3 ms·

Im aware of several AI assistance tools and using LLMs to generate (hopefully) runnable code but has anyone used the available tools to significant impact? I me

Im aware of several AI assistance tools and using LLMs to generate (hopefully) runnable code but has anyone used the available tools to significant impact? I mean can I delegate a sprint ticket to an AI yet? I can see us getting there. A suitably detailed feature ticket isn't much more than a prompt. The ticket would need to get down to a code level in specification of what needed to be done. Its possible on paper sure. Fun question: AI generates some code and its deployed to production. Theres a bug found in the logs and a dev sits down to analyze the fix. Does the dev correct the prompt that generated the code and commit that prompt, or does the dev make a code fix and commit that?

quantified 2y ago
Right, training the models will still be a thing, for libraries, systems, service semantics, etc. Like with robot tenders, a smaller number, but still there. I find one of them to be pretty good at autocompleting for-loops and simple statements for me. Because it's driven from the same tab key that I use to format code with, I also spend time dismissing suggestions with the escape key. As a power tool it's OK.
wackycat 2y ago
| A suitably detailed feature ticket isn't much more than a prompt. The ticket would need to get down to a code level in specification of what needed to be done. This is the part I think folks underestimate when saying that "Coding as a career might be over". Gen AI can write decent code, but constructing software from code, and coming up with implementation ideas and turning those into tickets still feels far away at least from my experience using ChatGPT and CoPilot. Also those systems were trained on code but I don't know that they have the best idea of quality of the code they were trained on.
- SavageBeast 2y ago
  In my experience and at this point in my career ark I can say that "coding" itself hasn't been the hardest part of the job in ... 20 years? Coding was never difficult. Understanding the rats nest of code that comprises a product is a big barrier. The whole "I know this looks batshit crazy but we do it this way for a business reason" and it goes on and on. Not to mention the database full of tables that are gigantic and nobody knows what they're for anymore (This company maybe your cell carrier I kid you not). In all seriousness though, an AI at this point can do some cool things with some handholding, prompt refinement, etc such that its almost a Jr. Dev. Both more or less went to school for CS and both are minimally experienced but have a lot of code to look at and draw conclusions from. Neither know anything about the business domain. Both generate candidate solutions that will at best require some additional guidance. Personally, I don't see things getting too far beyond this in the near term. Perhaps AI becomes a very good Jr Dev who can handle "write me a function with inputs Foo,Bar that returns the [biz logic] as Baz" and the "coders" modern job is to generate maybe 50% of the necessary code where the other 50% is the integration. Who knows ... but I dont see people saying the things I think people would say if anyone was actually doing this for realsies.
  - flembat 2y ago
    Coding well is very difficult, that is one of the reasons why all of our software is still unreliable and ridden with bugs. I agree its not the hardest problem in product development.

quantified 2y ago

Right, training the models will still be a thing, for libraries, systems, service semantics, etc. Like with robot tenders, a smaller number, but still there. I find one of them to be pretty good at autocompleting for-loops and simple statements for me. Because it's driven from the same tab key that I use to format code with, I also spend time dismissing suggestions with the escape key. As a power tool it's OK.

wackycat 2y ago

| A suitably detailed feature ticket isn't much more than a prompt. The ticket would need to get down to a code level in specification of what needed to be done. This is the part I think folks underestimate when saying that "Coding as a career might be over". Gen AI can write decent code, but constructing software from code, and coming up with implementation ideas and turning those into tickets still feels far away at least from my experience using ChatGPT and CoPilot. Also those systems were trained on code but I don't know that they have the best idea of quality of the code they were trained on.

SavageBeast 2y ago
In my experience and at this point in my career ark I can say that "coding" itself hasn't been the hardest part of the job in ... 20 years? Coding was never difficult. Understanding the rats nest of code that comprises a product is a big barrier. The whole "I know this looks batshit crazy but we do it this way for a business reason" and it goes on and on. Not to mention the database full of tables that are gigantic and nobody knows what they're for anymore (This company maybe your cell carrier I kid you not). In all seriousness though, an AI at this point can do some cool things with some handholding, prompt refinement, etc such that its almost a Jr. Dev. Both more or less went to school for CS and both are minimally experienced but have a lot of code to look at and draw conclusions from. Neither know anything about the business domain. Both generate candidate solutions that will at best require some additional guidance. Personally, I don't see things getting too far beyond this in the near term. Perhaps AI becomes a very good Jr Dev who can handle "write me a function with inputs Foo,Bar that returns the [biz logic] as Baz" and the "coders" modern job is to generate maybe 50% of the necessary code where the other 50% is the integration. Who knows ... but I dont see people saying the things I think people would say if anyone was actually doing this for realsies.
- flembat 2y ago
  Coding well is very difficult, that is one of the reasons why all of our software is still unreliable and ridden with bugs. I agree its not the hardest problem in product development.