From the article

> We also don't have continuous deployment (CD) set up. That's a weird thing for me to say as a positive, because for a very long time that was a hallmark of good software engineering practices... but with me shipping 108 PRs in a week, I actually have been thinking differently.

It’s still a hallmark of good software engineering.

We are past the point where these claims are impressive, anyone can open a shitload of PRs with AI.

It's the context around review, safety, etc that might make your headline interesting, but unfortunately i have no way of seeing that because the website times out.

PR count matters as much/little as lines of code count.

Here's a backup link for this article since it looks like Offprint might be down at the moment: https://brittanyellich.com/a/3mrjj34puva23-108-prs-in-eight-...

> unless you're shipping nuclear safety code, you probably don't need a human reviewer for your PRs")

There’s an enormous spectrum between nuclear safety code, and free consumer-facing software like Bluesky that can be down in some feature at any given point in time and not matter too much.

We have signed contracts with many large enterprise clients, which include assurances about security, data residency, data retention, and SLAs. Having even a part of the system go down or violate the contract is a big deal for us.

The way I would phrase it is “unless you’re working on systems that matter”. It’s not only nuclear safety code that matters.

108 PRs in a week, no mandatory code review or CI.

Power to you friend!

Just make sure to do some phases of slowing down, reviewing and refactoring. The agents code so defensively and add a lot of unnecessary code that we will all have to trawl through when the bubble bursts and codex/cc subs are $2000 a month.

Accidentally discovering how to push slop. Even though AI is super helpful assistant, I think we must read source code before pushing to github.

What is the difference between loop engineering and hyperparameter search?

I mean, sure, LLMs add a new textual dimension to hyperparameter search. You used to write a program to write a program from some parameters (how many layers, hidden dimension, which inputs, how often to drop, what momentum, ... note: in practice limit the choices), and now you just tell an LLM to try them. Then run and collect results. People did this in 1990.

Hell, I've done "compile options search" with this principle. Just a big python array with options, run through them, benchmark. Then decide it's taking to long and find some way to "intelligently" run through them.

Boom, "loop engineering". I hate that everything keeps getting renamed. Give me back my HAAR cascades!

[dead]