this is what actual nuance about AI looks like - Dorian Taylor @doriantaylor on how LLM coding is vaguely possible, but of limited usefulness for real work
-
@craigduncan @davidgerard @doriantaylor It's also possible to complete preparing dinner using only an assortment of power tools from the garage, or to perform open heart surgery using only tools found in a kindergarten classroom.
@prietschka @davidgerard @doriantaylor
Seriously, the article by Dorian Taylor is a really well written summary of the state of things, and some needed perspective:
"In other words, we shouldn’t be excusing this technology for getting things wrong, but impressed that it ever gets anything right."
The point about ELIZA is spot on. If you happen to have written such a program in primary school it's harder to invest too much in the LLM output.
-
@prietschka @davidgerard @doriantaylor
Seriously, the article by Dorian Taylor is a really well written summary of the state of things, and some needed perspective:
"In other words, we shouldn’t be excusing this technology for getting things wrong, but impressed that it ever gets anything right."
The point about ELIZA is spot on. If you happen to have written such a program in primary school it's harder to invest too much in the LLM output.
@craigduncan @davidgerard @doriantaylor I'm just being silly, pay no attention to me.
-
@craigduncan @davidgerard @doriantaylor I'm just being silly, pay no attention to me.
True for all of us. Thanks for being silly.
-
this is what actual nuance about AI looks like - Dorian Taylor @doriantaylor on how LLM coding is vaguely possible, but of limited usefulness for real work
Slop-Machine Future
The arc of large language models is mediocre, and it bends toward “target procurement”.
(buttondown.com)
@davidgerard @doriantaylor even these nuanced takes get a lot wrong and assign too much usefulness where none is deserved
-
@davidgerard @doriantaylor even these nuanced takes get a lot wrong and assign too much usefulness where none is deserved
@cap_ybarra @davidgerard cool, what do they get wrong
-
@cap_ybarra @davidgerard cool, what do they get wrong
"it's good for writing unit tests"
is it really though? how are you sure the tests fail in the circumstances you need them to? are all unit tests equally valuable?
"i didn't have to read the documentation for x"
so you understand all the ways it can fail, and the circumstances under which you expect it to succeed? you're sure it didn't miss some critical detail?
the answer to all of these questions, btw, comes from grokking the code, which will mean you need to read the code and toy with it enough to sufficiently inhabit it, in which case how much time did you save, really?
-
@cap_ybarra @davidgerard cool, what do they get wrong
@doriantaylor @davidgerard these takes reveal an immature level of engineering understanding, and while this article isn't their source, i do get mega fatigued of hearing them.
again, i'm just some asshole who has done this forever and no one of any consquence, my take may be ignored, i'm used to it, folks gotta get that bag
-
"it's good for writing unit tests"
is it really though? how are you sure the tests fail in the circumstances you need them to? are all unit tests equally valuable?
"i didn't have to read the documentation for x"
so you understand all the ways it can fail, and the circumstances under which you expect it to succeed? you're sure it didn't miss some critical detail?
the answer to all of these questions, btw, comes from grokking the code, which will mean you need to read the code and toy with it enough to sufficiently inhabit it, in which case how much time did you save, really?
1) at no point did i say it was *good* for writing unit tests, i just said it was *possible* to generate them; whether they're any good is a separate consideration
like these things are known to make bad tests and even alter tests to pass; my point was you can't get away with skipping test coverage whether you write the tests by hand or it generates them, because of the way it works ("works")
but as we both pointed out, no guarantee generating tests will save any time
-
1) at no point did i say it was *good* for writing unit tests, i just said it was *possible* to generate them; whether they're any good is a separate consideration
like these things are known to make bad tests and even alter tests to pass; my point was you can't get away with skipping test coverage whether you write the tests by hand or it generates them, because of the way it works ("works")
but as we both pointed out, no guarantee generating tests will save any time
@doriantaylor @davidgerard you're right, my read was insufficiently close. i'm a jerk
-
1) at no point did i say it was *good* for writing unit tests, i just said it was *possible* to generate them; whether they're any good is a separate consideration
like these things are known to make bad tests and even alter tests to pass; my point was you can't get away with skipping test coverage whether you write the tests by hand or it generates them, because of the way it works ("works")
but as we both pointed out, no guarantee generating tests will save any time
2) web api client boilerplate is a consummate pain in the ass, because every vendor does theirs ever so slightly differently. it's also likely one of the things that's super well-represented in the training data, plus it'll either work when you run it or it won't; it's pretty low-risk. the failure mode is i have to correct it by hand (and it isn't like i didn't have the reference docs open contemporaneously anyway).
-
2) web api client boilerplate is a consummate pain in the ass, because every vendor does theirs ever so slightly differently. it's also likely one of the things that's super well-represented in the training data, plus it'll either work when you run it or it won't; it's pretty low-risk. the failure mode is i have to correct it by hand (and it isn't like i didn't have the reference docs open contemporaneously anyway).
@doriantaylor @davidgerard but, and bear with me here, what if we decided to not accept boilerplate as an inevitable thing that must be prompted through and instead wrote some code to ease the burden
-
2) web api client boilerplate is a consummate pain in the ass, because every vendor does theirs ever so slightly differently. it's also likely one of the things that's super well-represented in the training data, plus it'll either work when you run it or it won't; it's pretty low-risk. the failure mode is i have to correct it by hand (and it isn't like i didn't have the reference docs open contemporaneously anyway).
@doriantaylor @davidgerard the boilerplate could be made unnecessary deterministically instead of accepted as a price of doing business
-
@doriantaylor @davidgerard the boilerplate could be made unnecessary deterministically instead of accepted as a price of doing business
@cap_ybarra that i agree with completely; my own work is completely deterministic and i'm a big proponent of open standards, and resent the fact that every web API is ever so slightly different from every other.
i guess my observation is that people *are* using LLMs to generate code, and will likely continue to (despite being a blunt tool in my opinion), but i suspect when measured from the outside, the net gains are going to vary dramatically.
(@davidgerard has written similar-ish things)
-
@cap_ybarra that i agree with completely; my own work is completely deterministic and i'm a big proponent of open standards, and resent the fact that every web API is ever so slightly different from every other.
i guess my observation is that people *are* using LLMs to generate code, and will likely continue to (despite being a blunt tool in my opinion), but i suspect when measured from the outside, the net gains are going to vary dramatically.
(@davidgerard has written similar-ish things)
@doriantaylor @davidgerard i'm perhaps over sensitive to the "llms are ok sometimes" takes. they're overrepresented and you do not, under any circumstances, gotta hand it to em
-
@doriantaylor @davidgerard i'm perhaps over sensitive to the "llms are ok sometimes" takes. they're overrepresented and you do not, under any circumstances, gotta hand it to em
@cap_ybarra @davidgerard the empirical reality i'm seeing is a) people seem to believe they're useful and b) they're already expensive, and those people are not going to want to pay 10-20x what they're currently paying, and so the situation will equilibrate eventually around locally-runnable models that are "good enough" for rote coding tasks over existing mundane procedural languages.
like i would bet money that the halo of LLM as panacea is eventually going to dissipate, but that will remain.
-
@cap_ybarra @davidgerard the empirical reality i'm seeing is a) people seem to believe they're useful and b) they're already expensive, and those people are not going to want to pay 10-20x what they're currently paying, and so the situation will equilibrate eventually around locally-runnable models that are "good enough" for rote coding tasks over existing mundane procedural languages.
like i would bet money that the halo of LLM as panacea is eventually going to dissipate, but that will remain.
@doriantaylor @davidgerard i will remain steadfast in my belief that rote coding tasks are a skill issue
-
@doriantaylor @davidgerard i will remain steadfast in my belief that rote coding tasks are a skill issue
@cap_ybarra @davidgerard i mean maybe but i find most people just flatly do not understand higher-order programming and cannot be made to
-
R relay@relay.an.exchange shared this topic
