Still try­ing to put my fin­ger on some­thing both­er­ing me. Very sub­jec­tive, no doubt ill-​​considered… but still there and not quite stated clearly enough.

This is some­thing about busi­ness, project man­age­ment, plan­ning and imple­men­ta­tion: About how a cer­tain class of man­ager views the spec­i­fi­ca­tion of goals, the sense that goals met cre­ate busi­ness value, and how those peo­ple deal with the real peo­ple whose work it is to con­nect the two ide­al­iza­tions (goal, value) to one another by apply­ing their expe­ri­ence, insight, and abil­ity to com­mu­ni­cate.

In my expe­ri­ence, soft­ware devel­op­ers are appro­pri­ate to the task; “com­puter pro­gram­mers” can­not as a rule reli­ably deliver value from their work.

This is some­thing about ped­a­gogy, grad­u­ate train­ing, the Acad­emy and spe­cial­iza­tion: About how grant appli­ca­tions are writ­ten years before monies are acquired; how “real” aca­d­e­mic projects are spelled out in grant appli­ca­tions as if fore­sight were per­fect and explo­ration was ratio­nal, while the work is done by sub­sti­tutable and inex­pe­ri­enced stu­dents and young fac­ulty; how “home­work” projects and eval­u­a­tions are treated as if indi­vid­ual peo­ple can learn in a vac­uum of read­ing and self-​​direction and wordy lec­ture, as if text­books were help­ful with­out con­ver­sa­tion; as if the cost, util­ity, qual­ity and dura­tion of schol­ar­ship were all per­fectly fun­gi­ble with one another, per­fectly liq­uid… sub­ject to insignif­i­cant exchange costs not wor­thy of note.

In my expe­ri­ence, stu­dents learn when they work col­lec­tively on a shared goal, sup­port­ing one another, and in the process learn by dis­cov­er­ing and shar­ing their nonover­lap­ping skills: when they “cheat”. “Stars” who can­not explain their work, who can­not col­lab­o­rate, who dis­dain “cheat­ing” (by the stan­dards of most mod­ern Honor Pledges and tenure review com­mit­tees) by sit­ting qui­etly by them­selves and doing what their mas­sive insight has revealed is the path to what you (mere peo­ple) need… these folks can­not as a rule reli­ably deliver value from their work.

This is some­thing about the the­ory and prac­tice of arti­fi­cial intel­li­gence, oper­a­tions research, machine learn­ing, and meta­heuris­tics: About the unwill­ing­ness or inabil­ity to treat tech­niques pre­scrip­tively except as a form of self-​​promotion of one’s own research or per­sonal bias; about the strangely per­sis­tent short­fall in com­mu­ni­cat­ing the util­ity of those thou­sand vari­ant meth­ods from lin­ear pro­gram­ming to fic­ti­tious play to genetic pro­gram­ming or graph­i­cal model learn­ing, any one of which might poten­tially answer ques­tions, iden­tify pat­terns, and help peo­ple invent soft­ware or phys­i­cal engi­neer­ing designs; about a cul­ture of “prac­ti­tion­ers” who can­not be both­ered to learn enough the­ory to explain why their approach is suf­fi­cient for their par­tic­u­lar tasks, and a sep­a­rate cul­ture of “the­o­rists” who can­not be both­ered to learn enough of best prac­tice to explain why their approach is nec­es­sary for any task.

In my expe­ri­ence, the aver­age time an algo­rithm is expected to run may be of inter­est, but as far as my par­tic­u­lar prob­lem is con­cerned it has no bear­ing until I have run it for a while to see some results, see how it’s going, suss out what “kind” of prob­lem this spe­cific instance is—to see what value comes from “how long” it will take to run, as opposed to see­ing any answer at all. I do work, I cre­ate stuff, to bet­ter under­stand the path from ide­al­ized goal to real­ized value. Things like speed, accu­racy, ease of use and under­stand­abil­ity, these are things I try to mea­sure, not assume before­hand for some com­bi­na­tion of prob­lem and approach, and I want infor­ma­tion with which to update my assess­ments as quickly and accu­rately as pos­si­ble. Because for some strange rea­son I am unable to tell before­hand how dif­fi­cult an inter­est­ing instance of a prob­lem will be, even with the most famil­iar approach.

I have a great deal of both prac­ti­cal expe­ri­ence and the­o­ret­i­cal back­ing in these mat­ters, and all that has hap­pened for me (your mileage may vary) is that I am more uncer­tain about my prej­u­dices, and yours, all the time.

On aver­age, doing some­thing small, imme­di­ately, is bet­ter than talk­ing a long time about the many things you could do, about poten­tial­i­ties and lim­its and aver­age behav­ior. And per­haps bet­ter than doing “just any­thing” is con­sid­er­ing the small set of sim­ple incre­men­tal improve­ments, select­ing the one that seems it will pro­vide the most value for that scale of effort, and try­ing it.

In too many domains we con­flate ratio­nal­ity with rigor, and treat the straight­est path between them as a recipe for suc­cess. But isn’t “ratio­nal­ity” an inten­tion­ally bounded thought process? a strat­egy of fully dis­miss­ing alter­na­tives as greed­ily and thor­oughly as possible?

But I don’t want to spend my time with rig­or­ous peo­ple. They’re fuck­ing annoy­ing, when you get right down to it. When I’m actu­ally try­ing to solve a prob­lem, I would pre­fer to col­lab­o­rate with ten expe­ri­enced peo­ple (some “the­o­rists”, some “prac­ti­tion­ers”) who can speak quickly, approx­i­mately, and explore oh so many alter­na­tives. I want peo­ple who can use sim­ple, stu­pid, non-​​optimal tools all of us poor fools can under­stand… but who in using those tools dis­cover many paths by which we might col­lec­tively trace our way—any god­damned way as long as we arrive—from our imme­di­ate goal to our desired value.

Because value trumps method.

And value (as I’ve said) is some­thing that may not be ratio­nally pre­dictable. Value comes along the way, it emerges. Value in so many cases is con­tin­gent on mul­ti­ple scales of expe­ri­ence, long and short term, on con­stantly revised and dis­carded mod­els, on alter­na­tive hypothe­ses eas­ily exchanged. Achiev­ing value depends on my tools, my incli­na­tion, my habit. On what I’ve done so far.

And all these change from per­son to per­son, from prob­lem to prob­lem. From moment to moment. In my expe­ri­ence, on a shorter scale than any—any—problem-​​solving method, whether it’s a busi­ness project, a the­sis or grant, a sin­gle “sim­ple” appli­ca­tion of heuris­tic to instance.

Some­thing deep is miss­ing out there.