I was thinking of going into accountancy. But then the bots are probably coming for them, too.
adamu
BartOtten
I let AI write the quartely review of those using AI. Every time its exiting to see how they perform! Fired 2 already purely bases on AI slob in the review reports.
0
Lucassifoni
What if this was more of an organizational issue than an individual one, the same as always, just exercised in a leveraged way ?
Developers have « responsibility » without stakes (being, at least a tiny bit, personally tied to the project success, and thus rewarded accordingly or in danger accordingly in case of full failure) nor trust (micro management and surveillance) in a lot of orgs, so no real responsibility nor ability to contribute something more than LoCs
This creates incentives where the org wants to extract as much value as possible from the dev who will in return try to minimize effort for a fixed output (salary).
Software coops ?
Siloing devs in an only technical spot interfaced with the world via tickets also increases this IMO. AI reducing development time is fine, if this leads only to « build more » the mismatch between needs, architecture and product can only grow.
I think we should use the liberated time to actively speak with downstream consumers of our work, whether this is software users or other developers consuming a library. This would avoid a lot of useless building. AI is great at removing code too.
2
Asd
You’ve completely missed my point. I can state it once more again:
Reviews (including self-reviews of generated code) work when developers have knowledge about the problem. Developers used to learn much more about the problem when they were writing the code themselves and thinking about it. Now developers generate code, what doesn’t make them think about the problem and as a result they don’t learn. My reviews on this forum are a proof of it.
Again, I am not talking about AI in general and the quotes you provide are funny, but they are off topic. The topic on this discussion is “how to stop the slop” or “how to ensure the pre-AI level of code quality in a team which generates code”, not the discussion whether the AI is useful in the first place or if we all should adapt to the new era, or whatever. And my point is: I think that reviews and super accountability are not the answer. And I am looking for the solution to the problem.
To expand my point there is also this idea. I think that no team can 100% rely on the most experienced engineer to perform the reviews. Because their knowledge is limited too.
Here’s an example to my point, imagine you had to implement some rare cryptography code (like some version of sha hash function validation) in order to integrate with some old system. You could not find a dependency, so you generated the code. Code contains some very critical bugs, but you can’t find them, because you have never implemented a cryptography algorithm. Reviews passed (because your most experienced engineer has happened to never implement a cryptography code before), code merged, that’s it. But if a person would have tried to learn and read about cryptography algorithm implementation, which they would have to do if they were to write the code, they would learn how to handle out-of-bounds errors, how to perform byte sequence comparison, how to derive keys from passwords, etc., they would be able to find bugs in the code they generated.
4
mudasobwa
Creator of Cure
Well, it’s a great question who actually misses whose point. Your reviews on this forum prove nothing, because they have nothing to do with the real-life products by any mean.
There is no such reality in what the team of professional developers, delivering the code for years, forgets everything they knew and starts to ship a slop. Wannabe unicorn startup founders?—Maybe, but I’m not interested in guiding them on the right path. Grown up developers would not lose the knowledge of the problem because they used assistants.
Now developers generate code, what doesn’t make them think about the problem and as a result they don’t learn.
Ugh. Yesterday developers copy-pasted stuff from StackOverflow.
The topic on this discussion is “how to stop the slop” […] I think that reviews and super accountability are not the answer.
That I understood. My point is, they are (it’s great different people are still allowed to have different points of view, isn’t it?) As yesterday and as ten years ago, when the crap has been attempted as a merge request. Are you saying that the problem of not-so-experienced developer trying to push the spaghetti wackyslop arose out of the blue three months ago?
no team can 100% rely on the most experienced engineer to perform the reviews
Nobody ever demanded the reviews to be done by the most experienced engineer. You’ve said earlier that developers don’t understand their code/slop anymore. During CR they should, even if they are trainees.
imagine you had to implement some rare cryptography code (like some version of sha hash function validation) in order to integrate with some old system. You could not find a dependency, so you generated the code. Code contains some very critical bugs, but you can’t find them, because you have never implemented a cryptography algorithm
I see here a logic gap. How come I have to implement some rare cryptography code if I never dealt with any cryptography code before? Even if I have to, for some reason, what makes you think my hand-written code is to be bug-free? And, the last but not the least, why do you think I would copy-paste the generated slop without understanding it?
The statement “if they were to write the code, they would learn how to handle out-of-bounds errors [and so forth] would be able to find bugs in the code they generated” requires a proof, at best.
NB I agree that writing code teaches better than reading code, that’s evident. What I strongly disagree is that there is the brand new problem of slop leaking to codebases all around. We learned hard way how to stop shitty code to get to prod (to some extent,) and these methods are applied to AI-slop all right.
1
tfwright
Well, your point seems fairly clear: “AI changes nothing. Some devs were talented before, some not. Now it’s the same. Talented devs will use AI well in the same way they have always made good use of their tools. Devs without talent should be janitors etc. The slop will be stopped in the same way it always has. When Good Devs Make a Stand (against God if need be)“
If I have missed something there, then apologies. And I can’t speak for @asd but from my POV, that last parenthesis is not just a joke because we are dealing with market forces here, not just individual choices. No matter how talented of a dev I am, the pressure is currently immense to ship slop even if everyone involves knows better. Because the alternative increasingly looks like closing up shop, maybe as an individual developer, but also maybe as a whole org.
Of course I can tell my CEO “no” to some individual PR he is opening. And I did. And then I was on an unscheduled 1 hour call, and then an unplanned 10+ hour refactor of his PR which was rushed because, unplanned. And after that I had to start preparing a memo about how we are going to change processes to accommodate these AI PRS because they are not stopping (thus my attention to this thread). That is not something I can say just say no to and still plausibly claim to be doing my job. I told him what he was doing wasn’t going to work and he said ok tell me what will work.
Individual dev responsibility is going to be part of that. But it’s got to be more than “get gud", because while I have a good amount of political capital I don’t have that much. Maybe you do (maybe you are the CEO!), but I think more people are in a similar boat as me.
Another point I think you brushed aside much too quickly is the issue of speed. Copying from stack overflow might be quicker than properly grasping and solving the problem, but it’s not so much faster, or nearly good enough, that I can’t deal with it within the scope of the traditional review process (not that it has ever really come to that–I’ve had some hiring mistakes but not that bad). These AI PRs are simply not the same thing. They are not a poor implementation of a ticket. They are massive new feature sets, massive new refactors, not even in a ticket yet because we haven’t even had the time to plan them. I am not telling the CEO, do what I say and you will get the same features, less bugs, barely any more upfront resource use and a hell of a lot shorter tail, as before. I am asking him to just not implement that wishlist feature he’s been dreaming about, along with 5 others.
We are not talking about graders vs shovels here. Technological change is a spectrum, but there are also paradigm shifts that do occasionally happen, which are specifically defined by their incompatibility with existing practices. I think perhaps we in the tech industry are a bit jaded about them because every marketing hack loves to sell their product in such terms. And I could be wrong but this sure as hell seems like it might be one of them.
edit just looked and there are 12 new PRs on our main active project opened this weekend by AI
8
dimitarvp
I don’t think him or anyone else is saying that. The real problem here is that the ability to drown people in slop has been hugely amplified by LLMs. And that’s visible in many places.
In the world I inhabit, I have to follow orders. If you are not living in that one, I want a ticket to yours.
CEO calls and says: “I want this delivered tonight”.
2
egeersoz
I think it is a bit silly to try to review AI-written code without your own AI agent assisting you, frankly. Because, as it has been noted, it’s simply too much code to review, and the stakes are often very high.
My workflow starts with asking an AI agent in Cursor to pull the PR branch, read the diffs, compare them against what the Linear/Jira ticket says should be done, and giving me a summary. The benefit of this over using one of those AI-assisted PR review solutions is that Cursor has a semantic index of your codebase and so the coding agents have a lot more context (plus memories of past conversations). They can also run code using Tidewave to double- and triple-check the PR’s logic themselves.
If it is a larger PR I might use one orchestrator agent that can launch multiple sub-agents and have it combine their reviews for me to consume. In Cursor I have defined a reviewer sub-agent for each of the SOTA models, so I can say “launch the codex-reviewer and opus-4.6-reviewer on this PR, validate their feedback and give a summary. From there, I can do additional verification and use my own knowledge of things like business requirements, codebase history and future goals/plans and draft detailed feedback.
This has been working really well. Team members frequently remark on how useful my reviews have been, and some have started to DM me on Slack asking if I can review their PRs, even though I’m not on their “squad”.
3
tozz
Don’t forget all the AI BS people are reading, like how CharGPT cured a dogs cancer (it didn’t). People without skills nor ability to do proper research will point and say ”I want that”.
This wasn’t a problem with stack overflow because you didn’t open your newspaper and read about how it did x only using y on the almost daily. There weren’t workshops on stack overflow in every major city, VCs weren’t forcing you to be ”stack overflow native” etc etc, there just is no comparison really.
3
mudasobwa
Creator of Cure
Well, in such a case I would write a long email to my CEO mentioning the reasons stated above in this thread, explaining how this is a threat for “the future us,” and why we should not act this way.
Then I’ll deliver the code and change my status to “looking for a job.”
There are many companies buring themselves six feet under, with AI, or without it. The quality of their codebase does not affect anything save for their own existance. The quality of OSS does, and I am positive there is no “must be delivered tonight” in OSS narrative yet. So we are fine, IMHO.