Discussion
Everyone hates AI, but do you even know how a coding agent actually works? It's hilarious:
LLMs mess up indentation, constantly mix tabs and spaces, get line endings wrong, hallucinate diff line numbers and produce completely broken patches. So the harness throws increasingly desperate fuzzy matching at the problem and blindly tries to apply whatever sticks.
This is software engineering now. Apparently.
@fribbledom For a long time I assumed that LLMs would be better at coding in languages where the most examples are scrapable. That would make decades of published C/C++ code favoured over newer languages such as Rust.
It turns out it works differently. Rust is apparently popular among Slopshippers because its compiler is very good at telling you what went wrong. That makes it very useful for LLMs, just throw code variations at it until it compiles.
@fribbledom Sounds like the harness could be a useful tool if it can make sometimes-functional code out of the nonsense.
@fribbledom
Working with an LLM is like working with an intern. Nobody let's an intern push code to production.
@fribbledom It’s funny when the AI doesn’t realize that comments can’t include em dashes and it trips over it constantly. It won’t figure out how to correct, so it’s like a Roomba with a mouth full of electrical cord.
@fribbledom "This is software engineering now. Apparently."
I'd rather argue that the long tedious debate on whether software engineering is "real" engineering is finally settled. It is not.
@fribbledom For decades, I wished developers would understand how wrong and ridiculous machine translation is. I guess I wished too hard? 😅
@fribbledom
First, the premise isn't factually right. Not *everyone* hates AI.
Second, you got to change the LLMs and harness, or at least manage context and instruction.
Vibe Coding ≠ Spec-Driven Development ≠ AI-assisted coding
@fribbledom Yeah, agentic coding probably came out of an LLM company brainstorm about how to get people to spend more tokens. “Wrap the chatbot in a for-loop and call it an agent!”
@fribbledom Truly, the "enough monkeys on typewriters" design ethos of our era.
@fribbledom@mastodon.social also they keep throwing that broken code at the wall over and over until it compiles, that's why sloppers are flocking to languages like rust 
@fribbledom@mastodon.social @jay0@alico.nexus
>
keep throwing that broken code at the wall over and over untił it compiles
that's how I write C++ template and constexpr code by hand too tbh 😁
@jcm @fribbledom @jay0 But only for of fraction of the cost, energy and resources of an LLM! No no, that can't be serious, numbers must go high.
@fribbledom I honestly dont know how to code! I know a bit a bout basic logic and math. The people I follow? My family members? They know much more! And they tell me that this thing we use to weave together so much?
It can't even reliably tie its shoes when you pose it dynamic questions that require it to WORK. It just lies and hopes it's correct like some fucked up manager
@fribbledom There's also buzz over on LinkedIn about reducing test coverage to "improve the development workflow".
I mean, sure, I've seen over-testing killing productivity and morale but I don't think they're approaching this from the human standpoint.