I worked at a fairly big tech co years before the AI boom. People did large scale refractoring across huge code bases back then. With refactoring tools. And properly written robots.

Applying changes to code at scale, opening PRs automatically, basic interaction with human reviewers, making sure tests pass, getting things merged when ready. All that already existed before LLMs. And it was actually reliable and not capable of hallucinating terrible things.

It's like we've forgotten how to automate things without LLMs and openclaw now...

@lina Proper automation has an upfront cost, and I think this is what ppl try to avoid. It's not uncommon that people wait for frontier models to get better at deciding on logical conditions instead of writing two if's...

@lina

Had to sit through an AI booster talk yesterday I couldn't get out of. At one point, the speaker was like "AI can write boilerplate code for you, accessors and modifiers, and other boring stuff, change names..."

I'm like "Bro, emacs has done that for decades"

@emc2 Small local LLMs can do that too. I don't know why nobody is working on those more...

@lina We also had code-generation tools, called compilers, that took input from a highly formal language with its own task-optimised syntax and semantics, and when these so called compilers generated something that didn't exactly match what was specified in the input, we would consider this an enormous problem.

In fact, compiler faults would be so uncommon that it wouldn't even be one of the error causes anyone normally thought of, but sometimes people would joke about it being the cause.

@ahltorp And compilers also had AST parsers that could be used as part of refactoring pipelines!