Rut­ter ar­gues for this point in ⁠Web Ty­pog­ra­phy, which is con­sis­tent with my own A/B tests⁠ where even lousy changes are dif­fi­cult to dis­tin­guish from zero ef­fect de­spite large n, and with the gen­eral sham­bolic state of the In­ter­net (eg. as re­viewed in the 2019 Web Al­manac⁠). If read­ers ⁠will not in­stall ad­block⁠ and load­ing times of mul­ti­ple sec­onds have rel­a­tively mod­est traf­fic re­duc­tions, things like align­ing columns prop­erly or using sec­tion signs or side­notes must have ef­fects on be­hav­ior so close to zero as to be un­ob­serv­able.

Para­phrased from Di­a­logues of the Zen Mas­ters as quoted in pg11 of the Ed­i­tor’s In­tro­duc­tion to Three Pil­lars of Zen⁠:

One day a man of the peo­ple said to Mas­ter Ikkyu: “Mas­ter, will you please write for me max­ims of the high­est wis­dom?” Ikkyu im­me­di­ately brushed out the word ‘At­ten­tion’. “Is that all? Will you not write some more?” Ikkyu then brushed out twice: ‘At­ten­tion. At­ten­tion.’ The man re­marked ir­ri­ta­bly that there wasn’t much depth or sub­tlety to that. Then Ikkyu wrote the same word 3 times run­ning: ‘At­ten­tion. At­ten­tion. At­ten­tion.’ Half-angered, the man de­manded: “What does ‘At­ten­tion’ mean any­way?” And Ikkyu an­swered gen­tly: “At­ten­tion means at­ten­tion.”

And also, ad­mit­tedly, for es­thetic value. One earns the right to add ‘ex­tra­ne­ous’ de­tails by first putting in the hard work of re­mov­ing the ac­tual ex­tra­ne­ous de­tails; only after the ground has been cleared—the ‘data-ink ratio’ max­i­mized, the ‘chartjunk’ re­moved—can one see what is ac­tu­ally beau­ti­ful to add.

Good de­sign may be “as lit­tle de­sign as pos­si­ble” which gets the job done, to para­phrase Di­eter Rams⁠; the prob­lem comes when de­sign­ers focus on the first part, and for­get the sec­ond part. If a min­i­mal­ist de­sign can­not han­dle more con­tent than a few para­graphs of text & a generic ‘hero image’, then it has not solved the de­sign prob­lem, and is merely a sub-genre of il­lus­tra­tion. (Like pho­tographs of el­e­gant min­i­mal­ist Scan­di­na­vian or Japan­ese ar­chi­tec­ture which leave one won­der­ing whether any human could live in­side them, and how those build­ings would learn⁠.) And if a min­i­mal­ist web­site can­not even present some text well, you can be sure they have not solved any of the hard prob­lems of web de­sign like link rot or cross-referencing!

The de­fault pre­sen­ta­tion of sep­a­rate pages means that an en­tire page may con­tain only a sin­gle para­graph or sen­tence. The HTML ver­sions of many tech­ni­cal man­u­als (typ­i­cally com­piled from LaTeX, Doc­Book, or GNU Info) are even worse, be­cause they fail to ⁠ex­ploit prefetch­ing⁠ & are slower than local doc­u­men­ta­tion, and take away all of the use­ful key­bind­ings which makes nav­i­gat­ing info man­u­als fast & con­ve­nient. Read­ing such doc­u­men­ta­tion in a web browser is Chi­nese water tor­ture. (That, decades later, the GNU project keeps gen­er­at­ing doc­u­men­ta­tion in that for­mat, rather than at least as large single-page man­u­als with hy­per­linked table-of-contents, is a good ex­am­ple of how bad they are at UI/UX de­sign.) And it’s not clear that it’s that much worse than the other ex­treme, the mono­lithic man page⁠ which in­cludes every de­tail under the sun and is im­pos­si­ble to nav­i­gate with­out one’s eyes glaz­ing over even using in­cre­men­tal search⁠ to nav­i­gate through dozens of ir­rel­e­vant hits—every sin­gle time!

Also known as “back­ward or re­verse ci­ta­tions⁠”, “what links here”⁠, & “back­links”.

This fixes the biggest prob­lem with the Me­di­aWiki⁠ wiki sys­tem’s ‘what links here’ im­ple­men­ta­tion of back­links—which is the sim­plis­tic way of im­ple­ment­ing it so has be­come the stan­dard wiki soft­ware ap­proach to dis­play­ing back­links.

The WhatLinksHere page (⁠eg. En WP⁠) will tell you that sev­eral hun­dred other Wikipedia ar­ti­cles link to your cur­rent Wikipedia ar­ti­cle, yes, but you have no idea what the con­text is (on ei­ther page!), and if it is an im­por­tant link or a minor link, or even where in the ar­ti­cle it might be—it might be hid­den under some un­pre­dictable dis­played text, and you have to search the Me­di­aWiki markup it­self just to find it!

This is only par­tially fixed by tools like Lupin’s Tool⁠ which try to lo­cate the link by load­ing the other page, be­cause those are used by few ed­i­tors, and still re­quire ef­fort. Be­cause Me­di­aWiki ren­ders every­thing server-side, there is no rea­son it could not do some­thing sim­i­lar and dis­play con­tex­tu­al­iz­ing ex­cerpts next to each link. It just doesn’t. (It doesn’t need true bidi­rec­tional links—even a heuris­tic hack of as­sum­ing the first link in each ar­ti­cle is the ‘real’ link, and ig­nor­ing du­pli­cates, would be a major im­prove­ment.)

Roam ap­par­ently might do some­thing like our ‘in­lin­ing’, but I know too lit­tle about it to say. Mag­gie Ap­ple­ton mocks up such a ⁠“spec­u­la­tive in­ter­face”⁠, but ap­pears to not know of any im­ple­men­ta­tions.

A lim­ited ex­am­ple is ⁠Greater­Wrong⁠, which does back­links on ⁠posts⁠ and ⁠in­di­vid­ual com­ments⁠. How­ever, while back­links on in­di­vid­ual com­ments are rea­son­ably atomic, they do not show the call­ing con­text, and the pop­ups on the links only show the stan­dard whole-item view. (GW’s back­links were in­tro­duced in 2019 at the re­quest of Wei Dai, well be­fore Gwern.net’s back­links were in­tro­duced in 2021 to take ad­van­tage of the new tran­sclu­sion fea­ture, and they are mostly in­de­pen­dent in de­sign.)

The orig­i­nal rea­son I began au­to­mat­i­cally gen­er­at­ing IDs on all Gwern.net hy­per­links was minor: I wanted to use the within-page pop­ups (like the lit­tle up/down-arrows) to re­move re­dun­dant links & show more con­text.

A re­search paper might be dis­cussed at length in one sec­tion, but then cited else­where; it would be bad to not hy­per­link it, so usu­ally, I would make a re­dun­dant hy­per­link. How­ever, if the first dis­cus­sion had a unique ID, then I could sim­ply link later ref­er­ences to the ID in­stead, and the reader could hover over it to pop up that dis­cus­sion, read it, and then click through. (So it would look like this in Mark­down: [Foo 2020](URL){#foo-2020} proved ABC, which is interesting because of DEF … [thousands of words & many sections later] … see also [Foo 2020](#foo-2020).)

One could do this man­u­ally on a case-by-case basis, but there are so many links, and the ID can be in­ferred from the meta­data, so why not gen­er­ate them au­to­mat­i­cally, so one could al­ways be sure that #foo-2020 was valid?

And once most links had IDs were unique within pages, that meant they could be unique across pages as well… So the pop­ups led to the bidi­rec­tional back­links.

Mul­ti­ple links to an­other URL is not un­usual on Gwern.net, par­tic­u­larly when mak­ing use of IDs for pre­cise links so one might eas­ily link not just /foo but /foo#bar, /foo#quux, and maybe even /foo#baz, why not?

In fact, this can be a good way to han­dle com­plex an­no­ta­tions: you can break them up into mul­ti­ple an­no­ta­tions & link each ver­sion. For ex­am­ple, imag­ine a com­plex, in-depth ma­chine learn­ing paper like the Big­GAN paper⁠, where the ab­stract is im­por­tant but omits some key parts on page 6⁠ and also page 8⁠ I want to high­light for other pur­poses.

I could set­tle for not an­no­tat­ing them at all; or I could try to jam them all into just one an­no­ta­tion; or I could link to the exact pages in the paper PDF using the #page=n⁠ trick & set­tle for the PDF pop­ping up with no an­no­ta­tion pos­si­ble (this also works if you cre­ate ar­bi­trary IDs solely for the pur­pose of writ­ing mul­ti­ple dis­tinct an­no­ta­tions); or I could cre­ate an­no­ta­tions for the exact page links & sim­ply cross-reference them! The back­links en­able cross-referencing at a glance, and nav­i­gat­ing at a hover. And since this is all fully re­cur­sive, an­no­ta­tions are first-class cit­i­zens, the tar­gets can be ar­bi­trary IDs of ar­bi­trary URLs or <div>/<span>s, back­links & links in­ter­op­er­ate etc, it all Just Works™ seam­lessly on the part of both au­thor & reader.

But a sys­tem which threw away the meta­data of an­chors & IDs would strug­gle to do any of this: the 1:1 links or the dis­tinct URL-anchors would col­lapse down to hopelessly-ambiguous many:many maps.

I use au­thor names for my IDs, be­cause that meta­data is usu­ally avail­able due to an­no­ta­tions and is eas­ily guessed & writ­ten. But other im­ple­men­ta­tions might pre­fer to in­stead gen­er­ate consistent-but-unique IDs by sim­ply strip­ping or es­cap­ing the URL in ques­tion (eg. into Base64⁠ or URL-encoding), or by feed­ing it into a web-browser-supported hash func­tion like SHA-256⁠ (trun­cated to 8 chars—there are not nearly enough URLs on any page to worry about col­li­sions).

Show­ing the con­text for the back­link re­quires down­load­ing ei­ther the an­no­ta­tion or page, to nar­row down to the ID’s con­text. Show­ing back­link con­text can use up a lot of space, and ren­der­ing all that HTML is ex­pen­sive, par­tic­u­larly for back­link sec­tions which have scores of back­links.

So the col­laps­ing serves as lazy eval­u­a­tion⁠, and avoids doing that un­less the reader re­quests it. (Since back­links are all known at compile-time, it would be pos­si­ble to pre­com­pute the con­text, but not too easy.)

This re­places an ear­lier ⁠Hakyll-based tag sys­tem⁠. The Hakyll ap­proach was quite sim­ple and in­tended just for small blogs, and had no way to han­dle tag­ging local files, much less ar­bi­trary URLs. (The tag code was also black magic I couldn’t mod­ify.) Mean­while, the evolv­ing filesys­tem hi­er­ar­chy for my local files al­ready looked like a tag sys­tem, and the evo­lu­tion was easy.

I abuse an­chors in this way to track ‘af­fil­i­a­tion’ of URLs, both for eas­ier ref­er­ence/search and for set­ting link-icons. For ex­am­ple, I re­gard Deep­Mind au­thor­ship of a paper as being a help­ful thing to know, and so I ap­pend #deepmind to any DeepMind-related URL (eg. https://arxiv.org/abs/1704.03073#deepmind). I find it par­tic­u­larly help­ful for track­ing Chi­nese AI re­search, where they have a habit of qui­etly drop­ping re­veal­ing pa­pers on Arxiv with no PR or West­ern at­ten­tion.

When I link that URL, the link will get the Deep­Mind logo as its link-icon, and it is eas­ier for me to search if I can re­mem­ber that it was DM-related. This will not break the link, be­cause the an­chor is client-side only (un­like if you wanted to abuse query-parameters this way—many servers would ig­nore a mal­formed url like foo?deepmind, but many oth­ers would throw an error); thus, I can copy-paste back and forth be­tween Gwern.net and Red­dit or Twit­ter, and the lat­ter will con­tinue to work nor­mally (they track the full URL but usu­ally drop the an­chor for pur­poses like search­ing). Be­cause it’s over­load­ing the an­chor, I can de­fine new af­fil­i­a­tions any time I please, and am up to 51 af­fil­i­a­tions as of 2023-04-19; and I could en­code any­thing else I might want to en­code by using a new con­ven­tion. I could en­code the date by writ­ing #2023-04-19, or the au­thor, #john-smith, or lit­tle notes like #todo. As long as they do not hap­pen to be a real an­chor, they will work. (This is why a num­ber of past web de­sign hacks like `#!’⁠ URLs⁠ or “Text Frag­ments”⁠ (crude content-addressable URLs⁠) have also ex­ploited an­chors, for their backwards-compatibility.)

This hack does come with costs. First, it cre­ates spu­ri­ous an­chors which my linkcheck­ers will warn about but must be ig­nored as de­lib­er­ate er­rors. Sec­ond and more se­ri­ously, while it works fine on ex­ter­nal URLs, it be­gins to cause prob­lems with local URLs: con­sider a case like /doc/reinforcement-learning/model-free/2016-graves.pdf#deepmind—this URL is not it­self a prob­lem for an­no­ta­tions, which does every­thing by URL, but it is a prob­lem for any­thing at the file level, which sees only .../2016-graves.pdf. There is no #deepmind at the file level! This re­quires flaky hacks like look­ing up every an­no­ta­tion with the file as a pre­fix to see if there is an an­no­ta­tion with some an­chor. I in­tend to re­move this hack in favor of stor­ing af­fil­i­a­tions in­side the an­no­ta­tion meta­data; how­ever, I may keep it as a con­ve­nient way to input af­fil­i­a­tions.

The tag pop­ups used to sim­ply tran­sclude/load the tag page into the popup. This turned out to be un­pre­dictable & slow for the mega-tags like psychology⁠ which had hun­dreds to thou­sands of en­tries (and are in des­per­ate need of refac­tor­ing), which were also not tran­sclu­sions at the time, and so could take >10s to load.

Amus­ingly, while use­less for their os­ten­si­ble in­tended pur­pose, In­sta­gram tags were use­ful for early land­mark re­sults in deep learn­ing scal­ing (par­tic­u­larly Ma­ha­jan et al 2018⁠), help­ing to es­tab­lish that neural nets could learn from bil­lions of im­ages when the ex­pert con­ven­tional wis­dom was that mil­lions was all that was use­ful & neural nets were ‘fun­da­men­tally un­able to scale’.

Em­bed­dings which are en­hanced by in­clud­ing their in­ter­nal links, back­links, manually-curated similar-links (the ‘see-also’ sec­tions of an­no­ta­tions), and their prior tags, it’s worth not­ing: all these should en­hance the tag­ga­bil­ity & clus­ter­abil­ity of the em­bed­dings.

You can use LLMs di­rectly for tag­ging, with tricks like fine­tun­ing or in­clud­ing a list of valid tags to choose from, but these will prob­a­bly be less ac­cu­rate than a clas­si­fier, will tend to be slower & less suit­able for ⁠real-time active-learning of tags⁠, and em­bed­dings are reusable for other pur­poses.

An al­ter­na­tive would be k-medoids⁠, which would con­struct clus­ters whose ‘cen­ter’ is a spe­cific dat­a­point (with k-means, there is not nec­es­sar­ily a dat­a­point at the cen­ter of a clus­ter), mak­ing in­ter­pretabil­ity eas­ier for the user and pos­si­bly cre­at­ing higher-quality clus­ters. We wouldn’t want to use DB­SCAN⁠ be­cause it would ig­nore many points as ‘out­liers’; this is rea­son­able with real-world data where dat­a­points may be un­pre­dictable or out­right garbage, but in tag­ging, we can as­sume that all data is valid, and so we want to keep ‘out­liers’ and con­sider as­sign­ing them their own tag—maybe they are sim­ply im­ma­ture.

For ex­am­ple, my last few ses­sions’ au­to­matic ti­tles/sum­maries: “In­su­lated con­tainer tem­per­a­ture”, “Dan­ger­ous meta­data dates”, “Haskell Re­cur­sive File List­ing”, “Plas­tic Pipes Re­duce Build-up”, “Tem­plate for un­sup­ported ci­ta­tions”, “Buffer rewrite con­di­tion mod­i­fi­ca­tion”, “Git pull with au­to­matic merge”, “Grep for non-printable char­ac­ters”, “Al­ter­na­tive to (keyboard-quit)”, “Re­verse Mark­down List Order”, “Goal Set­ting The­ory & Con­sci­en­tious­ness”, “Odin’s Fa­vorite Spice⁠?”, “Cat loung­ing against chair”, ⁠“De­crypt Poem Mes­sage”⁠, “La­dy­bug Fork­lift Cer­ti­fi­ca­tion”.

Why not do this for every­thing, like file­names, and end the ‘blink­ing 12’ prob­lem⁠ of Untitled (89).doc? After all, the prob­lem with these is due less to the in­trin­sic com­plex­ity or dif­fi­culty—it takes all of a minute to read the man­ual or fig­ure it out or de­cide what to call a file—so much as the irk­some­ness of tak­ing a minute to do so every time one would have to–every power out­age, Day­light Sav­ings, VCR up­grade, ran­dom doc­u­ment—for what is ul­ti­mately a tiny ben­e­fit per-instance. (A real cler­i­cal ben­e­fit, which adds up in ag­gre­gate, but still, small in­di­vid­u­ally, and thus eas­ily out­weighed by the has­sle.) But a LLM can do it eas­ily and won’t com­plain.

One idea I have not seen much, but which would be use­ful as au­toma­tion is added, is a con­cept of ‘neg­a­tive tags’ or ‘anti-tags’: as­sert­ing that an item is def­i­nitely not a tag.

Tags are typ­i­cally pre­sented as a two-valued bi­nary vari­able, but be­cause the de­fault for tag sys­tems is typ­i­cally to be un­tagged, and be­cause most tag sys­tems are in­com­plete, that means that er­rors are highly skewed to­wards er­rors of omis­sion rather than com­mis­sion. An item with tag x but not tag y, is al­most al­ways in­deed an in­stance of x; how­ever, it will often be y too. So the ab­sence of a tag is much less in­for­ma­tive than the pres­ence of a tag. But there is no way to dis­tin­guish be­tween “this item is not tagged y be­cause no one has got­ten around to it” and “be­cause some­one looked closely and it’s def­i­nitely not y”.

In reg­u­lar tag use, this merely re­sults in some wasted ef­fort as users pe­ri­od­i­cally look at the item and double-check that it’s not y. With au­toma­tion, this can be a se­ri­ous ob­sta­cle to things like ac­tive learn­ing work­ing at all: if we have no way of mark­ing ‘it’s def­i­nitely not y’, then when we at­tempt to find in­stances of y which are not la­beled y, we will every time have to ig­nore the same false pos­i­tives. (And we also can’t train our clas­si­fier to ig­nore those false pos­i­tives, even though those would be the most valu­able to train on be­cause they are the ones which most fooled our clas­si­fier.)

This could, in fact, be how we do the clus­ter­ing refac­tor­ing to begin with: we sim­ply clus­ter by de­fault (ar­bi­trar­ily sort­ing across-clusters, and then semantic-sort within-cluster), and the user can hit a lit­tle but­ton to ‘can­on­ize’ a clus­ter as a new tag.

Imag­ine you are sort­ing a list of items A–Z ([ABCDE­FGHIJKMLNOPQRSTU­VWXYZ]), where you have pair­wise dis­tances like ‘A is closer to B than C or Z’. If you pick, say, ‘H’, and then sim­ply ‘sort by dis­tance’ to ‘H’ to form a 1D list, the re­sult would merely ‘ping-pong’ back and forth: [HIGJFKEM­DOCPB…]. (The more equidis­tant clus­ters there are, the worse this ping-pong ef­fect is.) This would be mean­ing­ful in that ‘P’ re­ally is slightly closer to ‘H’ than ‘B’, but the back-and-forth would look con­fus­ing to any reader, who wouldn’t see the un­der­ly­ing A–Z. How­ever, if you sorted pair­wise greed­ily, you would get a list like [HIJKMLN…YZ­ABCDEFG], and aside from the ‘jump’ at A/Z, this would be mean­ing­ful to the reader and much more pleas­ant to browse. It would also be eas­ier to cu­rate, as you can see the se­quence and also the ‘jump’, and, say, de­cide to edit A–G into its own tag, and then fur­ther re­fine what’s left into H–L & M–Z tags.

If you had a cor­pus of restau­rant re­views, you might want to clus­ter them by the type of food, but your de­fault em­bed­ding keeps group­ing them by ge­og­ra­phy when re­duced to a 2D GUI; no prob­lem, just sub­tract the tags for all of the cities like “New York City”, and then what’s left af­ter­wards will prob­a­bly be clump­ing by French vs Asian fu­sion vs Chi­nese etc and you can eas­ily lasso⁠ each clus­ter and name them and cre­ate new tags with hardly any toil.

This is ab­stract enough that I doubt it could be eas­ily ex­plained to a user, but I think it could be en­coded into the UI in a us­able way. Like one could present a 2D plot, where each dat­a­point and averaged-tag is pre­sented, and they can be clicked on to ‘strengthen’ or ‘weaken’ them (where each de­gree of strength cor­re­sponds to an em­bed­ding arith­metic op­er­a­tion, but weighted, like 10% per click). If you are in­ter­ested in the X-ness of all your dat­a­points, you sim­ply click on the X dot to ‘strengthen’ it a few times until the up­dated 2D plot makes sense.

An un­usual choice, as one does not as­so­ciate IBM with font de­sign ex­cel­lence, but nev­er­the­less, it was our choice after blind com­par­i­son of ~20 code fonts with vari­ant ze­roes⁠ (which we con­sider a re­quire­ment for code). An ap­peal­ing newer al­ter­na­tive is Jet­Brains Mono (which doesn’t work as well with Gwern.net’s style, but may suit other web­sites).

Side­notes have long been used as a ty­po­graphic so­lu­tion to densely-annotated texts such as the Geneva Bible⁠ (first 2 pages⁠), but have not shown up much on­line yet.

An early & in­spir­ing use of mar­gin/side notes.

IBM Plex Mono was cho­sen in part via using the Cod­ing­Font ‘tour­na­ment’; Adobe Source Code Pro⁠ also ranked high, and we used it ini­tially, but Plex Mono edged it out, with its use­ful al­ter­na­tives and a some­what bet­ter over­all ap­pear­ance. (Who knew IBM could com­mis­sion such a nice mono­space font?)

Or at least, so we think? Google Page­Speed keeps claim­ing that mini­fi­ca­tion would cut as much as half a sec­ond off total time.

Per­haps the re­turns to de­sign are also going up with time as In­ter­net de­sign­ers in­creas­ingly get all the rope they need to hang them­selves? What browser devs & Moore’s Law giveth, semi-malicious web de­sign­ers take away⁠. Every year, the range of worst to best web­site gets broader, as ever new ways to de­grade the brows­ing ex­pe­ri­ence—not 1 but 100 track­ers! newslet­ter pop­ups! sup­port chat! Taboola chum­box! ‘browser no­ti­fi­ca­tions re­quested’! 50MB of hero im­ages! lay­out shifts right as you click on some­thing!—are in­vented. 80-column ASCII text files on BBSes offer lit­tle de­sign great­ness, but they are also hard to screw up. To make an out­stand­ingly bad web­site re­quires the lat­est CMSes, A/B test­ing in­fra­struc­ture to ⁠Schlitz your way⁠ to prof­itabil­ity, CDNs, ad net­work auc­tion­ing tech­nol­ogy, and high-paid web de­sign­ers using only Apple lap­tops. (A ⁠2021 satire; note that you need to dis­able ad­block.) Given the sub­tlety of this creep to­wards degra­da­tion & short-term prof­its and the rel­a­tively weak cor­re­la­tion with fit­ness/prof­itabil­ity, we can’t ex­pect any rapid evo­lu­tion⁠ to­wards bet­ter de­sign, un­for­tu­nately, but there is an op­por­tu­nity for those busi­nesses with taste.

Which might ac­count for why im­prove­ments in Gwern.net de­sign also seem to cor­re­late with more com­ments where the com­menter ap­pears in­fu­ri­ated by the de­sign—that’s cheat­ing!

One anec­dote Steve Jobs⁠ cited for his per­fec­tion­ism, even in things the user would os­ten­si­bly not see, is his fa­ther’s hob­by­ist car­pen­try, where he cared about mak­ing even the backs of fences & cab­i­nets look good; from Wal­ter Isaac­son’s2011 Steve Jobs:

Jobs re­mem­bered being im­pressed by his fa­ther’s focus on crafts­man­ship. “I thought my dad’s sense of de­sign was pretty good”, he said, “be­cause he knew how to build any­thing. If we needed a cab­i­net, he would build it. When he built our fence, he gave me a ham­mer so I could work with him.”

50 years later the fence still sur­rounds the back and side yards of the house in Moun­tain View. As Jobs showed it off to me, he ca­ressed the stock­ade pan­els and re­called a les­son that his fa­ther im­planted deeply in him. It was im­por­tant, his fa­ther said, to craft the backs of cab­i­nets and fences prop­erly, even though they were hid­den. “He loved doing things right. He even cared about the look of the parts you couldn’t see.”

His fa­ther con­tin­ued to re­fur­bish and re­sell used cars, and he fes­tooned the garage with pic­tures of his fa­vorites. He would point out the de­tail­ing of the de­sign to his son: the lines, the vents, the chrome, the trim of the seats. After work each day, he would change into his dun­ga­rees and re­treat to the garage, often with Steve tag­ging along. “I fig­ured I could get him nailed down with a lit­tle me­chan­i­cal abil­ity, but he re­ally wasn’t in­ter­ested in get­ting his hands dirty”, Paul later re­called. “He never re­ally cared too much about me­chan­i­cal things. I wasn’t that into fix­ing cars”, Jobs ad­mit­ted. “But I was eager to hang out with my dad.”

Cyn­i­cal com­men­ta­tors point out that many great pro­fes­sional fur­ni­ture mak­ers did not put much work into the back of cab­i­nets, as it was a waste; I would point out that since it is rel­a­tively easy to judge fur­ni­ture com­pared to soft­ware, their crit­i­cism ac­tu­ally re­veals why this at­ti­tude could be bril­liant mar­ket­ing for soft­ware.

Whose soft­ware would you trust more (and pay a pre­mium for): the guy who clearly slacks off in the few places you can spot slack­ing, but swears he’s just ef­fi­ciently fo­cus­ing on the im­por­tant parts, “trust me!”—or the guy who is so neu­rot­i­cally per­fec­tion­ist that he costs his com­pany mil­lions of dol­lars at the last minute ‘fix­ing’ some minor ug­li­ness in­side the case where you might never have looked? (Or con­sider sig­na­ture Jobs moves like hav­ing ⁠“sig­na­tures” in­side the Mac­in­tosh case, where “no one” would see it; but of course, plenty of peo­ple would see it, as even Mac­in­toshes would be opened up rou­tinely, and those who saw it would tell oth­ers, and those would tell oth­ers, and the sig­na­tures be­came fa­mous enough that I am link­ing the story many decades later.)