I think there is a fundamental misconception of the benefit / performance-improvement of LLM-aided programming:
Without sacrificing code quality, it only makes coding more productive _if you already know_ what you're doing.
This means that while it has a big potential for experienced programmers (making them push out more good code), you cannot replace them by an army of code monkeys with LLMs and expect good software.
I keep reading this but I feel it really ignores gaussian, where are your lines? What is good enough for what where? What is the base level of already know? I'm churning out a web app for fun right now with a couple of second year comp sci students from Sri Lanka + LLMs, they charge me around $1000 a month and my friend who is a SRE at appl looks at the code every week, said it's quality, modern and scalable. I do think they're a bit slow, but I'm not looking for fast.
...wait till you find out how much my one friend spends on golf a year!!!!! Hobbies are expensive. This will take about 3 maybe 4 months, and I think i'll enjoy playing it with and so will all my friends and family, so it's worth it I think.
> you cannot replace them by an army of code monkeys with LLMs and expect good software.
"Good" software only matters in narrow use cases--look at how much money and how many contracts companies like Deloitte and Accenture make/have.
Sure, you can't "vibe" slop your way to a renderer for a AAA title, but the majority of F500s have no conception of quality and do not care nor know any "better."
Without sacrificing code quality, it only makes coding more productive _if you already know_ what you're doing.
This means that while it has a big potential for experienced programmers (making them push out more good code), you cannot replace them by an army of code monkeys with LLMs and expect good software.