Gregory's ServerI drew this 12 months ago, as LLM code assistants were just becoming available.
Has it held up?
 42 comments
 @Viss @forrestbrazeal @amyengineer a coworker of mine has more than once received an incorrect function definition (not even valid syntax) with a body that was just "// todo" and it finally convinced me that AI will replace us all one day.  @groxx @forrestbrazeal @amyengineer at least we can take solace in that nothing 'more' will get done in tbat case, hah @Viss @forrestbrazeal @amyengineer plus uncompilable code cannot be run, so by definition it has zero bugs :successKid:  @Viss @forrestbrazeal @amyengineer I once asked it to explain how the posix dlopen function worked and instead of an explanation or working example code it wrote a nearly empty c++ class with "call to dlopen goes in this method" in a comment in one of the stub methods.  @Viss @forrestbrazeal @amyengineer How long until they generate code that attempts to call out to an LLM to generate itself? @forrestbrazeal @kstewart I keep saying “I don’t need help writing code that looks right but isn’t”.  @forrestbrazeal I heard some feedback from a doctor the other day that they are amazing while writing papers to help with things like reorganizing and sorting tables, doing statistics and other data manipulation stuff. I was like "wait, what?", surely using it for math is awful?! Turns out with the latest versions generating code and using the code to process the data they are indeed extremely fast (and correct) at that. @forrestbrazeal Most of the time when I ask chatgpt for help on complex jq queries it fails completely or just regurgitates the stackoverflow page I was already looking at   @forrestbrazeal I dunno, it probably speeds up burnout and constant reprioritizing. And incurring technical debt. @forrestbrazeal Every time someone has given me AI code, it hasn't been full of subtle bugs - it's just been so jaw-droppingly wrong that it might as well have been the answer to a different question. God help anyone who integrates AI code into their work! @forrestbrazeal it has, yes. LLM code assistants are also super fun if you're working with a feature (like offscreen pages in browser extensions that require background processing be done in a service worker) that did not exist in 2021. @forrestbrazeal This is because youre asking a crumby speech copier to come up with original code. If the code had already been written you could find it by searching, youre looking for new. Get to coding or find a new job ... but i guess complaining on lame social media is your new job? @forrestbrazeal Thankfully, haven't had to mess with LLMs in my daily work (aside from Microsoft trying to smuggle one into the damn browser). ...But I'd say you're pretty spot on. @forrestbrazeal maybe your process is different from mine, but for me, actually writing the code takes a fraction of my time so small that it's not worth optimizing away. Specifying, discussing, testing, debugging, etc. take 95% or more of the time I spend on a piece of code before it leaves my hands. @forrestbrazeal On left: Things that are actually slowing down our dev teams: constant reprioritizing, burnout, pointless meetings, technical debt. On right: Things an AI coding assistant speeds up: Googling boilerplate code In intersection: Generating subtle, devastating bugs  @forrestbrazeal If advent of code is any indication, it's also good for helping an experienced programmer be productive in a language they don't know, writing low stakes code they wouldn't be comfortable shipping. As a gag, totally holds up. @forrestbrazeal I personally love llms. You just need to see them as highly intelligent but autistic interns. They cut down dev time by a lot. I don’t think ai can go back to a time before copilot. Although it should only be used by people who already understand the code llms generate. Because that over zealous intern is sometimes incorrect and you need to be experienced enough to detect that. @forrestbrazeal @udittlamba can we not use autism to describe being bad at stuff? Considering how many people both on fedi and in dev are in fact autistic. @yayroos @forrestbrazeal @udittlamba Indeed, this is pretty rude to a whole lot of people here (and also overlooks the fact that a shitload of excellent engineers are autistic). @udittlamba @yayroos @forrestbrazeal well as someone on the spectrum,I will politely ask you to stfu about things you have no experience. We have enoufh already with reminding people we are.not.idiot savant all the time @joel_falcou @yayroos @forrestbrazeal dude you have a phD can your comment be any less ironic? 😂 @udittlamba @yayroos @forrestbrazeal thanks for confirming you know Jack shit on the subject. @joel_falcou @yayroos @forrestbrazeal I don’t listen to any white man telling me what to do. Keep your privilege to yourself. @udittlamba @yayroos @forrestbrazeal well ausitic people don't care about fake outrage. We have other thing to do... @yayroos @forrestbrazeal you guys are so privileged to be offended at the stupidest shit. That’s what living in the first world gets you and I hate it. You trivialise issues by making everything an issue when there isn’t any.  @forrestbrazeal when ChatGPT subtly mistated the Law of Cosines, it sent me back to pencil, paper and protractor for two weeks. So, yes! (But I still like ChatGPT.)  @forrestbrazeal as a CS dropout the boilerplate code generation is actually helpful for me @forrestbrazeal I have avoided LLM technologies wherever possible, for reasons also including the reason in the intersection. I am not expecting positive results that would disprove my objections in this timeframe. (Also do not want to train someone elses model without receiving compensation or giving my consent.) Does anyone accept the challenge? @forrestbrazeal If you’re talking about it accumulating, you could reasonably put tech debt in the overlap. @forrestbrazeal oh thought it would be the sense of dread of being a gear in the destruction of humanity. |
@forrestbrazeal @amyengineer mostly!
about half the time i ask gpt to give me some bash or some python the codeblock it responds with has commented lines like:
# .. and the code goes here
which .. is sorta the opposite of actually writing the code, so in some cases its just too lazy - which i suppose means it was trained on that sort of code to begin with