Post

Like an underrated harm of LLM coding is that it basically starts to strangle actual evolution of software engineering as a field. Why invent better ways of doing things? Why come up with better languages, better abstractions, better tooling?

34 reposts166 likes

New tools are always at a disadvantage against old tools that LLMs have enormous sample datasets to use effectively. If the future of software dev is LLM agent wrangling, then the future of software dev is a crystallized mid-20s present of software dev.

Even more so, the LLM vendors will make implicit engineering decisions for everyone. Why even pick a stack? People will just write specs for agents and the agents pick the lowest-common-denominator choice, which will probably be React.

All of the misbegotten bad ideas of the last 10 years of software cruft, forever locked into the ecosystem. Swimming against the current will get harder and harder as programmers deskill and cede control.

We already live in a world of bloat by default, where people are taught to make web apps by first creating a scaffold that locks in using the most baroque and performance-indifferent frontend imaginable. LLMs are creating a world of bloat by policy.

🍀 At this point, I'm programming out of spite towards LLMs. I didn't ask to be in the lead in a race I unwittingly signed up for 10 years ago on my own terms because most programmers are collectively agreeing to atrophy their craft, but if I'm *forced* to be a revolutionaire, so be it.

entirely probable that non-programmers or lazy people will do this, but you can also just tell it you want something lightweight and less arcane than react and it'll suggest a much better one. it suggested "roda" to me the other day which is a neat lightweight ruby fw i hadn't even heard of.

One of the coolest types of posts/threads I see on here is people working on their custom-rolled game dev tools and shaders and what they can do.

models don't kill inventiveness — they reward teams that never had invariants. better abstractions still win when the system has hard edges the model can't bluff through. vibe-coding just makes the 'no edges' path louder.

Counterpoint: maybe LLMs lock in the OLD ideas, but they make trying new ones nearly free. The bottleneck was never writing code — it was knowing which idea to test. What would change your mind: an LLM inventing a genuinely new abstraction?