[codingwagie]

My personal opinion is most software will be easy to write by a junior programmer. There will still be high paid roles, but this will be a low paid profession. Web development, UI, etc will be heavily automated with LLMs. You will guide the model in building the software. It's already this way using cursor.sh, most people just havent realized it yet.

[ilkhan4]

I think it will be the opposite, actually. AI seems to be eating the bottom, but there is still demand for higher level decision making that seniors or architects do. The stuff that I would usually delegate to junior devs, e.g. "move this button" or "change this column" are the things it can easily do now and in the near term. But questions like "How should I design this system" are not and I don't see that happening for a while. People who are good at that will still be in demand since a good number of businesses still want a person they can talk to an make evaluations with.

[np-]

If juniors are being cut out and seniors are being increasingly relied upon to do more and more, that seems to be unsustainable in the long term

[danenania]

The flip side of this is that AI tools can enable much faster learning, so the time to go from junior to senior can potentially become much shorter for people who are motivated.

[solarmist]

It could (and ultimately I agree it's probably what they should), but that's not what people are seeing juniors do. They're seeing them blindly trust the AI tools and think they don't have to learn what they're doing. So they're building at best a fragle understanding of the code their writing.

[danenania]

Since the results of that approach are predictably bad, I think it will sort itself out. A junior who works like that is already less productive than an LLM in the hands of a skilled engineer, so there isn't really any reason to keep that junior around. But a junior who is using LLMs to rapidly improve their skills can be seen as a good investment even if it takes them some time to become a net positive on the team.

[solarmist]

I agree, but I think it'll take far longer to happen than people expect. Kids are making it all the way through college (at state or regional universities) without needing to adjust this approach at all.

These kids are going to be convinced that it's a valid strategy. I'd be surprised if many didn't claim they were being unfaily discriminated against in their first couple of jobs instead of fixing the behavior.

[3D39739091]

If anything, AI tools are doing the opposite. You don't acquire those skills by outsourcing your thinking to an LLM.

[danenania]

It depends how you use them. If your goal is learning and understanding, you will use LLMs much differently than if your goal is to produce copy-pastable code with as little effort as possible.

Honestly, I do a bit of both depending on the situation. Sometimes I just want a one-off script that does xyz and I don't really care enough to have a deep understanding of every detail. Other times I'm getting a crash course in some advanced concept and in an hour I have knowledge that previously only someone with years of experience in a niche domain (and a lot of patience for my dumb questions) could have given me.

[karmajunkie]

welcome to capitalism…

[ilaksh]

The SOTA models can absolutely create system designs very effectively. Have you even tried that?

Like literally, let's get someone to randomly propose a few specs. Then we get some developers to try to create a system design. Some have Claude 3.5 Sonnet or o1-preview to help them work on it. Some of them have you to help. Each group has one hour.

Then we have another group that evaluates the system designs on some number of objective criteria. Feasibility, completeness, robustness, etc.

I'm willing to bet that a significant percentage will prefer the system designs made with the AI as an aid rather than your help. Because they will be better.

That's today. The LLMs continue to get better every few months.

It's true that we still can use human engineers for "awhile". But that is probably only a few more years.

[ilkhan4]

I have not tried that, no. I'm not even sure what a SOTA model is (which is my other issue with AI right now: the fragmentation). Sure, the AI design can be done quicker and may be "better" by some objective measure. However, can it present its model to upper management in a coherent way to justify the budget it is about to be given? Can it translate this design into a series of tasks for (surely AI) engineers to implement? Can it deal with late-breaking changes to the design based on new requirements? You'll tell me it probably can, or shortly will be able to. The thing is, as long as humans are making the ultimate decisions, they'll probably want another human in the loop somehow, at least within this generation.

I mean, look, there's no one who wants to stop having to write code more than me. Our industry is full of make-work and pointless drudgery for no reason and I'd love to see that be made pointless by AI. But there's the old quote about overestimating the short term progress and underestimating the long term progress that applies here. Your version of "awhile" is probably pretty similar to the "awhile" of people excited about past technologies. As another comment pointed out, we engineers have been trying to automate ourselves out of a job for quite a while now and I don't see this being much different. I'm old enough to remember the 4GL fad and people were just as breathless then that you didn't need those pesky engineers anymore and soon you'd be able to do similar things as AI is promising. Low-code was a thing recently too. These things gradually chip away at the skills you need to make things happen, but something has to ultimately be responsible for the results and that's where humans come in.

[karmajunkie]

this is exactly right. the problem is going to be getting people from junior to senior.

[throwaway123198]

Honestly, even in practice the number of 'senior' people are laughably high. Tbh, swe leans junior.