@Tom Even from experienced ones.

The moment I see an AI generated PR description, I’m losing motivation. It rarely provides useful information, but always floods you with words

First point, I disagree: I'm a modest programmer and I learn from AI code when I review it. Quite a good bit in fact.

Second point is spot on: coding was never very "humanizing" to begin with, but this is a step further along the dehumanization process. The sense of ownership, of caring, is defining of humanity.

”Outsourcing your thinking is a fascist idea, because our power comes from thinking together.”
-Naomi Klein

youtube.com/watch?v=iEf-MNsyUi…

interesting observation.
I've seen similar detachment in the early 2000s when companies didn't employ people but used lot's of external staff. That was not a good time...
Overall mood was evil, quality got shittier, don't ask about loyalty...

P.s. I was a mechanic and machine operator back then

Exactly.

I'm forced to vibe code at work and I said "ok, I will vibe code you a new app, but I won't touch the code even with a stick", I'm only reviewing for security issues and the tests.

Once in a week, I ran some prompt like "the users says it's slow", "a coworker looked at the code and find it ugly", and the LLM give improvement.

I can't care less about the code quality, because it's crap, once I ask a patch for a person all project, it give me 200 loc ... My fix was 12.

A pattern with team members who are using "AI" to write their code. I would point out a few items that needed fixing. I would get back a new commit, that sure, fixed those items, but introduced a slew of new entirely unrelated items.

Turns out, they just went back to Claude and asked it to address the code review, which threw away the old commit and generated a new one. They didn't really even think about my feedback or how to incorporate it.

one of the things I enjoy about writing code is the flow state you get when deeply exploring your understanding of a problem & solution. That makes all the other bits of the job worthwhile - even the tedious tracking down of the last few lines of code test coverage. Writing the PR is a concise reflection of that journey.

I can understand how losing those moments of productive flow to a machine would be highly demotivating. (Along with the mentioned lack of ownership/pride/care...)

Alternatively: 'Own' the app (rather than the code) and let AI deal with javascript (/typescript) fatigue? We work less for the other developers, but more for users?

When the crucial stuff is buried within framework internals or 2 GB of node modules, how much of the code is mine anyway?

Maybe the way web development has taken in recent years, agency was on the way out before AI already. <Humming "Bootstrap killed the CSS-star…"> 🤷‍♂️

Just a remark: I tend to think that the "I am NOT my code" is the way to go

Code identification/possession (taking critics of code as critics to the coder/self), has much less merits than one may think

I like the Egoless code approach: it allows improvement to occur seamlessly

#egoless #softwarecraft #softwarecraftsmanship

💯. The process of thinking about the problem, the architecture, the solution and finally writing the code creates a strong emotional bond for me. I care about the result.

As you said. For me this does not happen with generated code. I simply do not care for it. It does not motivate me the slightest.

I would guess that this is especially true when people work in a so-called "feature factory" and can't really identify with the software as a whole.

If people can identify with the software as a whole, at least this aspect might look different.

Totally agree.

Another anecdote: a few weeks back I used the plan / code feature of Claude Opus to create a component which called a rather complex 3rd party API. Code looked great, but I realised I didn't even know how to invoke it for a test run.
Reverted back to reading the 3rd party API docs (which were good), then found lots of small loose ends in the generated code.

Mixed blessing, lots of boilerplate were done w/o any issue, but fixing & "making it my own" still much effort

There have already been more than enough studies that demonstrate, the more someone uses "AI" in their job, the worse and worse they become at their job.

We're SUPPOSED to grind over menial tasks, especially mental ones. That's what builds new and better pathways in our brains.

"AI" literally makes us dumber

I think there are two ways to work with coding agents and not hate everything about it:

1) Vibe code, don't touch the code unless absolutely necessary. That's okay if you build prototypes or simply don't understand the technology you're using.

2) Go through every line, make small amendments, try alternatives. You make the code your own and are mostly still faster than without AI.

IMO passive reviewing doesn't work because the cognitive load is similar to doing it yourself but it's much more demotivating and you'll tend to just approve.

If it's not an option to kick bots out entirely, one thing that's doable is to take PR feedback and shove it into AGENTS.md and docs linked from there. You *can* actually get the machine to "learn" repo conventions to reduce future PR load, even if the person piloting the machine stubbornly refuses to internalize any feedback.

(And you can throw tokens at this codification if you like).

I'm doing this on an internal project and it does seem to work a bit 1/2

I absolutely feel the lack of motivation reviewing LLM code.

I usually do code reviews as part of security audits, and when I look at the commit history of projects that rely on LLMs, the code base changes so fast that I have to expect my work to become outdated and basically meaningless within weeks or days.

@kjetil_kilhavn that was my experience

The other devs apparently found my comments somewhat entertaining at least. Lately my responses are single words to each "point" the AI makes. It sucks at trying to interpret code but is okish at catching some basic lint errors and at noticing the occasional copy/paste error.

Drives me nuts at how padded out its comments are when most could be single sentences.

@nikodil Problem is that when the only thing visible for the reviewer is the output of a 3rd party tool (without input parameters and specs), it becomes really difficult to debug even with time and energy.

Code generation itself can be useful, but only if the model itself is generating consistent & predetermined output structure. LLMs mostly fail at both.