[unzadunza]

> 81% of developers agreed that AI increases their productivity

I've had a few AI generated PRs come my way and the code-review process is, shall we say, not fun. It takes me a lot more time to review these PRs, there is way more back-and-forth between me and the 'developer', and it takes much more time to get the PR merged. That's not saying anything about the increased difficulty in modifying this code in the future.

I have a feeling these claims of being more productive don't account for the entire development cycle.

[mattas]

I think employees (myself included) often think, "I'm more productive" when in reality what they are actually experiencing is, "My job is now easier."

Easy does not necessarily mean more productive when you're trading ease for something else. In the case of coding, you're trading ease for things like understanding and maintainability.

[zingababba]

Yes, exactly.

[ModernMech]

Right, a machine that automates the planting of mines may make the user more productive at their job, but that doesn't really account for the time spent cleaning them up on the backend and how many limbs will be lost in the process. AI is an automated landmine planting machine.

Sure AI increases developer output, which is sometimes correlated with productivity -- but often times not. Insofar as AI is accelerating positive outcomes (we were able to write some tricky code that was blocking us), it's also accelerating the negative outcomes (we used the LLM to write 40k lines of code in an hour and no one know what any of it does). One of these things is "productive" the other is just performative work.

If "being more productive" is using an ai to write an email which is then summarized by AI on the receiving end, or students using AI to write papers which are graded by AI, or developers using AI to write code which is then reviewed by AI, then AI isn't actually making anything better.

[itsdrewmiller]

The recent RCT on open source ai bug fixing had most participants feeling like they were more productive but actually being less productive. It may just be perception error, and then the issue you identify makes it even worse.

[riku_iki]

AI search makes me personally way more productive, e.g.: write snippet how to do X using library Y I never touched before.

[bootsmann]

Maybe as a small anecdote: I had a coworker ask an AI to write them a function to auth to aws using boto3 and environment variables. The LLM happily complied and gave them a 30 line snippet but it failed to point out that if they were to open boto3 docs it would tell them that it will already default to environment variables when it has no credentials and the whole function is just reinventing the wheel.

If you’re throwing the LLM at APIs you don’t know, how could you possibly verify it is using them properly?

[riku_iki]

The question is how fast and well dev unfamiliar with that API would achieve the goal alone compared to being assisted by LLM: he could set up snippet, run test, see what are the error, ask LLM to check why error could happen.

[bootsmann]

I don’t believe that the time a dev spends on prompting an LLM over reading an API reference saves a significant chunk of the TCO for the lines of code they write.

[riku_iki]

I personally observe very opposite: it safes a lot of time.