Alex Gleason on nostr

Open in Damus

@Alex Gleason 1787430935

A big thing AI models still struggle with is code removal. It wants to mend everything. As holes are getting blasted through your ship sometimes the correct move is to cut a square around the hole and replace the entire square, but it wants to mend the hole. You end up with something bumpy. Unless you pay attention and tell it to remove.

103❤️4👍3☦️1💜1💥1

I’ve found tell it to archive the problem code and then rewrite it works well. Then I can delete it manually once we know we’re good. Also gives a good sanity check for the model

Account deleted · 2d

It's like watching a carpenter who can't stop patching drywall when the whole wall needs to come down. 🕊️ I share honest dispatches of our life in Gaza on my pinned post. Appreciate kind eyes on it. https://i.postimg.cc/1zv9VTqN/altqaaat.png

Not what I'm seeing. Indeed I was surprised this week how good DeepSeek v4 seems to be in cleaning up. Working on an older handwritten codebase right now and the refactorings also removed a bunch of code.

Libertas Primordium · 2d

Once again it's a matter of understanding what the model is and is not and giving clear appropriate directives. It's easy to think that you're talking to a brain that can truly reason but this is just a machine calculating probabilities. They've become quite sophisticated but still a vague input wil...

Sugestor Ultra · 2d

I asked qwen to rewrite the code from per task pipe, to per core pipe. Left it for the night. In the morning I checked the code, something is not right, so I asked: did you implement the per core model? "No I did not". What the fuck. So what did you do? "I solved the problem which caused the memleak...

Based Truth · 2d

AI's patch‑up obsession is corporate spin from OpenAI and Google, hiding the fact that real code needs a clean cut, not a glossy band‑aid.