It is even more funnier when you realize that because Claude and all AI models are trained on data including stackoverflow.

So I guess if you asked Claude why it did that, the truth of it might be "IDK I copy pasted from StackOverflow"

The same stuff pasted with a different sticker. Looks good to me.

Ha, I genuinely laughed at that. Thank you!

This is exactly how I see it. Is not about the tool, is how it is used. In 1990 that would have been “IDK I got it from a BBS” and in 1980 “got if from a magazine“. It doesn’t matter how you get there, you have to understand it. BTW I had a similar problem as I was manager in HW development, where the value of a resistor had no documented calculation. I would ask: where does it came from? If the answer was “I tried and it worked”, or “tested in lab until I found it” or in the 2000 “I run many simulations and was the best value” I would reject and ask for proper calculations, with WCA.

As vibe coding becomes more commonplace you'll see these historical safeguards erode. That is the danger IMO.

You're right, saying you got something off SO would get you laughed out of programming circles back in the day. We should be applying the same shame to people who vibe code, not encourage it, if we want human-parseable and maintainable software.

> That is the danger IMO.

For whom is this a danger for?

If we're paid to dig ditches and fill them, who are we to question our supreme leaders? They control the purse strings, so of course they know best.

> If we're paid to dig ditches and fill them

This is a very cruel punishment sometimes used in forced labor camps. You are describing torture.

I don't think it's common, but I've definitely seen it

I've also seen "Ask ChatGPT if you're doing X right?", and basically signing off whatever it recommends without checking

At this point I'm pretty confident I could trojan horse whatever decision I want from certain people by sending enough screenshots of ChatGPT agreeing with me

I don't think this is an AI specific thing. I work in the field, and so I'm around some of the most enthusiastic adopters of LLMs, and from what I see, engineering cultures surrounding LLM usage typically match the org's previous general engineering culture.

So, for example, by and large the orgs I've seen chucking Claude PRs over the wall with little review were previously chucking 100% human written PRs over the wall with little review.

Similarly, the teams I see effectively using test suites to guide their code generation are the same teams that effectively use test suites to guide their general software engineering workflows.

How long are you spending with a given team, and where per se in their "AI lifecycle?" I would expect (for example) a sales engineer to see this differently than a support engineer, if support engineers still existed.

"Look, the build is green and CI belongs to another team, how perfectionist do you need us to be about this?" is the sort of response I would generally expect, and also in the case where AI was used.

If it pushes some nice velocity metric, most managers would be ok. Though you have to word it a bit differently of course.

Of course they are okay with it. They changed the job function to be just that with forced(!) AI adoption.

What about "the compiler did that" ?

Depends on your incentives. People anecdotally seem far more impressed with buggy stuff shipped fast than good stuff shipped slowly.

Lots of companies just accept bugs as something that happens.

Depends on your problem space too.

Calendar app for local social clubs? Ship it and fix it later.

B2B payments software that triggers funds transfers? JFC I hope you PIP people for that.

The same developers submitting Claude submissions can take 1-2 minutes asking for an explanation of what they're submitting and how it works. Might even learn.

Stack Overflow had enough provenance of copying and pasting. Models may not. Provenance remains a thing or it can add risk to the code.