An1lam

• An1lam 5 Dec 2019 2:29 UTC

  2 points

  on: Con­nec­tome-Spe­cific Har­monic Waves

  FFNNs’ in­abil­ity to pro­cess time se­ries data was a con­tribut­ing fac­tor to the Uber self-crash­ing car.

  I don’t see the ev­i­dence for this in the linked post and don’t re­call see­ing this in any of the other few ar­ti­cles/​pieces I’ve read on the crash. Can you point me to the ev­i­dence for this?

• An1lam 5 Dec 2019 0:10 UTC

  9 points

  on: Paper-Read­ing for Gears

  I won­der how hard it would be to for­mal­ize this claim about me­di­a­tion in terms of the space of causal DAGs. I haven’t done the work to try it so I’m mostly spit­bal­ling.

  In­for­mally, I as­so­ci­ate me­di­a­tion with the front-door crite­ria in causal­ity. So, the use­ful­ness of me­di­a­tion should re­flect that the front-door crite­rion is em­piri­cally eas­ier to satisfy than the back-door (and other) crite­ria maybe be­cause real causal chains tend to be nar­row but long? Think­ing about it a bit more, it’s prob­a­bly more like the min cut of real world causal graphs tend to be rel­a­tively small?

• An1lam 18 Nov 2019 15:10 UTC

  4 points

  in reply to: cousin_it's comment on: The Power to Draw Better

  Any re­sources you’d recom­mend that de­scribe the con­struc­tive style fur­ther? I’ve read Draw­ing on the Right Side of the Brain and so would be cu­ri­ous to read about this other ap­proach more.

• An1lam 16 Nov 2019 19:20 UTC

  3 points

  on: Chris Olah’s views on AGI safety

  I’d be cu­ri­ous whether Chris and the other mem­bers of the Clar­ity team have thought about (in their most op­ti­mistic vi­sion) how “micro­scope AI” could en­able other sci­ences—biol­ogy, neu­ro­science, physics, etc. -- and if so whether they’d be will­ing to share. AIUI, the lack of in­ter­pretabil­ity of neu­ral net­works is much more of an im­me­di­ate is­sue for try­ing to use deep learn­ing in these ar­eas be­cause just be­ing able to pre­dict tends to be less use­ful.

• An1lam 14 Nov 2019 16:13 UTC

  1 point

  in reply to: RobinHanson's comment on: Robin Han­son on the fu­tur­ist fo­cus on AI

  Mostly un­re­lated to your point about AI, your com­ments about the 100,000 fans hav­ing the po­ten­tial to cause harm rang true to me.

  Are there other ar­eas in which you think the many non-ex­pert fans prob­lem is es­pe­cially bad (as op­posed to com­puter se­cu­rity, which you view as healthy in this re­spect)?

  Then the ex­perts can be rea­son­able and peo­ple can say, “Okay,” and take their word se­ri­ously, al­though they might not feel too much pres­sure to listen and do any­thing. If you can say that about com­puter se­cu­rity to­day, for ex­am­ple, the pub­lic doesn’t scream a bunch about com­puter se­cu­rity.

  ---

  Would you con­sider progress on image recog­ni­tion and ma­chine trans­la­tion as out­side view ev­i­dence for lump­iness? Ac­cu­ra­cies on ImageNet, an image clas­sifi­ca­tion bench­mark, dropped by >10% over a 4-year pe­riod (graph be­low) mostly due to the suc­cess­ful scal­ing up of a type of neu­ral net­work.

  

  This also seems rele­vant to your point about AI re­searchers who have been in the field for a long time be­ing more skep­ti­cal. My un­der­stand­ing is that most AI re­searchers would not have pre­dicted such rapid progress on this bench­mark be­fore it hap­pened.

  That said, I can see how you still might ar­gue this is an ex­am­ple of over-em­pha­siz­ing a sim­ple form of per­cep­tion, which in re­al­ity is much more com­pli­cated and in­volves a bunch of differ­ent in­ter­lock­ing pieces.

• An1lam 13 Nov 2019 23:52 UTC

  1 point

  in reply to: jimrandomh's comment on: An1lam’s Short Form Feed

  In an in­ter­est­ing turn of events, John Car­mack an­nounced to­day that he’ll be pivot­ing to work on AGI.

• An1lam 8 Nov 2019 1:05 UTC

  8 points

  in reply to: John_Maxwell's comment on: Open & Wel­come Thread—Novem­ber 2019

  First of all, I apol­o­gize, I think my com­ment was too snarky and took a tone of “this is so sur­pris­ing” that I re­gret on re­flec­tion.

  Un­der what cir­cum­stances do you feel in­tro­duc­ing new policy ideas with the word “maybe this could be a good idea” is ac­cept­able? To be clear, I think in­tro­duc­ing the idea is to­tally fine, I just have a de­cently strong prior against wide­spread bans on “busi­nesses that do this”.

  I don’t ex­pect any­one im­por­tant to be read­ing this thread, cer­tainly not im­por­tant poli­cy­mak­ers. Even if they were, I think it was pretty clear I was spit­bal­ling.

  Agreed, I was not wor­ried about this.

  Let’s not fall prey to the halo effect. Eliezer also wrote a long post about the ne­ces­sity of back-and-forth de­bate, and he’s us­ing a plat­form which is uniquely bad at this. At some point, one starts to won­der whether Eliezer is a mor­tal hu­man be­ing who suffers from akra­sia and bi­ases just like the rest of us.

  Fair enough. I agree the Eliezer point isn’t strong ev­i­dence.

  I didn’t make much of an effort to as­sem­ble ar­gu­ments that Twit­ter is bad. But I think there are good ar­gu­ments out there. How do you feel about the nu­clear diplo­macy that’s hap­pened on Twit­ter?

  I don’t have time to re­spond at length to this part at the mo­ment (I wanted to re­ply quickly to apol­o­gize mostly) but I agree it’s the most use­ful ques­tion to dis­cuss and will try to re­spond more later. To sum­ma­rize, I ac­knowl­edge it’s pos­si­ble that Twit­ter is bad for the col­lec­tive but think peo­ple may over­es­ti­mate the bad parts by fo­cus­ing on how poli­ti­ci­ans /​ peo­ple fight­ing about poli­tics use Twit­ter (which does seem bad) and that even if it is “bad”, it’s not clear that ban­ning short re­sponse web­sites would lead to a bet­ter long-term out­come. For ex­am­ple, maybe peo­ple would just start fight­ing with pic­tures on In­sta­gram. I don’t think this spe­cific out­come is likely but think it’s in a class of out­comes that would re­sult from ban­ning that seems de­cently likely.

• An1lam 7 Nov 2019 18:46 UTC

  8 points

  in reply to: John_Maxwell's comment on: Open & Wel­come Thread—Novem­ber 2019

  Woah, this seems like a big jump to a form of tech­noc­racy /​ pa­ter­nal­ism that I would think would typ­i­cally re­quire more jus­tifi­ca­tion than spend­ing a short amount of time brain­storm­ing in a com­ment thread why the thing mil­lions of peo­ple use daily is ac­tu­ally bad.

  Like, ban­ning sites from offer­ing free ser­vices if a char­ac­ter limit is in­volved be­cause high sta­tus mem­bers of com­mu­ni­ties you like en­joy such sites seems like bad policy and also a weird way to try and con­vince these peo­ple to write on fo­rums you pre­fer. Now, one coun­ter­ar­gu­ment would be “co­or­di­na­tion prob­lems” mean those writ­ers would pre­fer to write some­where else. But pre­sum­ably if any­one’s aware of “in­ad­e­quate equil­ibria” like this and able to avoid it it would be Eliezer.

  Re-read­ing my com­ment, I re­al­ize it may come off as snarky but I’m not re­ally sure how bet­ter to con­vey my sur­prise that this would be the first idea that comes to mind.

  ETA: it’s not clear to me that Twit­ter is so toxic, es­pe­cially for peo­ple in the broad bub­ble that en­com­passes LW /​ EA /​ tech /​ etc. I agree it’s not the best pos­si­ble ver­sion of a plat­form by any means, but to say it’s ob­vi­ously net nega­tive seems like a stretch with­out fur­ther ev­i­dence.

• An1lam 5 Nov 2019 14:00 UTC

  12 points

  on: An1lam’s Short Form Feed

  In­ter­est­ing Bill Thurston quote, sadly from his obit­u­ary:

  I’ve always taken a “lazy” at­ti­tude to­ward calcu­la­tions. I’ve of­ten ended up spend­ing an in­or­di­nate amount of time try­ing to figure out an easy way to see some­thing, prefer­ably to see it in my head with­out hav­ing to write down a long chain of rea­son­ing. I be­came con­vinced early on that it can make a huge differ­ence to find ways to take a step-by-step proof or de­scrip­tion and find a way to par­allelize it, to see it all to­gether all at once—but it of­ten takes a lot of strug­gle to be able to do that. I think it’s much more com­mon for peo­ple to ap­proach long case-by-case and step-by-step proofs and com­pu­ta­tions as te­dious but nec­es­sary work, rather than some­thing to figure out a way to avoid. By now, I’ve found lots of “big pic­ture” ways to look at the things I un­der­stand, so it’s not as hard. To pre­vent mis-in­ter­pre­ta­tion, I think peo­ple of­ten look at quotes like this (I’ve seen similar ones about Feyn­man) and think “ah yes, see any­one can do it”. But IME the thing he’s de­scribing is much harder to achieve than the “case-by-case”/​”step-by-step” stuff.

• An1lam 5 Nov 2019 3:57 UTC

  2 points

  in reply to: Matthew Barnett's comment on: Matthew Bar­nett’s Shortform

  Weirdly enough, I was do­ing some­thing to­day that made me think about this com­ment. The thought I had is that you caught onto some­thing good here which is sep­a­rate from the pres­sure as­pect. There seems to be a benefit to try­ing to sep­a­rate differ­ent as­pects of a task more than may feel nat­u­ral. To use the fi­nal exam ex­am­ple, as some­one men­tioned be­fore, part of the rea­son fi­nal ex­ams feel pro­duc­tive is be­cause you were forced to do so much prep be­fore­hand to en­sure you’d be able to finish the exam in a fixed amount of time.

  Similarly, I’ve seen benefit when I (hap­haz­ardly since I only re­al­ized this re­cently) clearly seg­ment differ­ent as­pects of an ac­tivity and ap­ply ar­tifi­cial con­straints to en­sure that they re­main sep­a­rate. To use your VAE blog post ex­am­ple, this would be like say­ing, “I’m only go­ing to use a sin­gle page of notes to write the blog post” to force your­self to en­sure you un­der­stand ev­ery­thing be­fore try­ing to write.

  YMMV warn­ing: I’m es­pe­cially bad about try­ing to pro­duce out­puts be­fore fully un­der­stand­ing and there­fore may get more band­width out of this than oth­ers.

• An1lam 5 Nov 2019 2:15 UTC

  3 points

  in reply to: johnswentworth's comment on: Book Re­view: De­sign Prin­ci­ples of Biolog­i­cal Circuits

  No prob­lem, also, in case it wasn’t ob­vi­ous based on the fact that I com­mented ba­si­cally right af­ter you posted this, I liked the post a lot and now plan to read this book ASAP.

• An1lam 4 Nov 2019 23:17 UTC

  8 points

  on: Book Re­view: De­sign Prin­ci­ples of Biolog­i­cal Circuits

  Note: the talk you men­tioned was by Drew Endy and the elec­tri­cal en­g­ineer­ing col­league was the one and only Ger­ald Suss­man of SICP fame (source: I don’t re­mem­ber but I’m very con­fi­dent about be­ing right).

• An1lam 3 Nov 2019 13:26 UTC

  1 point

  in reply to: waveman's comment on: Epistemic Spot Check: The Role of De­liber­ate Prac­tice in the Ac­qui­si­tion of Ex­pert Performance

  In this in­ter­view, Don Knuth gives the strong im­pres­sion that he works more than 4 hours a day not nec­es­sar­ily do­ing de­liber­ate prac­tice but definitely hard cog­ni­tive work (writ­ing a book that most peo­ple con­sider quite challeng­ing to read). That said, Knuth is kind of a mon­ster in gen­eral in terms of com­bin­ing re­ally high tech­ni­cal abil­ity and re­ally high con­scien­tious­ness, so it wouldn’t sur­prise me if he’s similar to the other out­liers you men­tioned and is not rep­re­sen­ta­tive.

  A few ex­am­ples to back up my con­scien­tious­ness point:

  • The fol­low­ing story about do­ing all the prob­lems in the text­book (from that in­ter­view):

  But Thomas’s Calcu­lus would have the text, then would have prob­lems, and our teacher would as­sign, say, the even num­bered prob­lems, or some­thing like that. I would also do the odd num­bered prob­lems. In the back of Thomas’s book he had sup­ple­men­tary prob­lems, the teacher didn’t as­sign the sup­ple­men­tary prob­lems; I worked the sup­ple­men­tary prob­lems. I was, you know, I was scared I wouldn’t learn calcu­lus, so I worked hard on it, and it turned out that of course it took me longer to solve all these prob­lems than the kids who were only work­ing on what was as­signed, at first.

  • Writ­ing an en­tire com­piler on his own over a sum­mer.

  • Finish­ing his PhD in three years while also con­sult­ing with Bur­roughs.

• An1lam 30 Oct 2019 22:18 UTC

  8 points

  in reply to: maximkazhenkov's comment on: The Tech­nique Taboo

  I agree with this crit­i­cism and have made it my­self be­fore. How­ever, I do think there’s some­thing to it when it comes to es­pe­cially slow typ­ing (ar­bi­trar­ily, ⇐ 35 WPM). At least for me, some­times I need to “clear out my men­tal buffer” by ac­tu­ally writ­ing out the code I’ve thought about be­fore I can think about the next thing. When I’m in this po­si­tion, be­ing able to type rel­a­tively quickly seems to helps me stay in the flow and get to the next thought faster.

  Also, frankly, some­times you just need to churn out 10s of lines of code that are some­what mind­less (yes, DRY, but I still think some situ­a­tions re­quire this) and while do­ing them fast doesn’t save you that much time over­all, it cer­tainly can help with main­tain­ing san­ity.

  Re­lated to that, I think the big prob­lem with hunt and peck typ­ing is that it isn’t just slower, it also takes your at­ten­tion off the flow of the code by forc­ing you to fo­cus on the char­ac­ters you’re typ­ing.

  ETA: all that said, I definitely agree with Said that it’s not nec­es­sary to learn typ­ing in a for­mal set­ting and definitely would not en­courage col­leges to teach it. I ac­tu­ally did have a com­puter class in el­e­men­tary school that taught touch typ­ing but still got much faster mostly by us­ing AIM in mid­dle school.

• An1lam 30 Oct 2019 14:32 UTC

  1 point

  in reply to: MakoYass's comment on: Turn­ing air into bread

  Thanks for shar­ing your per­spec­tive!

• An1lam 29 Oct 2019 22:31 UTC

  11 points

  in reply to: MakoYass's comment on: Turn­ing air into bread

  As crazy as the prophet Malthus sounded, some peo­ple con­tin­u­ously try to heed his warn­ing, thank­fully we’re learn­ing how to co­or­di­nate our pop­u­la­tion growth to sup­port a good life within the limited car­ry­ing ca­pac­ity of our nat­u­ral re­sources bet­ter and bet­ter over time. Time and time again, a wiz­ard makes a new gizmo and we all get away with it.

  The mod­ified ver­sion of your first para­graph from above feels as or more ac­cu­rate to me. I’d be cu­ri­ous to hear why you think it won’t always be like this (be­sides X-risk, which I to­tally un­der­stand would lead to a “not always like this” situ­a­tion).

• An1lam 29 Oct 2019 13:58 UTC

  5 points

  on: An1lam’s Short Form Feed

  Blockchain idea in­spired by 80,000 Hours’s in­ter­view of Vi­talik Bu­terin: a lot of pod­casts ei­ther have ter­rible tran­scrip­tions or pre­sum­ably pay a ser­vice to tran­scribe their ses­sions. How­ever, even these ser­vices make minor ty­pos such as “ASX” in­stead of “ASICs” in the linked in­ter­view.

  Now, most peo­ple who read these tran­scripts pre­sum­ably no­tice at least a sub­set of these ty­pos but don’t want to go through the effort of emailing pod­cast­ers to tell them about it. On the flip side, there’s no good way for hosts to scal­ably au­dit flagged ty­pos to see if they’re ac­tu­ally ty­pos. What we re­ally want is a mostly au­to­mated mechanism to ag­gre­gate flagged ty­pos and ac­cept fixes which mul­ti­ple peo­ple agree upon that only pays peo­ple (micro amounts) for cor­rectly iden­ti­fy­ing ty­pos.

  This mechanism seems like some­thing that could live on a blockchain in some sort of smart con­tract. Ob­vi­ously, like al­most ev­ery blockchain ap­pli­ca­tion, you *could* do it with­out a blockchain, but us­ing blockchain makes it easy to au­dit and dis­tribute the micro-pay­ments rather than hav­ing to pro­gram the vot­ing scheme from scratch on top of a cen­tral­ized database.

• An1lam 24 Oct 2019 19:57 UTC

  3 points

  on: An1lam’s Short Form Feed

  If alge­bra’s a deal with the devil where you get the right an­swer but don’t know why, then ge­o­met­ric in­tu­ition’s a deal with the devil where you always get an an­swer but don’t know whether it’s right.

• An1lam 24 Oct 2019 19:56 UTC

  1 point

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

  Uh, what if you for­get to do your habit trou­bleshoot­ing habit and then you have to trou­bleshoot why you for­got it? And then you for­get it twice and you have to trou­bleshoot why you for­got to trou­bleshoot for­get­ting to trou­bleshoot!

  (I’m jok­ing about all this in case it’s not ob­vi­ous.)

• An1lam 24 Oct 2019 19:35 UTC

  3 points

  on: An1lam’s Short Form Feed

  Some­one should write the equiv­a­lent of TAOCP for ma­chine learn­ing.

  (Ok, maybe not liter­ally the equiv­a­lent. I mean Knuth is… Knuth. So it doesn’t seem re­al­is­tic to ex­pect some­one to do some­thing as im­pres­sive as TAOCP. And yes, this is au­thor­ity wor­ship and I don’t care. He’s Knuth god­damn it.)

  Speci­fi­cally, a book where the the­ory/​math’s rigor­ous but the al­gorithms are de­scribed in their effi­cient forms. I haven’t found this in the few ML books I’ve read parts of (Bishop’s Pat­tern Recog­ni­tion and Ma­chine Learn­ing, MacKay’s In­for­ma­tion The­ory, and Tibr­ishani et Al’s Ele­ments of Statis­ti­cal Learn­ing), so if it’s already out there, let me know.

  Note that I don’t mean that who­ever does this should do the whole MMIX thing and write their own lan­guage and VM.