- cross-posted to:
- linux@lemmy.ml
- cross-posted to:
- linux@lemmy.ml
Imagine taking the lightweight, privacy-focused appeal of Linux and completely ruining it by hardwiring AI models straight into the OS. Nobody is switching from Windows for more forced AI features.
The thing is, nobody knows how much effort I put into it versus the AI. There’s just no scenario where I could proudly take credit for using AI. Either I put in tons of work but others have no idea how much work is mine vs not, or I don’t do much work and I’ll know I don’t deserve credit.
Like I understand where you’re trying to come from here but I just can’t see myself winning with AI, against my own morals.
Me collaborating with a human is different. Then the other human and I will know how much work went into it, and at least two real people will share the knowledge of who did what.
How about digging a ditch with a hand shovel vs. using an excavator? About a hundred years ago, you would have hired a bunch of men to chop down a forest, and dig ditches so that the land could be turned into a field. Digging the ditches was hard work, but excavators changed that completely. What used to require 10 strong men and a week, could be done by a single excavator operator in an afternoon. The result is still pretty much the same, even though the bucket of an excavator isn’t quite so precise. You still get the job done, but most of it was done by a machine. Should that one guy be proud of the work? Sure, it didn’t take much time effort, and to him it was probably just afternoon like any other. A few months later, that gig begins to blend in with hundreds of others like it, so it’s not going to be very significant to him. But can you still be proud of your work?
Similar kinds of shifts have happened before that, and the industrial revolution in the textile industry is clearly the most famous example of that. I’m thinking AI in similar terms, even though the details differ. There’s still enough to compare. They both produce trash if you skip quality control. They both reduce the amount of time and effort required. The cheap trash is sold on Temu and dumped in the landfills of wealthy countries. Same process with proper quality control results in all sorts of nice products you would never find on Temu. I think the same logic applies to AI generated things. So far, we’ve seen plenty of Temu grade trash, AKA slop. There’s no reason to be proud of that.
The difference here is that I know as well as everyone else that I couldn’t possibly do what the excavator does with my own two hands, so the skill here becomes operating the vehicle itself. But what the LLM can do, I can do.
This is so often the argument when talking to LLM apologists/proponents, that “this is a shift just like any other before it”, but it isn’t. It sneakily resembles one, but it is not exactly like one. There are major differences that set it apart from previous shifts. This isn’t just a “paradigm shift”. We are no longer thinking of solutions on our own, so we are/will be losing the ability to do so, yet we are expected to be the judges as to whether the generated solutions are good or not. That doesn’t add up.
Once a few generations of “developers” have grown up with LLMs, it’ll eventually be like using a calculator to solve divisions with large numbers, yet still you have to check your work. But if you keep using the calculator, or you never even trained yourself to do the divisions by hand in the first place, you won’t be able to check the solution is correct.
I can hold a shovel and dig a hole. Not as fast as an excavator, but I’ll get the job done sooner or later. I’m sure you can as well. If you want results faster, you can get a bunch of people or a single excavator. I’ve seen similar dynamics with programming too.
Yeah, totally agree with the rest though. Being qualified to evaluate the output is the main bottleneck with LLMs. Most users aren’t qualified, so they will approve all sorts of stupid nonsense while thinking it’s brilliant. If people constantly rely on LLMs to do the thinking, those skills will begin to disappear, which obviously will have serious consequences.
An excavator can also lift rocks that I can’t. But that’s really beside the point.
It doesn’t really matter which analogy you use, the fact is that I can do what an LLM can do, in my field. I can probably do a better job, because I know more current web APIs than an LLM has had time to train on because of limited human output.
You don’t get faster results with more developers though. You don’t make a baby in one month with 9 mothers either. You might have seen a small increase in development time with more than one dev on a single project but that doesn’t scale past more than a handful of devs, usually. At least not per feature.
Actually, that’s a good point. Proper programmers are impossible to replace. LLMs have the greatest impact for people who can only write bad code slowly or not at all. They will see a huge boost in output. However, they are also least qualified to take care of quality control, so the output ends up being 100% slop code. People like that will end up slopping together software projects that are impossible to maintain.
And it’s super worrying, because everyone I know and almost every colleague I have is suckling on the LLM teets right now, but they don’t realize that their enthusiasm affects the youth that we will have to rely on to maintain our projects once we retire. They’re gonna be so bad at it.
Adopting and maintaining software is hard enough as a good programmer. 😅😅
The current generation is throwing a massive slop grenade at the next. Here, catch!
Technically, you can use an LLM to learn what code does, how a function works, why some convoluted workarounds are necessary, etc. If you really take the time to do that, you should also add lots of relevant comments to the code, something the LLM certainly won’t do. When you look at the code two weeks later, you’ll still be able to debug it on your own or tell an LLM to carry out very specific tasks for you. Either way, this route takes time and effort, which is why most won’t do it. The temptation to skip vital steps is too great. That’s how we’re currently driving to a software maintenance dead end.
Sooner or later, the code is such a mess that you can’t fix it by throwing more AI at it. Humans have to understand exactly what the code does and why it works that way. When people start dealing with these slop grenades every month, they realise how important it is to have proper guidelines about using LLMs responsibly. It’s about time someone starts teaching these guidelines that haven’t been written yet
Preach! 🙌