计算机程序员

Some of the best programmers in the world are working daily to help improve ChatGTP and other models. The capabilities are quickly increasing and given some time they will be better than most programmers. So yes some very good programmers will have a job, but most will not be needed. You wrote the comment 2 months ago, there is now a new model which you thought was not possible 2 months ago. What will happen in another 2 months, or 2 years?

We don't know the trajectory. We only know that for now, progress has plateaued.

But though AI (= LLMs) has gotten better in the sense of creating more complex outputs, it otherwise suffers from exactly the same problems as early versions: no compositionality, no continual learning, no consistency, and no self-correction.

If you ask it to fix a certain bug fix in a complex codebase, you have a high chance that it also starts to change something completely unrelated. It doesn't really grasp how precisely elements make up the complex whole; instead, it applies pattern-matching, by which it gets misled. You can waste days with AI on what would be a five-minute manual fix.

Also, junior devs' job isn't to churn out mediocre code for generic, long-solved tasks or create the 1000th to-do list app. Instead, they get onboarded and are supposed to familiarize themselves with your codebase to improve it. But you know that "PT" in ChatGPT stands for "pre-trained?" So AI doesn't learn anything new. AI is like suffering from anterograde amnesia.

Anyway, neither apps nor code are like a commodity where "more is better". In the end, the point is to create something new. Otherwise there is not much point: just use an already existing and tested library. And it's exactly the "new" part where AI sucks because it struggles to transcend its training data.

So no, AI is incapable of replacing junior devs.

It would have a lot of immediate positive effects if AI worked like you describe: it would insanely empower open-source projects (often lacking maintainers) to fix all their long lists of bugs or introduce new features. But this is not what we observe at all. Because AI is overhyped and underdelivers.

Model Context Protocol helps give AI better context. AI can work across many files now (way better IDE integration has happened in the past 6 months). The limits of AI have dwindled and will continue to dwindle as time goes on. Vibe coding is now a thing and AI can follow directions to build complete mvp's in under an hour.

My worry is if future code-generating models can modify itself or the programs it works on towards a specific goal, sort of like an intelligent self-modifying quine.

Currently. What about in 10 years though

Getting better every week. MCP. Augment Code. LLM rules.

You can convert from one language to another with AI and a developer is not protected by knowing Java vs C#. This further erodes job protection. Recreating the wheel will drastically stop in programming and I'd imagine known problems will converge to one solution.

To be honest, it's quite scary that most people think this job will be entirely replaced. If AI ever fully becomes "human enough" (as you said! That's a brilliant way to put it) to take over every single Computer Programmer job... well, I don't think we have to worry about whether CS Majors will get employed.

How does this relate to programmers? Programmers program not write prompts. Evaluation to code is already possible

But if AI becomes as good as humans then why programmers would be ever needed if not those who work on AI itself? Maybe for some complex and big programms people would still be needed.

No it is not. LLM's don't replicate how human brains work at all.