@ploum Nailed.

@ploum Once Upon A Time, in the days before agile, when people did things like "design" and "project management", our typical project spent about 10% of its time on coding.

Which would occasionally get our customers really quite nervous when they could see that we were nearly half way through the timescale but hadn't written any code yet.

@ploum Software development is in almost all cases a collaborative effort, requiring *at least* one developer to be able to explain or document the code for the rest of the team (as well as the future team). When code is generated by an AI, this is missed and no-one really understands the code so no-one can really maintain it making all post-1.0 work very HARD.

@ploum once the "AI" bubble pops, we software engineers are gonna have so much work.

let's just hope I have enough health left by then.
and if I do, that i haven't found a more sensible job

@ploum

Also hiring less competent engineers will lead to no one even knowing about step two and only performatively participating in step four dooming the whole process to failure but the process will be doomed much faster and more cost effectively so AI is the future.

@ploum I think 1 is the absolute hardesrt. Of course, followed closely by 5. And your main point about the consequence of using chatbots to replace competence still stands.

@ploum I'm not a _typical_ user for certain, but I find a good LLM assistant to be quite valuable to me in all 5 of the steps listed, plus at least 2 others (0. Decide if I want to build vs buy, and 6. Document). It won't do the hardest work for me, but it saves me a lot of time and results in a better product by helping me think through hard problems (best rubber-duck ever).

@ploum It can make the first one easy if you don't mind terrible ideas that have a 30% chance of being either illegal or simply impossible to accomplish.

@ploum somewhat easier is huge stretch considering I cannot trust it.

@ploum

Hmm.

I think that step 1 is actually the hardest.

Normally, in fact step 6 is redoing step 1 because you find that the software is not supposed to do that at all, or at least, the no-one wants the software that you have written.

At any rate, your points about chatGPT are bang on.

#Ai #chatgpt

@ploum I dunno I usually use copilot snippet as a starting point,like a sightly better intellisense. Nothing that can replace programmers, but not completely useless

@ploum very well summarised! Thank you!

@ploum I have on occasion suggested 'WaterBoarding' the customer to get the full software requirements.

@ploum meanwhile I want to see the ChatGPT code go fecking rogue and sudo rm -rf --no-preserve-root /

@ploum Using AI for code generation adds:
6. Make sure to publish the result with the correct license (literally impossible)

@ploum This is what I have been saying. Writing the code is a small portion of the job.

I still wonder what those 'LLM makes me so many times more productive when developing' do differently.

@ploum Writing code is easy.

Writing good code - which means having gone through 1 and 2 first and making 5 as easy as possible - is very very difficult.

@ploum As act of self-critique I have to admit that for 99% of my personal projects I stuck at step 1.

@ploum one friend, who only writes code when (2) doesn't matter (i.e. relatively simple/throw-away software), says chatGPT allows him to skip straight into debugging. Which is handy, but only if you're already good at writing software.

@ploum Precisely.

@ploum
So, you integrate steps three and five in an iteration loop, and the person who does it will be called Prompt Engineer.