Solution is clear, just let an LLM review the PRs!

The article actually covered that. And that was growing over time too.

It should be noted that the data cutoff was March 2026 and that significantly better models have been released since then. The study also only covered a 12 month window after AI adoption and that companies of these sizes generally have a lot of internal inertia for new tools

Normalise the Prompt to Prod workflow. Just give the AI a vague description of what you want, and your credit card details, and let it publish to prod on the first try.

Every time I test and LLM solution and am pleased with how well it works, and I look at the code and it looks reasonable, it’s a lie. It may work, and even look okay, but unless the change was minuscule it’s always doing weird things when I look closely. By the time I’m done fixing the code I could have done it significantly faster myself.

Edit: in case anyone is wondering I only ever use frontier models (the anthropic and openAI ones), I have unlimited access to them from work and mandates to use them.

Not the case for me. I even let an AI basically generate the whole project and it mostly worked. I found it didn’t work well with IPv6, but my server doesn’t have an IPv6 address. Nevertheless, I just asked the AI to fix the bug.

No chance I actually do all the work myself

You clearly don’t know remotely enough to understand the generated code. People with more than half a brain also don’t call a slop bot “AI”

And you know this how, exactly?

You let it generate the whole project likely because you are incapable of doing it yourself.

The OP you’re responding to, like most of us that dislike AI slop, don’t use it because it can do things we can’t. We used it because it was supposed to be a productivity booster.

While it does generate code quickly, it’s not at the level of quality we would write, and because of that, we’re constantly having to tweak and refactor it afterwards, losing any productivity gains.

I wrote the same thing as shell scripts, but to use capabilites instead of running it as root, I needed to generate an executable. So it’s not like I didn’t have a worse solution already in place. Whether or not I can write it is irrelevant, I didn’t want to spend the time to do it myself if AI didn’t generate it for me

and how does larping as a programmer work out for you?

If I’m LARPing as a programmer, then ask the more reason to use AI

nest.pijul.com/iopq/proxy

This is the project in question. It’s not a secret or anything

There’s no readme so I don’t really understand what this is for. Plenty of proxy solutions already exist…

that website gave me eye cancer

The only way it helps me is when I basically use it as a search engine, and it generates some wonky code that I have to fix. For example, instead of checking how you override the mouse in c# in a godot node, it gives me a basic blueprint of that (usually with unnecessary stuff added), that I then change and use.

Or recursively scan a folder in python.

It never gives you exactly what you want, but it often has the underlying structure. It’s like going to stack overflow but a bit faster.

I do agree it’s pretty helpful for remembering things I don’t use often or finding APIs for things I haven’t done before. Also not a bad rubber duck. Not worth burning the planet for, but smaller local models that can handle this job I’d be happy to have in my workflow

The moment AI can take the dumbest, completely lacking detail, uninformed, uneducated, unmotivated idea from a (Project/Program/Moron) Manager and turn it into something worth value… I’ll know that my job is in danger.

Oh, it can. It’ll do a crap job of it, but lots of people think that’s “good enough”

Yeah this rings true to my experience.

Any efficiency increased during the actual coding phase, the study authors find, is “absorbed by downstream constraints in the production process”; as “the code review process significantly increases in length, pull requests are more likely to require revisions, and reviewers leave more comments.”

Yep. Almost like actually understanding problems and computers was the job, and not just typing.

My job recently has definitely been to fix the UI that the robot makes because for the life of it, it cannot follow our UI library components and the way we set up our React files.

I don’t use the bot, but I end up having to clean up after it from other people. And it does take a while to just parse what the thing did, come up with a game plan, then fix.

Some people might say, “well just have the bot fix it!” No. I don’t want to use the bot.