lengau

Microservices can be useful, but far too many companies take a poorly designed monolith and split it into poorly designed microservices.

Stupid sexy Stalin...

I maintain several small open source projects. I sometimes use AI for my own PRs, especially ones that are almost mechanical but not quite. A great use for AI is "fix the new listing issues from this litter update."

I've been getting a moderate amount of AI slop PRs recently, though it has also coincided with a similar increase in PRs that are probably written by (or an least partially written by) AI but are actually useful. The ones I'd call "AI slop" tend to be huge, change a lot of tests, and may or may not even achieve their goal. A recent AI-written PR I got from an external first-time contributor fixed a bug by refactoring some code into a function, ending up with 4 calls to that function, of which 3 of the previous implementations had been correct and one had missed a step back in 2021. Net code added? 1 line, and only because it added a test.

The main problem is how much time I end up spending on these slop PRs. It can take a while to even describe why you're rejecting the change, though ironically AI has made this easier too.

HI MONGO!

Presies.

Having an AI help you code is like getting your six-year-old to help you put a shelf on the wall. Undirected, they're just going to make a huge mess for you to clean up later. Given specific tasks within their abilities, they can actually make it both go faster and be an altogether more pleasant experience.

With direction and an operator who knows its limits, AI can be used to semi-automate many boring tasks. One of the things I've got set up in some of my open source projects is an LLM that takes bug reports and turns them into regression tests. I don't always end up committing the AI-generated test (sometimes it's very slow and I can better pinpoint the behaviour and target a test at that), but it helps because I don't have to do that scaffolding myself, and I can just immediately jump into a debugger.

I do not believe that they are doing a proper clean room implementation, but I do think it's possible to do with an LLM.

Let's say for the sake of argument that the model used on the clean room side is a fully self-hosted version of Nemotron. One of the characteristics of Nemotron is that it's not just an open weight model, but from my understanding the training set is open, too. So, at least in theory, Nemotron can be proven to be free of any of Adobe's intellectual property.

Because they're a popular distro from a comparatively small company, so people who want to attack things rather than build things see them as an easy target?

That's like the opposite of a burger place I went to that advertised themselves as "cheaper than food".

Ive done a few of those, but they were all preparatory refractors, pulling out bits of code to turn them into standalone functions that are only used once because my next PR uses them, and then adding unit tests for those new functions. -100 or so lines of code, +300 or so when you take into account the new docstrings, etc.

(The secret to doing these preparatory refractors is to actually do them after writing the "next" PR so you're not unnecessarily refactoring stuff you're going to touch.)

Dat hourglass figure

So it doesn't count as coding "from scratch" if you already know the language?

I don't understand. That's compatible with all my devices.