# The most important meta-skill

Note: This ar­ti­cle un­der­went a sig­nifi­cant re­vi­sion on 5/​28/​2015. Thank you to es­ti­ma­tor for all your feed­back.

The most im­por­tant meta-skill that any­one can learn is how to learn skills. With prac­tice, you learn how to pick up new skills as they are needed, which is in­finitely (quite liter­ally) more effi­cient than try­ing to learn each skill in­di­vi­d­u­ally in ad­vance.

There are two ba­sic premises that this method re­lies on:

1. A skill can be even­tu­ally be bro­ken down into a se­ries of triv­ial sub-skills.

2. The skill and its sub-skills fol­lows a Pareto dis­tri­bu­tion.

The Pareto prin­ci­ple states that typ­i­cally, 80% of a sys­tem’s effects can be linked to 20% of their causes. Or in this case, learn­ing 20% of the triv­ial sub-skills will make you 80% profi­cient at the over­all skill. Em­piri­cally, many sys­tems, both ar­tifi­cial and nat­u­ral have been proven to fol­low this dis­tri­bu­tion, and skills are no ex­cep­tion. This guide is in­tended to teach you how to iden­tify that 20%.

What lies be­low this is al­most 1,000 words to de­scribe some­thing that’s ul­ti­mately about con­dens­ing things and tak­ing short­cuts. So, to be true to this at­ti­tude, I’ll start with the “20% ver­sion”, and those so in­clined can con­tinue to read the other 80%.

1. Break the skill you want to learn into sev­eral sub-skills.

2. For each sub-skill, ask “Is this triv­ial?” If so, add that to your “triv­ial list”. If not, re­peat steps 1-2 for each sub-skill. Con­tinue to iter­ate un­til all you have left is a list of triv­ial sub-skills.

3. For each triv­ial sub-skill, ask, “How can this go wrong, and what can I do if it does?” Add this to your list of back-up plans, un­less it is re­dun­dant.

4. Sort your list of sub-skills by how easy they will be to learn, then start learn­ing and prac­tic­ing them. Any time some­thing goes wrong or you en­counter a situ­a­tion you did not ac­count for, use one of your back-up plans.

5. Re­peat steps 1-4 for any sub-skills you en­counter that you did not ac­count for.

So, that was the short ver­sion. If you find you need more con­text, here goes. Note that the first premise uses the word “triv­ial”, which then begs the ques­tion: “What makes a sub-skill triv­ial?” A con­ve­nient an­swer to that is: “If you per­son­ally feel suffi­ciently con­fi­dent that you can do it.” Or, in other words, “Can you look up how to do it on the in­ter­net?” Which means, if the prob­lem it­self is triv­ial, you don’t need to bother with this. Just look up a guide on­line.

Most skills are too com­pli­cated for some­one to sit down and an­a­lyze ev­ery pos­si­ble sub-skill needed to ac­com­plish it. For­tu­nately, you don’t have to. Your goal isn’t to learn all the sub-skills, its to learn the im­por­tant 20%. The over­all effi­ciency of a sub-skill is a func­tion of two things: how how in­te­gral it is to the over­all skill, and how easy it is to learn. You’re go­ing to let Sys­tem 1 do most of the heavy lift­ing here.

For­tu­nately, our brains are pretty good at pat­tern-match­ing. Goals are high-level con­cepts whose mean­ings are de­rived from the com­bi­na­tion of sev­eral pat­terns and archetypes that you’ve got stored away some­where. When you say, “I want to learn a for­eign lan­guage”, your brain im­me­di­ately starts filling in the pat­terns of what ex­actly that means. It starts iden­ti­fy­ing the things that are in­te­gral to your idea of the con­cept. Then it com­bines them into one co­her­ent con­cept, and that’s what you’re left with. The trou­ble is, most peo­ple don’t pre­serve these in­di­vi­d­ual pat­terns be­fore com­bin­ing them, and thus they’re left with some­thing that’s purely con­cep­tual, rather than ac­tion­able. “I want to learn a for­eign lan­guage” or “I want to learn to code” or “I want to learn so­cial skills”.

So just let your brain go to work do­ing what it already does, but pay at­ten­tion dur­ing the pro­cess and iden­tify the key com­po­nents be­fore they get mushed into a con­cept. Make Sys­tem 1 tell you “You want to be able to con­verse, in­ter­act, and func­tion in a so­ciety that speaks a differ­ent lan­guage,” in­stead of just, “You want to learn a for­eign lan­guage.” Re­mem­ber that you don’t need to iden­tify all the com­po­nents. Just the ones that are im­por­tant enough for Sys­tem 1 to dredge up on a mo­ment’s no­tice. Most likely, these will be the 20% that you’re look­ing for. Of course, chances are the ini­tial out­put is go­ing to be a high level con­cept unto it­self. There’s no “to-do list” for “be­ing able to con­verse in a so­ciety that speaks a differ­ent lan­guage”. So you put Sys­tem 1 to work again. What ex­actly do I mean by that? “Oh, what you mean is: you want to be able to ask and un­der­stand both ques­tions and an­swers, and be able to ex­press your thoughts.

Even­tu­ally you’ll reach the point of triv­ial­ity. You’ll have a siz­able list of triv­ial tasks such as “You want to be able to say the fol­low­ing twenty ba­sic sen­tences: XYZ”, and You want to know the fol­low­ing 100 ba­sic vo­cab­u­lary words: ABC.” and “You want to be able to iden­tify the most com­mon ar­ti­cles, prepo­si­tions and con­junc­tions.” Here’s where Sys­tem 2 goes to work: you look at this big list and ask your­self, which of these would be eas­iest for me to ac­com­plish? And then you sort the list ac­cord­ingly.

Now, all of this is fine and good, but at some point you will en­counter a situ­a­tion that doesn’t fall un­der this con­ve­nient lit­tle roadmap you’ve fol­lowed. So you want to make a backup plan. Sys­tem 2 needs to look over your roadmap and ask: “How can this go wrong, and what can I do if it does?” If you do this for each item on your list, chances are there will be a lot of du­pli­cates and re­dun­dan­cies, which you can pare down. When all is said and done, you’ll have a few plans of ac­tion in case things go wrong.

So, you have a roadmap to guide you through the 20%, and a gen­er­al­ized plan for the other 80%. What now?

Well, there’s always room for im­prove­ment. If you do things right, you’ll be pretty well im­mersed in the nitty-gritty of what­ever skill you are try­ing to learn, which means you will be get­ting loads of first-hand ex­pe­rience as to all the differ­ent ways things can go wrong which you prob­a­bly never could have an­ti­ci­pated. And you’ll run in to sce­nar­ios that make you say, “I can’t be­lieve I didn’t think about that.”

For­tu­nately you don’t need to get things perfect on the first try. If you en­counter a situ­a­tion you didn’t ac­count for it, then ac­count for it. Ask your­self what hap­pened, and let Sys­tem 1 go to work on break­ing it down. If some­thing goes wrong in a way you hadn’t thought about, come up with a sep­a­rate plan for that. Even­tu­ally your model will be­come more and more ro­bust as you start to learn many of the fun­da­men­tals that you prob­a­bly skipped over when you made your roadmap.

There seem to be two differ­ent types of learn­ing styles, the “aca­demic” way of start­ing with the fun­da­men­tals and build­ing from the ground up, and the “im­mer­sion” method of just throw­ing some­one into the deep end of the pool and work­ing from the top-down. This method com­bines both: you learn the fun­da­men­tals of the things that are nec­es­sary to im­merse your­self. In­stead of be­ing “top-down” or “bot­tom-up”, this is more like, “start at the bot­tom, skip to the top, then work your way back down through the mid­dle.”

• I think that all self-help /​ “learn­ing to learn” /​ etc. ar­ti­cles should con­tain a short sum­mary tel­ling us some rea­sons to ac­tu­ally be­lieve any­thing writ­ten be­low. Like refer­ences to rele­vant re­search, or au­thor’s real life achieve­ments, or some­thing. Gen­er­ally, one shouldn’t rely on per­sonal anec­dotes; but for self-help, even hav­ing a sin­gle data point is of­ten too high a stan­dard.

In your ar­ti­cle, I couldn’t find a sin­gle bit of ev­i­dence in sup­port of your claims.

• Sure, I could, but would that make you any more likely to ac­cept it? Gen­er­ally I’ve found that the more some­one ex­pounds on their own cre­den­tials, the less cred­ible (and lik­able) they sound.

If my own per­sonal achieve­ments would gen­uinely make a differ­ence to you per­son­ally, then I’d be glad to tell you them. If not, then I don’t quite see the point.

• Ar­ti­cles on such top­ics are no­to­ri­ous for their av­er­age bad qual­ity. Re­for­mu­lat­ing in Bayesian terms, the prior prob­a­bil­ity of your state­ments be­ing true is low, so you should provide some proofs or ev­i­dence—or why I (or any­one) should be­lieve you? Have you ac­tu­ally checked if it works? Have you ac­tu­ally checked if it works for some­body else?

I don’t think that per­sonal achieve­ments are a bul­let-proof ar­gu­men­ta­tion for such an ad­vice. Still, when I read some­thing like this, I’m pretty sure that it con­tains valuable in­for­ma­tion, al­though it is prob­a­bly a mis­take to fol­low such ad­vice ver­ba­tim any­way. So, if you have Ham­ming-level cre­den­tials, it will help.

As for your ar­ti­cle, prob­a­bly the only way to fix it is to add proofs to your state­ments. What ev­i­dence sup­ports them? Is there any psy­cholog­i­cal re­search to back up your claims? Why do you think it is op­ti­mal (or near-op­ti­mal) way to learn skills?

This is a good self-help ar­ti­cle. Can you see the refer­ence list? :)

• Ar­ti­cles on such top­ics are no­to­ri­ous for their av­er­age bad qual­ity.

That’s in­ter­est­ing, I wasn’t aware of that rep­u­ta­tion. That’s good to know and cer­tainly jus­tifies your skep­ti­cism.

All that said, I think one can still eval­u­ate your point (and in my case, my Less Wrong post) based on its in­ter­nal logic and how con­sis­tent it is with one’s own ob­ser­va­tions, with­out need­ing re­search to back it up. It would be easy enough to dis­miss your own post for the very rea­sons you cited. Con­sider the fol­low­ing:

In gen­eral, peo­ple new to a com­mu­nity are no­to­ri­ously bad at gaug­ing the pulse of said com­mu­nity. To re­for­mu­late in Bayesian terms, based on the length of time you’ve been post­ing here, the prior prob­a­bil­ity of your state­ment be­ing true is low, so shouldn’t you provide some proofs or ev­i­dence—or why should I (or any­one) be­lieve you?

But to me, your logic checks out, and is fairly con­sis­tent with my own ob­ser­va­tions (that most self-help pub­li­ca­tions tend to be garbage), so that shifts the prob­a­bil­ities sig­nifi­cantly in your fa­vor. I’m hop­ing that peo­ple will eval­u­ate my own post by similar crite­ria, rather than im­me­di­ately dis­miss­ing it.

• I’ve started com­ment­ing here re­cently, but I’m a long time lurker (>1 year). Also, I was speak­ing about self-help ar­ti­cles in gen­eral, not con­di­tional on whether they are posted on LW—it makes sense, be­cause pretty much any­one can post on LW.

Now I found a some­what less ex­treme ex­am­ple of what I think is an OK post on self-help al­though it doesn’t have sci­en­tific refer­ences, be­cause a) the au­thor told us what ac­tual re­sults he achieved and, more im­por­tantly, b) the au­thor ex­plained why he thinks that the ad­vice works in the first place.

Per­son­ally, I don’t find your post con­sis­tent with my ob­ser­va­tions, but it’s not my main ob­jec­tion—my main ob­jec­tion is that throw­ing an in­struc­tion with­out any jus­tifi­ca­tion is a bad prac­tice, es­pe­cially on such a con­tro­ver­sial topic, es­pe­cially in a ra­tio­nal­ist com­mu­nity.

• That’s to­tally fine, like I said, your post made sense and was con­sis­tent with what I’ve seen.

I still don’t re­ally think that stat­ing my qual­ifi­ca­tions would do much. In this con­text, it still just seems too much like brag­ging. “I helped build a multi-mil­lion dol­lar com­pany, I com­pete in bar­be­cue com­pe­ti­tions and con­sis­tently place in the top 10% of the field and was spon­sored by a ma­jor bar­be­cue web­site, was ranked in the top 100 play­ers in the world for a pop­u­lar col­lectible card game, learned how to code with no for­mal ed­u­ca­tion (and used that knowl­edge wrote a some­what well-re­ceived cal­ibra­tion test, and also write a bunch of bor­ing busi­ness plat­forms), wrote an ar­ti­cle about a base­ball statis­tic I co-de­vel­oped and was pub­lished in a pub­li­ca­tion that’s im­por­tant for peo­ple who care about base­ball stats, learned how to be a car­pen­ter, at one point was a li­censed phar­macy tech­ni­cian, blah blah blah”

Even though I’m sure there’s a less crass way to phrase it, to me it still sounds ex­ceed­ingly ar­ro­gant. I might be over­re­act­ing though. You tell me: if I pref­aced my post with that, would you be more or less in­clined to take me se­ri­ously?

I do like the idea of ex­plain­ing why I think the ad­vice works in the first place. I will start writ­ing some­thing up about that and ap­pend it to the origi­nal post.

• I do like the idea of ex­plain­ing why I think the ad­vice works in the first place.

If I may sug­gest spend­ing some space on ex­plain­ing why do you think your ex­pe­rience gen­er­al­izes—that is, why do you think that your meth­ods will work for peo­ple who are not you.

• I took your ad­vice as well as es­ti­ma­tor’s into ac­count and added two para­graphs at the be­gin­ning to offer 1. Some re­search show­ing that many sys­tems fol­low a dis­tri­bu­tion where a small por­tion of work ac­counts for a large por­tion of re­sults, and 2. and ex­pla­na­tion as to why it’s gen­er­al­iz­able.

• Also, I’d like to com­pare your sys­tem against com­mon sense rea­son­ing baseline. What do you think are the main differ­ences be­tween your ap­proach and usual ap­proaches to skill learn­ing? What will be the differ­ence in ac­tions?

I’m ask­ing that be­cause that your guide con­tains quite long a list of recom­men­da­tions/​ac­tions, while many of them are used (prob­a­bly in­tu­itively/​im­plic­itly) by al­most any sen­si­ble per­son. Also, some of the recom­men­da­tions clearly have more im­pact than oth­ers. So, what hap­pens if we ap­ply the Pareto prin­ci­ple to your learn­ing sys­tem? Which 20% are the most im­por­tant? What is at the core of your ap­proach?

• As I men­tioned in an­other com­ment, the differ­ence be­tween this and the “com­mon sense” ap­proach is in what this sys­tem does not do.

As for what the 20% of this sys­tem that gives you the most bang for your buck? That’s a good ques­tion. Right now my “safe” an­swer is that it’s de­pen­dent on the type of skill you’re try­ing to learn. The trou­ble is that the com­mon threads among all the skills (“Find the 20% of the skill that yields 80% of the re­sults”) doesn’t have a lot of prac­ti­cal value. Like tel­ling some­one that all they need to do to lose weight is eat less and ex­er­cise more.

Let me think about it some more and I’ll get back to you.

• So, af­ter some cur­sory thought, nat­u­rally the part of the sys­tem that gives you the most bang for your buck are the first 4 steps. The last 3 steps are de­signed to help you im­prove, which is a much slower pro­cess than just learn­ing the ba­sics.

So, now to figure out how to re­cur­sively ap­ply the the skill of learn­ing a skill quickly to the skill “learn­ing skills quickly”.

• Okay, so I made a sig­nifi­cant re­vi­sion of the post. The origi­nal ideas are all there, just writ­ten in a much less ob­tuse man­ner.

• A much more log­i­cal ar­gu­ment is pre­sented at the be­gin­ning, along with con­straints.

• “Archetypes” and “Pro­cesses” have been re­placed by sub-skills and triv­ial sub-skills.

• The lengthy dis­course on strat­egy has been re­placed by sim­ply sort­ing your list of triv­ial sub-skills, which ac­com­plishes the same effect.

• The “im­prove­ment” has been stream­lined greatly.

• Meta-anal­y­sis has been re­moved be­cause it’s re­ally a sep­a­rate sub­ject.

• One piece of in­for­ma­tion you can use to de­ter­mine what is most im­por­tant is the num­ber of other skills which re­quire a cer­tain skill as a pre­req­ui­site. Pr­ereq­ui­sites should ob­vi­ously be learned first, and it makes sense to learn them in or­der of how many doors they open. This is how I pri­ori­tize at the mo­ment if I’m not con­sid­er­ing sub­jec­tive mea­sures of “use­ful­ness”.

For my learn­ing goals, I’ve started mak­ing con­cept maps, partly as it helps me un­der­stand a sub­ject by un­der­stand­ing how con­cepts are re­lated, and partly to iden­tify what to learn next as de­scribed above. It be­comes fairly ob­vi­ous that I should learn X if I want to learn Y and Z and X is a pre­req­ui­site for both.

• In my ex­pe­rience, in math/​sci­ence pre­req­ui­sites of­ten can (and should) be ig­nored, and learned as you ac­tu­ally need them. Peo­ple who thor­oughly fol­low all the pre­req­ui­sites of­ten end up bogged down in nu­mer­ous sci­ence fields which have ac­tu­ally weak con­nec­tion to what they wanted to learn ini­tially, and then get de­mo­ti­vated and drop out of their en­deavor. This is a com­mon failure mode.

Like, you need prob­a­bil­ity the­ory to do ma­chine learn­ing, but some you are un­likely to en­counter some parts of it, and also there are parts of ML which re­quire very lit­tle of it. It to­tally makes sense to start with them.

• I’m think­ing more speci­fi­cally than you are. Rather than learn­ing prob­a­bil­ity the­ory to un­der­stand ML, learn only what you de­ter­mine to be nec­es­sary for what ML ap­pli­ca­tions you are in­ter­ested in. The con­cept maps I use are very spe­cific, and they avoid the weak con­nec­tion prob­lem you men­tion. (It’s worth not­ing that I de­velop these as an au­to­di­dact, so I don’t have to take an en­tire class to just get a few facts I’m in­ter­ested in.)

• It sounds like both you and es­ti­ma­tor are ac­tu­ally both on the same page: es­ti­ma­tor seems to be talk­ing about the “pre­req­ui­site” in the sense of, “sys­tem­atic pre­req­ui­site”, as in, peo­ple say that you should learn X be­fore you learn Y. You seem to be talk­ing about “pre­req­ui­site” in the sense that, “skill X is a nec­es­sary com­po­nent of skill Y

Both of you, how­ever, seem to agree that you should ig­nore the stuff that is ir­rele­vant to what you are ac­tu­ally try­ing to ac­com­plish.

• This is a good way to put it. I may not have been clear.

To use an ex­am­ple, I have a con­cept map about fluid dy­nam­ics that I used in a class I took on tur­bu­lence re­cently. There were a few con­cepts that I did not un­der­stand well at some point, and I wanted to figure out which ones. To be more spe­cific, isotropic ten­sors are of­ten used in tur­bu­lence the­ory and mod­el­ing, but I didn’t re­ally un­der­stand how to con­struct isotropic ten­sors alge­braically. It be­came pretty clear this was some­thing I should learn given the num­ber of links isotropic ten­sors had to other con­cepts.

• On the other hand, if you don’t have a solid grasp of lin­ear alge­bra, your abil­ity to do most types of ma­chine learn­ing is se­ri­ously im­paired. You can learn tech­niques like e.g. ma­trix in­ver­sions as needed to im­ple­ment the al­gorithms you’re learn­ing, but if you don’t un­der­stand how those tech­niques work in their origi­nal con­text, they be­come very hard to de­bug or op­ti­mize. Similarly for e.g. cryp­tog­ra­phy and ba­sic in­for­ma­tion the­ory.

That’s prob­a­bly more the ex­cep­tion than the rule, though; I sense that the point of most pre­req­ui­sites in a tra­di­tional sci­ence cur­ricu­lum is less to provide skills to build on and more to build habits of rigor­ous think­ing.

• Read what is a ma­trix, how to add, mul­ti­ply and in­vert them, what is a de­ter­mi­nant and what is an eigen­vec­tor and that’s enough to get you started. There are many al­gorithms in ML where vec­tors/​ma­tri­ces are used mostly as a handy no­ta­tion.

Yes, you will be un­able to un­der­stand some parts of ML which sub­stan­tially re­quire lin­ear alge­bra; yes, un­der­stand­ing ML with­out lin­ear alge­bra is harder; yes, you need lin­ear alge­bra for al­most any kind of se­ri­ous ML re­search—but it doesn’t mean that you have to spend a few years study­ing ar­cane math be­fore you can open a ML text­book.

• Who said any­thing about a few years? If you paid at­ten­tion in high school, the lin­ear alge­bra back­ground you need is at most a few months’ worth of work. I was pro­vid­ing a sin­gle coun­terex­am­ple, not say­ing that the full pre­req­ui­site list (which, if mem­ory serves, is most of a CS cur­ricu­lum for your av­er­age ML class) is always nec­es­sary.

• if you don’t have a solid grasp of lin­ear alge­bra, your abil­ity to do most types of ma­chine learn­ing is se­ri­ously impaired

That de­pends on whether you’re do­ing re­search or purely ap­plied stuff. For ap­plied use, do­main ex­per­tise trumps know­ing the in­ter­nal de­tails of the al­gorithms which you usu­ally just call as pre-build func­tions—as long as you un­der­stand what do they do and where the limits (and the traps) are.

Not many peo­ple can in­vert ma­tri­ces by hand any more and that’s not a prob­lem for a higher-level un­der­stand­ing of lin­ear alge­bra. Similarly, you don’t nec­es­sar­ily need to un­der­stand, say, how sin­gu­lar value de­com­po­si­tion works in or­der to do suc­cess­ful higher-level mod­el­ing of some do­main.

• I wasn’t point­ing strictly to re­search, but I was point­ing to low-level im­ple­men­ta­tion. It now oc­curs to me that I might be un­usual in this re­spect—much of my ML ex­pe­rience is in the con­text of a rather weird en­vi­ron­ment that didn’t have any ex­ist­ing libraries, leav­ing me to cut a lot of code my­self.

So I might have to back off from “abil­ity to do ma­chine learn­ing”. You can, in ret­ro­spect, use ML perfectly com­pe­tently in a lot of set­tings even if the clos­est you’ve ever got­ten to a simu­lated an­neal­ing al­gorithm is plug­ging the cost func­tion into a Python library; but I have a hard time call­ing some­one an ex­pert if they’ve never writ­ten any­thing lower-level, just as I’d ex­pect a com­pe­tent soft­ware en­g­ineer to be able to write a hash table by hand even if ev­ery en­vi­ron­ment they’re likely to en­counter will have built-in im­ple­men­ta­tions or at least effi­cient libraries for it.

• just as I’d ex­pect a com­pe­tent soft­ware en­g­ineer to be able to write a hash table by hand even if ev­ery en­vi­ron­ment they’re likely to en­counter will have built-in im­ple­men­ta­tions or at least effi­cient libraries for it.

I have a feel­ing that’s a bit of a relic.

Long time ago pro­gram­ming en­vi­ron­ments were prim­i­tive and Real Men wrote their own sorts and hash ta­bles (there is a canon­i­cal story from even more An­cient Times). But times have changed. I can’t imag­ine a se­ri­ous situ­a­tion (as op­posed to e.g. a pro­gram­ming con­test) where I would have to write my own sort rou­tine from scratch—similarly to how I can’t imag­ine need­ing to con­struct a wash­ing ma­chine out of a tub, an elec­tric mo­tor, pul­leys, and belts.

I cer­tainly still care about perfor­mance prop­er­ties of var­i­ous sorts, but I don’t care about their in­ter­nal de­tails as long as the perfor­mance prop­er­ties hold. I sus­pect that the in­ter­view ques­tions of the “im­ple­ment a bub­ble sort on this piece of pa­per” va­ri­ety if any­thing aim more at “have you been pay­ing at­ten­tion dur­ing your CS classes” and less at “do you have a deep un­der­stand­ing of what your pro­gram is do­ing”.

The ca­pac­ity of hu­man minds is limited and I’ll ac­cept climb­ing up higher in ab­strac­tion lev­els at the price of for­get­ting how the lower-level gears turn.

• I can’t imag­ine a se­ri­ous situ­a­tion (as op­posed to e.g. a pro­gram­ming con­test) where I would have to write my own sort rou­tine from scratch

You can’t? I’ve had to do that sev­eral times. The usual sce­nario is that there are search/​sort rou­tines, but they have in­con­ve­nient prop­er­ties—ei­ther they don’t perform well in the spe­cific prob­lem do­main I’m deal­ing with (hap­pens a lot in simu­la­tion; func­tions for effi­ciently do­ing cer­tain types of cat­e­go­riza­tion on spa­tially ar­ranged data are rare out­side graph­ics libraries), or they don’t work on the data types I need and a re­duc­tion is im­prac­ti­cal for one rea­son or an­other, or they ex­ist but can’t be used for le­gal rea­sons. Un­less you always situ­ate your­self in the most pop­u­lar sub­fields, which I frankly find bor­ing, you can’t always count on there be­ing a library that does ex­actly what you want—all the more so in a still-emerg­ing space like ML.

(I’ve never had to build a wash­ing ma­chine, in­ci­den­tally, but I’ve had to fix wash­ing ma­chines—twice this year for two differ­ent ma­chines, in fact. I could have hired a me­chanic or bought a new ma­chine, but ei­ther one would have cost me hun­dreds of dol­lars.)

• You can’t? I’ve had to do that sev­eral times.

Well, I was talk­ing about stan­dard sort rou­tines—the ones where you have a vec­tor of val­ues and a com­para­tor func­tion. Now, search is quite a differ­ent beast al­to­gether.

The thing is, most sort­ing is brute-force where you just sort with­out tak­ing into ac­count the spe­cific struc­ture of your data. That ap­proach works well with sort­ing—but it doesn’t work well with search. The ob­vi­ous prob­lem is that we are in­ter­ested in search­ing very large search spaces where brute force is nowhere close to prac­ti­cal. The sal­va­tion comes from the par­tic­u­lar struc­ture of the space which al­lows us to be much more effi­cient that brute-force, but the same struc­ture forces us into cus­tom solu­tions.

Be­cause the struc­tures of search spaces can be very differ­ent, there is a LOT of search al­gorithms and fre­quently enough you have to make be­spoke ver­sions of them to fit your par­tic­u­lar prob­lem. That’s en­tirely nor­mal. Plus, of course, op­ti­miza­tion is a sub­type of search and cus­tomiz­ing op­ti­miz­ers is also quite com­mon.

but I’ve had to fix wash­ing machines

Sure, so have I. In fact, I prob­a­bly would be able to con­struct a wash­ing ma­chine out of a tub, an elec­tric mo­tor, and some parts. It will take a lot of time and will look ugly, but I think it’ll work. That doesn’t mean I’ll feel a need to do this :-)

• Yes, this this this this this this this. “The ca­pac­ity of hu­man minds is limited and I’ll ac­cept climb­ing up higher in ab­strac­tion lev­els at the price of for­get­ting how the lower-level gears turn.” If I could up­vote this mul­ti­ple times, I would.

This is the crux of this en­tire ap­proach. Learn the higher level, ap­plied ab­strac­tions. And learn the very ba­sic fun­da­men­tals. For­get learn­ing how the lower-level gears turn: just learn the fun­da­men­tal laws of physics. If you ever need to figure out a lower-level gear, you can just de­rive it from your knowl­edge of the fun­da­men­tals, com­bined with your big-pic­ture knowl­edge of how that gear fits into the over­all sys­tem.

• That only works if there are few lev­els of ab­strac­tion; I doubt that you can de­rive how do pro­grams work at the ma­chine codes level based of your knowl­edge of physics and high-level pro­gram­ming. Some­times, gears are so small that you can’t even see them on your top level big pic­ture, and some­times just climb­ing up one level of ab­strac­tion takes enor­mous effort if you don’t know in ad­vance how to do it.

I think that you should un­der­stand, at least once, how the sys­tem works on each level and re­fresh/​deepen that knowl­edge when you need it.

• The defi­ni­tion of “fun­da­men­tals” differs though, de­pend­ing on how ab­stract you get. The more lay­ers of ab­strac­tion, the more ab­stract the fun­da­men­tals. If my goal is high-level pro­gram­ming, I don’t need to know how to write code on bare metal.

That’s why I ad­vo­cate break­ing things down un­til you reach the level of triv­ial­ity for you per­son­ally. Most peo­ple will find, “writ­ing a for-loop” to be triv­ial, with­out hav­ing to go farther down the rab­bit hole. At a cer­tain point, break­ing things down too far ac­tu­ally makes things less triv­ial.

• Can I give a coun­terex­am­ple? I think that way of learn­ing things might help if you only need to ap­ply the higher-level skills as you learned them, but if you need to de­velop or re­search those fields your­self, I’ve found you re­ally do need the back­ground.

As in, I have been bit­ten on the ass by my own choice not to dou­ble-ma­jor in math­e­mat­ics in un­der­grad, thus re­sult­ing in my hav­ing to start climb­ing the tow­ers of con­tin­u­ous prob­a­bil­ity and statis­tics/​ML, ab­stract alge­bra, logic, real anal­y­sis, cat­e­gory the­ory, and topol­ogy in and af­ter my MSc.

• There’s a big differ­ence be­tween the fun­da­men­tals, and the low-level prac­ti­cal ap­pli­ca­tions. I think the lat­ter is what es­ti­ma­tor is refer­ring to. You can’t re­ally make a break­through or do real re­search with­out a firm grasp of the fun­da­men­tals. But you definitely can make a break­through in, say, physics, with­out know­ing the ex­act ten­sile strength of wood vs. steel. And yet, that type of “Ap­plied Physics” was a pre-req­ui­site at my school for the more ad­vanced fields of physics that I was ac­tu­ally in­ter­ested in.

• And yet, that type of “Ap­plied Physics” was a pre-req­ui­site at my school for the more ad­vanced fields of physics that I was ac­tu­ally in­ter­ested in.

Oh. Really? Dang.

• You’re right; you have to learn solid back­ground for re­search. But still, it of­ten makes sense to learn in the re­versed or­der.

• Nice, but be­ware rea­son­ing af­ter you’ve writ­ten the bot­tom line.

As for the ac­tual con­tent, I ba­si­cally fail to see its area of ap­pli­ca­bil­ity. For suffi­ciently com­plex skills, like say, math, lan­guages or foot­ball de­ci­sion-trees & howto-guides ap­proach will likely fail as too shal­low; for iso­lated skills like chang­ing a tire com­plex learn­ing ap­proaches are an overkill—just google it and fol­low the in­struc­tions. Can you elab­o­rate lan­guages ex­am­ple fur­ther? Be­cause, you know, learn­ing a bunch of phrases from phrase­book to be able to say a few words in a for­eign coun­try is a non-is­sue. Ac­tu­ally learn­ing lan­guage is. How would you ap­ply your sys­tem to achieve in­ter­me­di­ate-level lan­guage knowl­edge? Any other non-triv­ial skills learn­ing ex­am­ple would also suffice. What skills have you trained by us­ing your learn­ing sys­tem, and how?

• Also, when you say “in­ter­me­di­ate level lan­guage knowl­edge”, what ex­actly do you mean? One of the key steps is defin­ing ex­actly what you want to ac­com­plish and why. I don’t want to cre­ate a whole write-up, only to re­al­ize that you and I have two differ­ent defi­ni­tions of “in­ter­me­di­ate level lan­guage knowl­edge”.

So if you’d tell me the “what” and the “why”, I’ll do the rest.

• I meant some­thing like this.

… take part in rou­tine con­ver­sa­tions; write & un­der­stand sim­ple writ­ten text; make notes & un­der­stand most of the gen­eral mean­ing of lec­tures, meet­ings, TV pro­grammes and ex­tract ba­sic in­for­ma­tion from a writ­ten doc­u­ment.

• I’ll give a more in depth break­down soon but for now, I’d prob­a­bly take a similar ap­proach that I took to learn­ing to read Ja­panese : learn ba­sic sen­tence struc­ture, learn top 150ish vo­cab­u­lary words, avoid books writ­ten in non-ro­maji. Prac­tice hear­ing spo­ken word by listen­ing to speeches and fol­low­ing their tran­scrip­tions. My ex­cep­tion pro­to­col for un­rec­og­nized words was to look them up. And for ir­reg­u­lar sen­tence struc­ture, to guess based on con­text. It worked for watch­ing movies and read­ing, mostly but as you can tell, yoi kak­ikomu koto ga deki­masen*. I’d have to do some think­ing on the writ­ing part, it would most likely in­volve stick­ing to sim­ple sen­tences.

*thats ter­rible Ja­panese for “I can­not write well”. I think. I hope.

• But these are the things pretty much ev­ery­body does while learn­ing lan­guages.

• Well of course they do. Be­cause these things are nec­es­sary to learn­ing a lan­guage. This is the 20% that’s most effi­cient. By defi­ni­tion some­one who puts in 100% of the effort will be do­ing what I did.

The effi­ciency of this ap­proach re­volves around what you don’t do. You’re ex­cis­ing the 80%. I didn’t spend long hours learn­ing katakana, hira­gana and kanji. I didn’t learn the more com­plex tenses and con­ju­ga­tions. I didn’t spend time on vo­cab­u­lary words that are highly situ­a­tional. Con­trast this to a typ­i­cal Ja­panese text­book.

• There seem to be two ma­jor ap­proaches to learn­ing lan­guage.

One is to go a lan­guage school /​ course where the teach­ers, in my ex­pe­rience, teach it like an aca­demic dis­ci­pline + the usual guess-my-pass­word bul­lshit, so you get tested and graded on things like gram­mar, like a test where you need to fill in con­ju­ga­tions /​ de­cli­na­tions into holes in a text. (Ob­vi­ously I am talk­ing about lan­guages that have those kinds of things, like Ger­manic or Ro­mance ones). Case in point: part of my B2 level Ger­man exam at the Univer­sity of Vienna was ex­actly that kind of hole-filling and it felt re­ally wrong as it has not much to do with com­muica­tion, it is a more aca­demic ap­proach.

The other ap­proach is to do some­thing like this for a while, but when you get to that ba­sic point where you can say “Jack would have or­dered a beer yes­ter­day if he had money on him” ditch it and pretty much learn from im­mer­sion. Screw gram­mar, just read a lot of books, figure out words from the con­text, and con­duct imag­i­nary or real con­ver­sa­tions no mat­ter how bad the gram­mar is. Real peo­ple pre­fer to com­mu­ni­cate with peo­ple who talk fast, not cor­rect. Talk­ing with some­one say­ing at a nor­mal speed who is talk­ing like “me no want buy house, me want rent house now” is far bet­ter than some­one who is like “I no… (long pause) do not? want … (long pause) want to? buy a house, rather… (long pause)… in­stead? I want to rent it… (long pause) rent one”. I used to be that sec­ond guy in 2 lan­guages and it sucked.

(Now of course you may think “but ev­ery­body knows im­mer­sion is bet­ter it is not even new” yeah ap­par­ently that ev­ery­body does not in­clude the huge Euro­pean lan­guage school chains like Ber­litz and their who knows how many stu­dents… )

• Bas­ket­ball is an ex­am­ple. I’m about to head home so I’ll do the ul­tra-ab­bre­vi­ated TL;DR ver­sion:

1. Goals: Score points, pre­vent op­po­nent from scor­ing points.

2. Archetypes: Offense (2-point), Offense (3-point), Defense

3. Pro­cess How-To: Googled “how to layup”, “how to shoot a 3-poin­ter”, and “how to steal a ball” 3a. Pro­cess Failure Points: Miss­ing a shot, get­ting the ball stolen, miss­ing a pass. 3b. Pro­cess Difficul­ties: Any­thing in­volv­ing ball han­dling or drib­bling. Defense.

4. Ex­cep­tion Pro­to­col: On offense: Pass the ball to a bet­ter player than my­self, or set a pick. On defense: play very close to my op­po­nent. 5a. Avoid any­thing in­volv­ing drib­bling but not scor­ing. 5b. Pre­pare and prac­tice two-point shots. 5c. Fo­cus on get­ting open for a 3-point shot. Prac­tice con­sis­tently shoot­ing from 3-point line.

5. Get bet­ter by play­ing.

I would say bas­ket­ball is fairly com­plex. One thing I didn’t men­tion in the origi­nal post (mainly be­cause it starts to get into the “how do in­di­vi­d­ual peo­ple learn”) but for me, I don’t get good at a com­pet­i­tive skill by com­pet­ing against peo­ple who also suck. By get­ting good enough to be able to play with peo­ple who are ac­tu­ally good, it made it eas­ier for me to learn the ad­vanced part of the game faster.

Also, this post has a list of (at least what I think to be) fairly non-triv­ial skills that I have trained us­ing this method.

• Gen­er­ally I’ve found that the more some­one ex­pounds on their own cre­den­tials, the less cred­ible (and lik­able) they sound.

This is true, par­tic­u­larly be­cause peo­ple are differ­ent, so if you tell us a lot about your­self, you’re tel­ling us com­par­a­tively lit­tle that ap­plies to, well, us. That’s why we of­ten want some kind of or­ga­nized, even if in­for­mal, study on a group of peo­ple, par­tic­u­larly a group in it with enough vari­a­tion that per­sonal unique­ness washes out.

• Per­sonal achieve­ments are pretty bad as ev­i­dence. But they can still be bet­ter than no ev­i­dence.

(And per­sonal achieve­ments aren’t all he asked for. I don’t think refer­ences to rele­vant re­search would have the same prob­lem as per­sonal achieve­ments here.)

• I agree that in some cases, it’s bet­ter than noth­ing to in­clude per­sonal achieve­ments (as I did when I was dis­cussing so­cial­iz­ing in an­other thread). I just don’t re­ally think that’s the case here. I’ll say the same thing that I said to es­ti­ma­tor: if you gen­uinely think that my per­sonal achieve­ments would make a differ­ence to you, I’ll be glad to tell you.

As for rele­vant re­search, well, (and I might be wrong on this) I thought one of the pur­poses of LW was to pro­duce origi­nal con­tent. Again, I might be mis­in­ter­pret­ing things here. But if there was re­search that said, “Such-and-such ap­proach to skill-learn­ing works well”, why not just link to that in­stead of try­ing to para­phrase it?

• Here’s an ex­am­ple of this pro­cess ap­plied to learn­ing a for­eign lan­guage:

Define the goal: I want to be able to a. in­ter­act and b. con­verse and c. func­tion in a so­ciety that speaks a differ­ent lan­guage.

Archetypes: Most of the time I spend talk­ing with oth­ers in per­son is spent a. eat­ing/​drink­ing/​buy­ing things, b.ask­ing for as­sis­tance, c. meet­ing new peo­ple. To break those down into sub­types, I’d say:

• Order­ing food

• Order­ing drinks

• Need help

Pro­cesses: Learn to say the fol­low­ing sen­tences for each archetype, along with the var­i­ous vo­cab­u­lary words. Also list out pos­si­ble re­sponses to each sen­tence and learn to un­der­stand them.

• Food: I’d like to or­der [food], Is [food item] [ad­jec­tive]?, [Com­pli­ments], [Com­plaints],

• Drinks: Can I have a [drink], Another [drink], Do you have [spe­cific drink]?, No thanks.

• Prod­ucts: Do you carry [product]?, I am look­ing for [de­scrip­tion].

• Land­marks: Where is [lo­ca­tion]?, Thank you.

• Help: I need help, I have [con­di­tion], I am sick, Where is the hos­pi­tal?, I am from [coun­try], I am stay­ing at [place], My emer­gency con­tact is [per­son].

• Small Talk Ques­tions: Hi, How are you?, Where are you from?, How do you like it here? What do you do?

• Small Talk An­swers: I’m do­ing well, I’m from [coun­try], It’s great here, I like [thing], I don’t like [thing], I am a [oc­cu­pa­tion].

Ex­cep­tion Pro­to­col: There are two main non-re­dun­dant failure points: “Some­one uses a vo­cab­u­lary word I don’t rec­og­nize” and “There is a com­plete gap in un­der­stand­ing.” Two differ­ent ex­cep­tion pro­to­cols can han­dle these, de­pend­ing on the situ­a­tion:

• Learn to say, “Sorry, I’m from Amer­ica and don’t speak [lan­guage] as a first lan­guage. Do you know English… or, can you say that again?”

• Use Google Trans­late to trans­late what I am try­ing to com­mu­ni­cate.

Strat­egy: Avoid situ­a­tions that have the po­ten­tial to be­come high-stakes. Pre­pare for situ­a­tions where it may be loud/​fast-paced and hard to un­der­stand peo­ple. Fo­cus on sce­nar­ios that in­volve fairly low-stakes in­ter­ac­tions and sim­ple con­ver­sa­tions. Prac­tice my “Sorry, I’m from Amer­ica” line un­til I can say it and sound good. Make sure my Google Trans­late app is read­ily available.

Im­prove­ment. Any time I have to use my ex­cep­tion pro­to­col, make sure to look up the sen­tence that I was try­ing to say or was be­ing said to me, and re­mem­ber it. Take note of any re­cur­ring sce­nar­ios that I haven’t learned an ap­pro­pri­ate sen­tence for. Ex­pand my vo­cab­u­lary list for the sce­nar­ios that are most com­mon.

• So, tak­ing a look at what you ac­tu­ally pro­pose to do, this re­duces to a) learn some phrases from the tourist phrase­book and b) learn the rest of the lan­guage while c) avoid­ing high-stakes situ­a­tions where you need lan­guage knowl­edge. Re­minds me of this.

• That may be a bit more snarky than is helpful. Your re­duc­tion loses use­ful in­for­ma­tion; Nanashi’s longer de­scrip­tion of the pro­cess in­cludes use­ful, spe­cific pro­ce­du­ral de­tails that could oth­er­wise trip peo­ple up.

• Yup, pretty much. To quote myself

TL;DR: The fastest way to learn new skills is to 1. Break it down into enough “recipes” or “how-to” guides that they cover most of what you might en­counter, and 2. Figure out how to elo­quently ask for help if you don’t know what to do.

(In­ci­den­tally, the link you posted does not work, it’s giv­ing me a 404).

• I’m con­fused about the “strat­egy” sec­tion; it seems largely re­dun­dant with the ear­lier parts.

• Bad edit­ing on my part. Ill up­date the post and in­clude the origi­nal here for pos­ter­ity

• Although you were fo­cus­ing on learn­ing skills, your lan­guage ex­am­ple made me think of so­cial ram­ifi­ca­tions, and I wanted to say that I call things like this ‘pre­pro­cess­ing’ more gen­er­ally. It’s sort of like a men­tal ver­sion of Crazy-Pre­pared. I think of writ­ing as a form of it, and it’s one rea­son I’m glad that most LW users are very far from me in meatspace. I can’t imag­ine hav­ing to en­gage in per­son on a level that I some­times must here. I think that there’s an ex­pe­ri­en­tial com­po­nent and an in­nate com­po­nent, like with most things, and Wei_Dai has brought up this differ­ence in pass­ing:

While I look for­ward to talk­ing to Eliezer and you, I do have a con­cern, namely that I find Eliezer to be much bet­ter (ei­ther more na­tively tal­ented, or more prac­ticed, prob­a­bly both) than I am at mak­ing ar­gu­ments in real time, while I tend to be bet­ter able to hold my own in offline for­mats like email/​blog dis­cus­sions where I can take my time to figure out what points I want to make. So keep that in mind if the chat ends up be­ing re­ally one-sided.

Some peo­ple seem to be able to perform what I con­sider very com­plex tasks on a merely per­cep­tual level, and I don’t feel like one of these peo­ple, so it re­ally helps to do a bit of Sys­tem 2 plan­ning be­fore situ­a­tions where oth­ers could just go with the flow. I also find that it’s a way to make it eas­ier to have more con­fi­dence and at­ten­tion in so­cial situ­a­tions; you’re not wor­ry­ing about non-so­cial de­tails.

• 28 May 2015 19:03 UTC
1 point

That was use­ful. I tried to com­pare it to my own at­tempts to learn a new tech­nique in microscopy and saw I had not for­mal­ized ex­cep­tions (though I still would be at a loss as to how do it, even hav­ing gained some ex­pe­rience). It went Luke this:

Goals: study my­c­or­rhiza, 1) in roots, 2) in the sur­round­ing soil. 2) quickly was shown to be un­fea­si­ble (too many stages in sam­ple prepa­ra­tions that re­quire ac­cess to equip­ment like a cen­trifuge, which is for me ridicu­lously in­con­ve­nient /​ no­body here available to check the end re­sult /​ peo­ple who do it in earnest use DNA anal­y­sis for species iden­ti­fi­ca­tion, which is an­other skill I don’t pos­sess and re­quires primers that cost a lot /​...), so I con­cen­trated on 1). In part, I did it in the hope that af­ter I have a cou­ple pub­li­ca­tions on the topic, I can find col­lab­o­ra­tors to whom I’d be able to out­source 2) - an analogue of your Google trans­late:)

Then, I chose the ways of prepar­ing roots to prac­tice (since there are about 10 stains com­monly used and they don’t give the same re­sults). Those were my ‘archetypes’, as I saw them ini­tially. How­ever, when I saw my first slides un­der the micro­scope, I un­der­stood that I did not even cut the roots evenly enough (and so some of them just didn’t have the cor­tex where the fungi should have been) and did not place the roots in clear par­allels on the slide (they over­lapped, etc.) The cut­ting is done the same way be­fore any kind of stain­ing, andthe plac­ing—the same way af­ter it.

So I sucked it up (in­ci­den­tally, 10% w/​v NaOH seems much worse on your nose than KOH when you make it out­side a hood) and chugged out sev­eral dozen well-packed never-mind-stain­ing slides. This part was eas­iest to self-cor­rect. Then I came back to my ‘archetypes’ - Ink af­ter Vier­heilig et al., Try­pan blue and acid Fuchsin, and now creep to­wards good stain­ing qual­ity.

There are ‘ex­ter­nal forces I can­not con­trol’ that limit my abil­ity to make prepa­ra­tions, and I do have a jour­nal of mis­takes, but other than that, I do not keep track of things. Your model is much more rigor­ous, and (to me) seems to as­sume quick feed­back (or an iron dis­ci­pline, or both:) I don’t think I could serenely wait for hours be­fore I knew what to cor­rect if I used such ap­proach.

• 19 Aug 2015 18:16 UTC
0 points

Equally or al­most as im­por­tant meta-skills/​skills any­one can learn are how to change your be­liefs, how to ask ques­tions and how to un­der­stand an­swers.

• Since we’re try­ing to be less­wrong here, I’ll risk seem­ing petty by point­ing out that “beg­ging the ques­tion” is a log­i­cal fal­lacy, not a syn­onym for “rais­ing the ques­tion”. Just sayin...

• Although tech­ni­cally you could say that the whole ar­gu­ment begs the ques­tion, de­pend­ing on how you in­ter­pret the logic. Be­cause it ba­si­cally fol­lows the form: “Learn­ing a skill is triv­ial be­cause you can break a skill down into triv­ial sub­skills.”

• Prac­tice my “Sorry, I’m from Amer­ica” line

Err.. that would be the “Why can’t you damned fur­riners speak ’Mer­i­can like all the reg­u­lar folk!” line. You may have been think­ing of an “I would like to ab­jectly apol­o­gize for be­ing a Cana­dian” line… X-D

• To non-na­tive speak­ers like my­self Brits can be fore more difficult be­cause their lo­cal ac­cents /​ di­alects can de­vi­ate far more from lan­guage schools stan­dards than Amer­i­can ones. I was se­ri­ously dumb­founded the first time some­one from Bir­m­ing­ham greeted me as “O’royt, moyt!”. One fa­vorite: Stock­port = shhtoffpfo, Stafford = shtaoffpfo so I had to be care­ful where to buy the train ticket to.

I used to think it must be a lower class thing, but I have seen en­g­ineers talk like this in e.g. Wolver­hamp­ton.

On top if it, there are eth­nic ver­sions. Chi­nese-Bir­m­ing­hamese is es­pe­cially hard to un­der­stand, the above greet­ing sounds like “Ohoy, moyh!”

Scot­land is very in­ter­est­ing, in one place peo­ple talk like the most difficult lines in Trainspot­ting (“Hazshu gozshda bish in shour shoightsh?”), and in other places in the usual Hol­ly­wood-Amer­i­can one from the movies (which prob­a­bly origi­nated from there).

Ire­land… friend of mine in­ter­viewed an Ir­ish DJ, it took a while to figure out “heiss mu­sic” means “house mu­sic”.

I have heard but did not ver­ify that ’Straya down there is pretty much Bir­m­ing­ham, ac­cent-wise, nice = noice etc.

Com­pared to this, Amer­i­cans are easy, ex­cept those who speak like Jay Leno.

I would re­ally like a com­edy scene when two guys ar­gue, one is talk­ing in Brum­mie, Scul­lie or a similar thicker Bri­tish di­alect, and other in ei­ther AAVE or Ja­maican like Lit­tle Ja­cob in GTA IV.

Note: there are similar things go­ing on with e.g. the Ger­man lan­guage as well, Frank­furt meet­ing Vo­rar­lberg is a similar com­edy.

How­ever in smaller Eastern Euro­pean coun­tries pretty much ev­ery­body speaks “TV lan­guage” (not the offi­cial but close, offi­cial + slang) and I still won­der how comes TV did not kill lo­cal ac­cents and di­alects yet in ev­ery coun­try. I was re­ally used to ev­ery­body talk­ing like TV and this not be­ing a prob­lem at all.

• Warn­ing: a time sink :-) Au­dio record­ings of Bri­tish di­alects.

• Serbo-Croa­t­ian used to be one lan­guage with di­alects. After the coun­tries split it be­came two. A lot of the Ser­bian lan­guages have mu­tual in­tel­ligi­bil­ity. Smaller Eastern Euro­pean coun­tries just named their most pop­u­lar di­alect it’s own lan­guage.

Com­mu­nism is also likely to play a huge part. Part of the Marx­ist idea of ac­cel­er­at­ing progress is to get rid of di­alects and make sure that ev­ery­body learns the lo­cal high lan­guage in kinder­gar­den and school.

Diver­sity in the Ger­man lan­guage is also down. I had a uni­ver­sity friend who came from a Bavar­ian town that had a di­alect that wasn’t able to be un­der­stood in Mu­nich. Maybe <10,000 peo­ple could un­der­stand it. It did weird things like do­ing new word con­struc­tion by for new in­ven­tions like skate­boards by adding syl­la­bles to­gether in a way that Ger­manic lan­guages or even An­glo lan­guages usu­ally don’t do. They had es­sen­tially their own gram­mar. But that di­alect seemed to be dy­ing and not used by the young any­more.

• Do we even know that TV kil­led lo­cal ac­cents and di­alects in any coun­try? Be­cause I’m not sure that Slavic lan­guages ever had such rad­i­cal di­alec­tal va­ri­ety as English and Ger­man.

• The Slavs just pro­moted their rad­i­cal di­alects to full-blown lan­guages, mostly due to poli­ti­cal frag­men­ta­tion.

• Anec­do­tal ev­i­dence—dated a woman in Bir­m­ing­ham who was a pro­fes­sional in­ter­preter from English to Pol­ish and she men­tioned she some­times of­ten hired by vis­it­ing Slo­vak busi­ness­peo­ple who can­not find an English to Slo­vak in­ter­preter and it works all right. Not with­out some con­fu­sion, but works. I was sur­prised, since that kind of poli­ti­cial frag­men­ta­tion be­gan a good 1000 years ago.

• Slavic lan­guages re­sem­ble each other enough that peo­ple have tried to con­coct pan-Slavic bridge lan­guages for a long time. It has oc­curred to me that it might just be eas­ier for some­one to learn a rel­a­tively in­ter­op­er­a­ble nat­u­ral Slavic lan­guage (Serbo-Croat or Slo­vak, maybe, judg­ing by this?) and try that as a bridge.

• I think the best re­turn on in­vest­ment in this re­gard is learn­ing Rus­sian. It is widely use­ful even in places like Kaza­khstan or Bul­garia, due to it be­ing taught at schools in the Soviet era. Of course that gen­er­a­tion is ag­ing, but not so much—I’d say over 45 peo­ple were still taught it and this is the age when they get into lead­er­ship po­si­tions and not re­tire for an­other 20-25. Be­sides, there are lots of sci­en­tific pub­li­ca­tions, SA of SSC men­tioned med­i­cal re­search that never got trans­lated and so on. Liv­ing in Cen­tral Europe, the most use­ful lan­guages here look like English > Ger­man > Rus­sian. This is of course widely lo­ca­tion de­pen­dent.

Fun story: a friend of mine was in­vited to a pro­ject to the Sili­con Valley and came back shak­ing his head say­ing “Next time I will learn some Span­ish in ad­vance so that I can have a chat with peo­ple like the news­pa­per guy.”

Less fun story: I re­ally like the sound of the Ital­ian lan­guage, but it does not feel like it worths in­vest­ing into. I would not re­ally want to live there, lovely cul­ture but crazy poli­tics and the “im­por­tant” peo­ple speak English any­way. There is an enor­mous differ­ence in ROI be­tween say learn­ing Span­ish and learn­ing Ital­ian.

• Liv­ing in Cen­tral Europe, the most use­ful lan­guages here look like English > Ger­man > Rus­sian.

How use­ful to you see Ger­man out­side of Ger­many, Aus­tria and Switzer­land?

• With 50+ peo­ple in Czech Repub­lic, Slove­nia, Croa­tia, even Hun­gary, so ba­si­cally ex-Hab­s­burg places, it tends to be use­ful. They picked up some amount of it from their grand­par­ents who re­mem­ber when it was the lin­gua franca of the monar­chy and their English is not very good usu­ally. The typ­i­cal old Czech tourist in Bu­dapest will try to com­mu­ni­cate in more or less bro­ken Ger­man.

Also for young peo­ple, in this re­gion young peo­ple usu­ally learn English but if they have ca­pac­ity left, the sec­ond one is usu­ally Ger­man. And for this rea­son, “east­ern” sub­sidi­aries of DE/​AT/​CH com­pa­nies of­ten keep the in­ter­nal re­port­ing lan­guage in it. Which re­in­forces peo­ple want­ing to learn it.

The most cre­ative us­age I saw was some guys from South Ty­rol who offered bil­in­gual SAP con­sult­ing to DE/​AT/​CH firms hav­ing sub­sidi­aries in Italy.

I also know a lady who went to East Belgium to work and it worked out well for her.

• Note, though, that you ba­si­cally need to achieve na­tive-like lev­els of profi­ciency in or­der to use one Slavic lan­guage to un­der­stand an­other. So you may well never be able to col­lect this re­turn on the in­vest­ment at all. My Rus­sian isn’t any­where near this level and, liv­ing in Cen­tral Europe, the only use I ever make of it is talk­ing to Rus­si­ans. Sig­nage in Slavic-speak­ing coun­tries be­comes some­what com­pre­hen­si­ble, but that’s about it oth­er­wise. Rus­sian does, of course, al­low you to com­mu­ni­cate in Cen­tral Asia, the Cau­ca­sus, and the Baltic coun­tries, which is nice if you want that, but also likely to be ir­rele­vant for most peo­ple.

• True, but also: I smell some sig­nifi­cant busi­ness op­por­tu­ni­ties in the gen­eral di­rec­tion of Cen­tral Asia. It is a very vir­gin mar­ket for e.g. tech prod­ucts and they have nat­u­ral re­sources /​ fos­sil fuels to pay with.

• Slavic lan­guages di­vide into three groups which are fairly differ­ent: Eastern (e.g. Rus­sian, Ukrainian), Western (e.g. Pol­ish, Czech), and South­ern (e.g. Bul­gar­ian, Ser­bian). I sus­pect that a nat­u­ral “bridge” lan­guage might work within a group, but not be­tween groups.

• If you are a tourist in Ukraine, know­ing ba­sics of Rus­sian is use­ful, the kind that the OP refers to. If you want some busi­ness done, just speak English; there would be enough trans­la­tors. If you need to read some­thing writ­ten in Ukrainian, and it is suffi­ciently com­plex and your Rus­sian is not very good, have it trans­lated; and even if your Rus­sian is that good, but not your first lan­guage, it will cost you a lot of nerves.

(Also, if you know Ukrainian more or less well, it can help sig­nifi­cantly to ac­cept Rus­sian, Byelorus­sian, Pol­ish and Czech. I am not say­ing you will au­to­mat­i­cally un­der­stand those lan­guages, just that adapt­ing to them should be eas­ier with Ukrainian as the base.)

• Ha ha, when I first read that, I thought “fur­riner” was an­other nick­name for Fur­ries and I was very, very con­fused.