Post

"I will not touch anything with ai-generated code" is going to mean, in pretty short order, "I will not touch software" this is just functionally what software development is now and it's only going to accelerate from here

ceej@ceej.online

I think it’s important for anyone who (completely understandably) checked out on AI the second the crypto/nft/metaverse/snake oil chuds started pushing it to realize that, in at least the domain of software development, we are very much living in a scenario where the snake oil actually works

there's an enormous amount of poorly vibe-coded slop out there and unfortunately that is also going to accelerate but minus some individual passion projects almost nothing is written entirely by hand these days

19 reposts472 likes

Depends who you ask and what they're working on. It's easy for me to avoid, but one of the frameworks I use for conveniently making stuff in the JUCE gui layer has gone heavily AI. So what I add to it isn't, but the tools I'm using can't be AI-free. I'm watching to see if it stays good.

I don't know why anyone would write a passion project that way either, unless they are brand new to coding. There is always boring work that does not require much thought, like logging or unit tests.

This is true and I'd also point out software made "entirely by hand" hasn't been a thing for a long time before ai turned up. Most modern coding frameworks "scaffold" the basics of your project and have done for decades

And the reason vibe-coded stuff is bad is that the humans responsible neither know nor care how it works or if it does so properly. Which is just a function of whether they care and their human-created work would be bad for the same reason.

yeah, I think this is a good place to draw the line - replacing human artists with the machine is bad but the ship has one hundred percent sailed on the code side

gubbbo.bsky.social@gubbbo.bsky.social

I guess I can't avoid the code. But I can certainly choose to avoid paying money for image gen assets, voice, images etc.

This is also my personal boundary with it. I ask Google questions about things that would be hard to find otherwise, but I'm never going to use it to generate images. Art is art, and it's about the process and not just the end product. Plus apparently image generation is orders of magnitude worse

for the environment. I will say, I miss the days when AI slop was mostly Obama, Trump, and Biden playing Wii Sports. It was AI slop, but it was CLEVER AI slop.

I don’t need code to have interesting things to say, i just need it to work I don’t just need pictures on my wall or books on my shelf, i want them to have interesting things to say

It's also going to be hard to draw a clear line about AI art, for much the same reasons. If an artist sketches a character and asks the AI to clean up the lines, does that count? How about drawing a detailed 2d image and asking the AI to make the 3D model?

Then I'll still do everything I can to sink that stolen ship. To be clear, slop coding does not NEED to involve mass theft from writers and artists and everyone else's personal data, but right now, anything based on Claude or ChatGPT still does.

Why do people give up the "all the input is stolen" argument when we're talking about code? The enormous theft at bottom here is very salient when talking about image generation, but the coding is from swallowing github whole. Is it just that people are less sentimental about code?

Training an LLM on all of human literature may or may not be legal; court cases continue about fair use and while the labs have lost several they've won others. Training an LLM on open source code is obviously legal and allowed by the license. They are not trained on "stolen code'.

Moral questions are harder. The open source licenses were drafted in the pre-LLM era and the vast majority of the code was written then too. Nobody who contributed to the Apache web server *intended* to allow training on their work. But... they did. Whereas authors (probably) did not allow it.

I think we're going to need to keep evaluating What It All Means and What We Demand as more and more ppl come in to the industry without ever learning to do things by hand and thus understanding AI's output. Best case is it's like car mechanics, where fewer ppl need to actually know how a car works.

I think we are, to a certain degree, going to find out it's not so good for code as well. Currently companies are pumping out vulnerability fixes that AI is finding (and usually fixing) but we don't have the long term data to know if those fixes are any good.

What do you mean "long term data if those fixes are any good"? Did they fix them or not? Accelerating the detection and mitigation of security vulnerabilities is pretty huge.

another piece of this for gaming specifically is that games are full of weird-ass corner cases and these tools have gotten shockingly better at diagnosing those - I work in a large (non-gaming) codebase and recently I have been able to hand it a bug report and say "diagnose and fix this" and it will

Dealing with an asshole at work, whose job description is NOT web designer, vibe coding a web app that I now have to deal with on a project not a single person warned me was coming. The "app" is in ??? code (🤷‍♀️) and isn't branded with our colors etc, and with zero UI testing.

I will be treating it on a MAYBE wireframe v1 basis, I'm certain once I dig in I'm gonna find bad UI shit that I'm going to have to convey we don't want with the REAL devs I'm working on the main website redesign with RIGHT AS THIS IS HAPPENING. But no one saw fit to include me on this convo to...

I just wish it didn't cause so many weird issues also. I've been inundated with new bugs in areas that shouldn't have been touched in the code but AI messed with anyway. Another agent will fix what the first one broke, sure, but why is the first one breaking things to begin with?

Which tools? LLM-based debuggers struggle with nondeterministic physics and frame-timing bugs—corner cases that static analysis can't reproduce.

At the end of the day, a product is polished -- no glaring bugs, no 7-fingered humans, no plot holes, no gaping flaws in gameplay, a good and clever idea at its core, a good implementation of that idea -- or it is not. There are areas in development where AI can be utilized, and where it shouldn't.

But the difference between a product that uses it but maintains proper standards throughout, and one where they just said "let the AI do the work" and stamp their name on it will always be obvious.

To be fair, this has been a big problem for over a decade I worked somewhere that about 25 years ago a small group of the devs quit and started a consulting company, then told the original company "You're going to hire us, because no one else knows how this code works" And they all got raises

Yeah this is absolutely not a problem in games alone. I have seen some suggested solutions for this, such as having the AI document its own code and have humans review it as part of the review process, so there are potential methods to deal with this. No idea if they are effective to be fair.

Arguably, outside of live service games, tech debt feels less consequential to me for games than for enterprise stuff which probably has a longer maintained lifecycle. Caveat that I’m an engineer but not a coder. AI compounded tech debt in things like ATC and financial stuff scares me shitless.

Yeah we recently had a case where the devs looked at some logs, said hm maybe we had a deadlock, then punted the whole thing to Claude saying "we suspect a deadlock" it said "yeah this routine is fucked, mate" Then we still had to figure out where this routine came from and fix it, but it helped

As a dev, I gotta say I hate this. Fixing a bug is sometimes like tightening a screw. Other times it’s like replacing the whole roof. Telling AI to fix a bug might mean it replaces the roof when you just needed to tighten a screw, and vice versa.

This whole shit has not freed me up to “become an architect”, it has stuck me in an endless stream of code review. Mine, LLM additions, other people’s PRs - code review is what I spend most of my “freed up” time on.

This btw is my theory why we haven’t seen great productivity gains in software despite AI. We have these tools that lower the barrier for code generation, but as any dev knows, lines of code has no correlation with productivity.

automated unit testing and stress testing being handled by an agent is leagues ahead of where it was even two years ago it’s so incredibly helpful

"wait how the fuck did this change break this other completely unrelated system" is a fun thing you get to ask with every patch

You can also tell it to generate a program which prints "hello, world" and then tell it to check over its own work and it'll hand you about 12000 lines of code suggestions.

because is/ought posting is in fine form on bluesky today, a pop quiz: as someone who works in software and had what seemed to be ironclad job security only a couple of years ago, do you think I want anything in this thread to be true

I regret to inform everybody that, yes, AI DOES largely seem to be the future and really ISN’T a bubble/fad in any meaningful capacity, especially in the ways that most people think about those phenomena.

I take some relief knowing my job is secure. But I sure damn hope IT isn't pulling out their hair more than they do normally. I've kinda accepted AI code is gonna be something I run into no matter what. Best to just keep making do as we can.

Maybe it's copium, but I don't think AI truly affects the security of those with the skill to understand the abstract issues and design, particularly in complex systems. I'm not writing code by hand any longer either, but so much of my job is still really understanding what needs to be done.

Maybe I'm just in a niche field, but is everyone else not doing really deep code reviews? My background is engineering and not computers, my coding is self taught, but it seems like it would be horrendous to review AI written code?

You might have too narrow a view of software. For example, AI remains pretty useless for firmware. There is also more conventional work, but which involves super sensitive IP. But, in general, yes, steaming piles of vibe is becoming normal. But since models collapse when recursively trained… 🤷‍♂️

I was a pretty top notch software engineer 3-5 years ago. I suspect I'm unemployable now. What I was great at (clean code, high quality, easy to read/modify, strong APIs) is almost meaningless in the age of AI. Glad I'm out and don't have to worry about it.

You might want to consider actually resisting, then. Rather than embracing the "oh well, I guess that's how it is" doomerism. Perhaps look at works of Emily Bender, Timnet Gebru, Ed Zitron, and many others here. You don't have to buy what the LLM pushers are selling.

I was already burnt out on software development and IT for a variety of reasons, but the rise of AI definitely signaled to me that now was the time to get out.

Yeah, as a software dev it is truly bizarre. I do think there will continue to be a period of uncertainty around our job security but ultimately AI will also serve to make orgs that thought throwing people at things and churning out spaghetti was the solution 10x “better” at that.

Worked on SAP/ABAP before I retired years ago and in this case, I would have gladly welcomed these tools. I haven't kept up since I left the field, but I imagine that my former employer's IT department is now much smaller.

Yeah, I really hate where we are with this, but for things like tracing a gnarly bug, it’s hard to argue with it’s ability to be relentless in a way that’s difficult for me to be, at least without completely burning out.

I think part of this is that software engineers are typically not organized/unionized (by design. larger salary, larger carrot). Add on to this the demanding and entitled attitude toward free software (just look at the wave of decomps) and it's a perfect shitstorm for this

Gemini helped me quite a bit with rewriting a budget spreadsheet to eliminate some bad older dependencies and solve some bug reports from beta testers. I didn't let it write my formulas, but it was very helpful in diagnosing why they weren't doing what I wanted.

I’ve been able to do this with my regulatory reporting. Just take my 120 pg summary report for the FDA and do a comprehensive gap assessment against the most recently released guidance documents. It’s saved me WEEKS of work, in some cases. And I’m not even on the clinical research side.

A friend at a small 10 man shop were estimating it'd take a man-year of time to do an Unreal engine upgrade because they were badly out of date. AI did it in two weeks, during which the CTO that guided it was mostly able to do other work. With engineering, that genie's not going back in the bottle.

I knew we were fucked by AI the day all my artist friends went “I will never replace my work with a machine” and all my coder friends went “Oh thank fuck, I never have to code again”.

I have Thoughts and Opinions on vibe coding, but since the landscape is "AI, everywhere" it boils down to holding people's feet to the fire that they are responsible for what the AI slops out. I can't avoid using AI powered stuff, but I can avoid directly using AI (no Claude, chatgpt, etc for me)

I think on the art side, too, we’re going to see AI used by artists - the actual ones who physically (or digitally) paint. They’ll train it on their own styles and teach it to do the work of production. It’s no different, fundamentally, than using procedurally generated items to design a level.

Is using AI to help code like importing libraries these days? (Haven't coded in 20 years). My concern is if coders are relying on AI, will they lose the skills to develop new coding methodologies?

No, it's like writing a requirement document and then having a developer implement it for you. It absolutely does mean you lose the ability to hand code, at least temporarily.