel on­tolo­gies lead­ing to too many buck­ets to check for users (though many obv. find this as­pect ad­dic­tive as well).

• romeostevensit 5 Feb 2019 18:14 UTC

  4 points

  on: Sam Har­ris and the Is–Ought Gap

  In epistemic struc­tural re­al­ism the bridge is all we have. One end of the bridge feels more ‘is’ like and one end feels more ‘ought’ like. Both are sub­ject to ex­ten­sion­al­ism in try­ing to figure out what’s ‘re­ally there’. I find this stance much much less con­fus­ing than the more stan­dard in­di­rect re­al­ism that typ­i­cally un­der­lies the is-ought dis­tinc­tion.

  There’s also the gen­eral pat­tern, see if by in­vert­ing the na­ture of the rep­re­sen­ta­tion (turn the edges into ver­tices and vice versa) a false di­chotomy dis­ap­pears.

• romeostevensit 5 Feb 2019 18:08 UTC

  4 points

  on: What makes a good cul­ture?

  Jung seemed to be the first to iden­tify more ex­plic­itly the punc­tu­ated equil­ibrium model of adult psy­cholog­i­cal de­vel­op­ment. I find it use­ful to com­bine with sub agents to ex­plain a lot of the psy­chotehra­puetic mod­els I see. Differ­ent parts can be at differ­ent stages of de­vel­op­ment and thus be open to differ­ent kinds of ev­i­dence and ex­pe­riences in or­der to ma­ture and gain ac­cess to bet­ter strate­gies for get­ting needs met.

• romeostevensit 23 Jan 2019 6:02 UTC

  2 points

  on: Book Sum­mary: Con­scious­ness and the Brain

  I re­ally want to read a book fo­cus­ing on vigilance. Con­scious ac­cess seemed like the least in­ter­est­ing of the three by far.

• romeostevensit 23 Jan 2019 5:54 UTC

  2 points

  on: Ca­pa­bil­ity amplification

  Men­tioned be­fore, but I won­der if pro­gres­sive/​iter­a­tive sum­ma­riza­tion would in­crease peer re­view be­tween AI re­searchers a bunch. I.e. marginal efforts to­wards sum­ma­riz­ing the ei­ther ad­ver­sar­ial or gen­er­a­tive ideas a post is a re­sponse to, which gives the origi­nal au­thors and out­siders more sur­faces to offer ideas along. I think this would also lead to more set­tling on shared ter­minol­ogy, which is a sig­nifi­cant frac­tion of progress in a field as far as I can tell from the his­tory of sci­ence. If it in­creased en­gage­ment then it would be di­rectly in­cen­tivized as soon as peo­ple knew/​ex­pe­rienced that.

• romeostevensit 23 Jan 2019 5:48 UTC

  7 points

  in reply to: Kaj_Sotala's comment on: Life can be bet­ter than you think

  +1 elimi­nat­ing cer­tain nega­tive emo­tions seems to tem­porar­ily get rid of any mo­ti­va­tion struc­tures that were us­ing those nega­tive emo­tions as a build­ing block. Things I re­ally value seem to re­build them­selves on bet­ter foun­da­tions.

• romeostevensit 23 Jan 2019 5:37 UTC

  12 points

  on: Life can be bet­ter than you think

  I have a hy­poth­e­sis for in­creas­ing trans­fer­abil­ity of in­sights. Their trans­fer­abil­ity be­ing by de­fault quite low. Lower than they feel like they should be from the in­side. I think what gen­er­ally hap­pens is some­thing like this: a per­son has an in­sight, this gen­er­ates a bunch of emo­tional en­ergy, some­times this gets chan­neled into the urge to share/​write about the in­sight, this is most of what we hear about. But the writ­ing is from the per­spec­tive of the in­sight, which tends to be dis­similar from the ma­te­rial that *trig­gered* the in­sight. I no­ticed this in my­self af­ter de­vel­op­ing a very de­tailed note tak­ing sys­tem. This al­lowed me to go back and trace the tra­jec­tory of past in­sights. It is much harder and less mo­ti­vat­ing (cur­rently) to share stuff from the pre-in­sight per­spec­tive. Harder be­cause of in­sight am­ne­sia, the ten­dency to for­get what your past think­ing pat­terns were like, and also be­cause most don’t have de­tailed enough notes. Less mo­ti­vat­ing be­cause pre-in­sight ma­te­rial just seems, well, wrong now. Why write about wrong things when you could write about *glo­ri­ous new cor­rect thing*?

• romeostevensit 23 Jan 2019 5:21 UTC

  3 points

  on: Disen­tan­gling ar­gu­ments for the im­por­tance of AI safety

  Pretty ran­dom, but the thought this post lead to was ‘let’s up­load Cu­ladasa and then have the up­load di­rect us in how to iter­a­tively up­grade him.’ Sort of like gun-to-the-head you have to pick the safest op­tion us­ing only the con­cepts we already have right now.