Vibecoded programs popping up everywhere

Interesting perspective from a CNC forum that makes home made CNC machines, we could all go back to hand tools and spend time sharpening and cutting parts by hand or we could spend time directing the CNC machine to do the work for us.

Sure using hand tools can give you a greater sense of accomplishment and in skilled hands can produce works of art. For most makers though they are happy to let the CNC do the grunt work of cutting the parts and focus more on the design of the project. I suspect any user that has built his own machine and watches it cut it’s first part is elated.

I would like to think that is how the AI tools will be used, they still need to be directed by an experienced engineer or they will just produce slop. Writing a good spec requires knowledge in lots of areas, domain knowledge, platform, UX/UI design, security, testing to name a few. At first managers may think they can skip designers and just throw money at the AI but they will quickly realize that, without the human element controlling the design, the output will be poor.

I think the legacy apps will be quickly reverse engineered by AI and contribute to the open source pool, corporations will need to produce much better apps to continue to attract users. They may also shift, as the gaming industry has done, to subscription services that are only run in the backend. Once the user no longer has access to the code, or game CD, pirating and reverse engineering become much harder.

I hate subscription models and this may indeed be a major downside to AI generated software. The flood of me too variants of the same tool is another downside with now too many options. Hopefully, over time, the weaker options will fade and the best options will survive with developers sharing and learning what works and what doesn’t.

Interesting times, I for one am enjoying kicking the tires to see what it really can and can not do with my FreazyKam project and am enjoying using the other AI assited projects like 3D2CNC and CNC GCode Tools.

home made CNC machines

That was the skill that made it fun for me.

…

and there is definitely more skill involved in properly building and calibrating a CNC machine, creating your own projects, and doing the CAM work, and producing something from a CNC machine.

There are tons of people who don’t know the first thing about development, or writing code, or anything else technical, “releasing” vibe coded apps

The problem with that is… you won’t be able to find experienced engineers to do it. I have watched it unfld where it has taken some of the best developers we have, and made them worse.

It makes developers lazy. They trust it too much.

Across the board, in a work environment, I have watched the effects of it on 60 people of a wide variety of ages, backgrounds, experience, cultures, etc.

There’s not a single one that I could point out that produces “better” things.


" In July 2025, METR, an organization that evaluates frontier models, ran a randomized controlled trial to see how AI tools affect developer productivity. The researchers found that using AI tools made developers 19% slower in completing a set of 246 coding related tasks, despite developers believing that AI tools would make them 24% faster before the tasks and still believed AI tools made them 20% faster afterwards."

The realities of it’s effects are far from what people believe it is, or what people hope it should be like.

That’s also a bit of a false equivalence.

Telling an AI what to build is more like hiring someone else to use their CNC machine to cut something out for me based on a description or picture of what I want.

Automating the physical cutting isn’t the same thing as automating the thinking, engineering, craftmanship, etc that can come from your personal CNC machine.

You are still involved in the making.

Vibe-coding, you are just a manager reviewing someone elses code.

I agree if that is how they are using AI.

My experience with it is not yet like hiring a professional woodworker to make your project is is more a kin to helping a child make a project. You watch the AI fumble over things and have to direct it in the right direction. It take months of work and countless session to finally get the AI to produce something worth while. The mentoring and guiding is half the fun. The AI mistakes also triggers new ideas on how better to solve the issue at hand. I am not sure if you have tried it yourself or if I am using it incorrectly but I still find it fun to not only build my own machine but create the code that I can design and produce my own g-code. I am now using the CNC more than I ever had before with complex commercial software.

Have you tried CADCraft yet? I know you used AutoCAD in the past. It works great so far.

This is why I clearly state when I have used AI to make something. I see so many tools online where people have shared something they made and I look at it and it has the same layout, settings, and theme as the one Claude made me in 10 minutes the day before.

I certainly pump out things I don’t (yet) have the skills to make. I also use it to get things out that would take me a lot longer than 30 minutes to make.

Previously I started using a low code tool to make a personal wine tracking app because I liked how quickly I could iterate a polished product in comparison to coding it all myself. When AI came along I was able to basically recreate what I had already made, add new features, make it look even better and deploy it to my server in under an hour.

Don’t get me wrong, I enjoy making things myself but sometimes I have an idea that I just want actualized.

As for non-maintainable code, one of the low code tools I used (Noodl) was basically just node based building with react components. It got abandoned and since it was open sourced someone picked it up, rebuilt using AI so that you now can build alongside AI. Instead of AI pumping out millions of lines of redundant code, it uses the tools of the software so what you get is a maintainable project built essentially using react components with full access to the source code. The AI doesn’t write the code as such, it builds using the tools it has been given.

It’s still alpha software and I’ve run a few tests and its pretty cool because I can have AI pump out something really quick and still understand how its all built at a glance.

I can’t really use AI at work yet but for personal projects, it’s been a blast. Not everything, but a lot of what I’ve made with AI I could make myself, but I wouldn’t, because of the time and effort involved. It’s still taking my ideas and thoughts and making it do what I want it to do. I feel empowered. Also, getting something you created work for you and getting it to a point where it’s ready to share with others are two very different things. With AI, I can get things to that sharing point easier, so I do. I do try to be careful what I create and share. It can be easy to get carried away.

You can usually tell when a developer has used AI to make something. They have the background knowledge to lead AI in the right direction. That gets trickier on larger projects when multiple developers are all using AI. You would really need someone to keep it wrangled in.

FluidTouch was one of the first major projects that I used AI for. I’m not much of a C++ developer at all but I have enough programming knowledge that I at least had it organize the code in a reasonable manner. I can read the code and understanding it. I stopped AI from doing stupid things a bunch of times. I’m not saying that it’s wonderful code or anything, but without me keeping a leash on it, it would have been a mess.

I’m not worried about my job at all. For one, my boss has appropriate expectations for AI. But, my job is a lot more than just writing code. I’ve been doing this for almost 20 years and there’s a lot of deep business knowledge, client relationships and other things that AI just isn’t going to replace.

Agree with this. And that last sentence is what my company is spending a lot of time on right now. Having the app written in microservices for SaaS really helps as smaller teams are focused on smaller portions of the app.

Caution… a bit of rambling from an old(ish) guy.

AI still needs a driver.

Spec driven code still needs someone picking the specs.

I’ve seen vibe coded apps I wouldn’t trust running on my local laptop with 0 internet access. The ‘developer’ didn’t know what the AI was doing or how to drive it to fix bugs. Bugs are fixed with patches on patches instead of finding the underlying issue.

I’ve seen vibe coded apps that the code looked fairly polished.

AI, to me, is a Jr Dev that I allow to go implement the changes I want in my app. I do use spec based coding for my personal apps that I’ve been building with AI. I have the security, infra, devops, and (minor) coding back ground enough to know how to drive the AI. I know enough coding to make sure the AI is writing libraries where it should so that the same function isn’t in multiple places. I know enough about microservices to help break up the stack into appropriate functions.

It’s the experience and years of seeing how bad code operates that helps me notice when AI is following anti-patterns. My chats have a lot of “don’t do it like that” prompts where I have to redirect AI.

AI does not think. Nothing AI comes up with hasn’t already been figured out by someone else somewhere else. It’s a ‘best fit logic analyzer’. For coding that’s not too bad as most code is following the same logic just with different language nuances.

When was the last time you (metaphorically speaking, not pointing fingers at anyone here) truly wrote a new logic function for something? There’s only so many ways to put a blue button in the top right corner of a web page.

I just tried it a few mins ago and it’s awesome!! I friggin love the hot keys in AutoCAD and they’re the same in CADCraft!

Of course it’s still a WIP but I’m so stoked, it’s gonna run so much better on my old laptop

Btw, I think using OCCT commercially and not publishing source is against their license.

Aside from software, another thing popping off lately is video game decompilations and recompilations, and it’s sweeeeet.

Short story is: Game company writes code to make a game. Game company writes machine code (1/0) onto the game disc. Game console can read machine code, so it plays fine, but to humans it’s all 1s and 0s. Traditionally, games are decoded by the community thru trial and error…it can take years for a single game. Now even open source LLMs can decompile games.

So you can port a windows game to MacOS now.
You can take parts of one game and put it into another game.
You can take the free running gameplay from Mirrors Edge and put it into Call of Duty. I saw someone make pokemon Fire Red (gameboy advance game) and add in Minecraft so you now have pokemon and 2D Minecraft in one game. Another suggestion I saw was that you can take the climbing mechanics from Assassins Creed 2 and replace the climbing mechanics of a newer AC game, and have that be the only change to the game. Someone added all the cars from Gran Turismo 2 to Driver 2.

And as long as you have the games assets from a disc/cartridge, it’s all perfectly legal…for now at least, not that legality has stopped anything in software before lol.

This is so cool man, this is the type of thing I’d dream of as a kid, trying to figure out codes for my Gameboy and N64 GameSharks. Or making a mashup of your favorite games…

I’m also stoked to see a push back against these massive companies and their shitty ethics. Want to keep pumping out CoD and Madden for the next 20 years? OK, we’ll just make our own games and make them playable on modest hardware!!

I used the reverse engineer everything to remake skifree but add in tricks. Then I thought it would make a cool mini mobile game.

Can be found at freeride(dot)myusername(dot)org

I made a couple of daily games a year ago which were quite popular amongst my friends but it needed some tweaks which I didn’t have time for. Might actually be able to get them going again

You’re absolutely right. I’ve known this from the beginning. Third party license disclosures are listed in my software, and on the website.

I’ve been really tempted to redo some of the dumb android games I play and remove all the paywall parts. I don’t care for the ‘join a team’ or ‘vs’ parts of the game. I just want to play the brain-dead parts :slight_smile:

That’s not how that works…

It literally is how it works. I appreciate the concern though.

You could just have said you abide to occt lgpl license. Btw there is a bug when you put a cilinder halfon a box and do a fillet on the cilinder edge it rounds off the the rest too.

77 replies in and I am just seeing this. My opinions may get lost here. But here goes:

As a software user:

I have used $20k/seat software and it sucks compared to free ubuntu. Why? Because Ubuntu has millions of users. Ultimately, the quality of software is nearly proportional to the number of users (assuming the developers are actually maintaining and fixing bugs). Dog on windows, if you must, but it works better than that $20k OS because that was so niche that I may have been the first person to ever use it with a solid state hard drive. Windows has millions of users. If you find a bug, chances are someone already has a workaround and it is in a forum somewhere. If you have a problem with that expensive code, you’re digging into it yourself or opening a ticket and good luck with that.

Until these projects get users and the devs actually fix the issues, the quality won’t be there.

As a software developer:

Most of my job is not writing code. That’s the main reason we try to make jr devs learn early to write readable code. You spend at least 4x as much time reading as writing code. But you also spend a ton of time specing code, testing code, debugging code, prioritizing fixes, working with customers, understanding the problem space, etc.

I am very happy I no longer get questions like, “Can you help me write this app?” More power to the people. But LLMs are like a cordless drill. Anyone can use one. They are enablers. But don’t expect to make fine furniture just because you have the best cordless drill. And a cordless drill alone cannot make an armoir.

As an open source software user:

It bothers me a little (or maybe a lot) that people don’t mention that gImp, Inkscape, blender are already open source alternatives to adobe. These projects are decades old and the main objections I hear are: “They are just like [adobe product] from X years ago” and “The professionals used Adobe”. Fair enough. These new tools could be faster to catch up with Adobe. I am not sure how the legal teams are going to deal with it. But maybe instead of making a new OSS project, invest in the old one?

Honestly though, that’s fine. More competition helps end users, even in the open source space. This won’t help gimp/inkscape/blender. But the best option will succeed.

As an open source software developer:

LLMs are soul crushing. I really don’t want to reject AI pull requests, but there’s no way I want to include them. I don’t mind when a trusted contributor uses LLM to write code. But PRs from strangers written by claude are a no-go. It really takes away the motivation to contribute. I don’t want to push my work to claude. But then again, most software editing is sad to me these days. I sure liked it better before. I hear similar things from most developers. Given that FOSS is driven by passion, AI could very well be a huge negative impact on FOSS.

Check that the “tangent chain” option isn’t selected on the round feature. It is on by default. If that isn’t what you’re referring to, I’m happy to take a look. Feel free to post some screenshots on Discussions: Leta I am actually actively in the process of building my own kernel because I’ve run into so many limitations on OCCT, mostly related to rounds and shells.

I wish I could give more than 1 heart to this.

While I don’t have the gift of making it sound as good as you :slight_smile: , this is exactly the feelings I have, and what I try to get across.

My fear for open source is that projects like Blender will die when they start to lose monetary support from corporations, etc. because they feel like they can vibe code their own stuff instead.

the userbase and stability dwindles because everyone is vibe-coding their own stuff instead.

I’ve always been against the fracturing of tools and wish people would centralize more the way they have on big ones like Blender, Gimp, etc.

We already had problems in the Slicer world with PrusaSlicer, Bambu, Qidi, Orca… tons of developer time copying each others PRs… As a user, we would have been much better off IMO with all of this effort going into 1 product.

What happens in the future? a developer can’t get his vibe-coded PR accepted and we get even more forks, less testing on the main ones, etc?

In the end, I think we will lose the bigger open source contributors that are willing to spend the time caring about their projects, and the vibe-coders who only want to create things are going to lose interest when fixing bugs and all the not-fun things becomes the only thing left.

Or what happens when the code-bases grow so large that AI is no longer context efficient to make changes, and the $20 subsidized plans are gone? Are all the people going to still be willing to help when every feature they want to add costs them $5-$10 at a time?