Programming

Was trying to be a sport and use GPT5 exclusively to write a python script that I could have whacked out in about 2 hours. The task was simply authenticating to an on-premise instance of sharepoint, reading a list of folders and documents recursively, authenticating to a cloud instance of confluence and recreating the structure as confluence pages while converting every docx to a confluence page and generating a link to the other ones.

After 4 hours of correcting and babying it it managed to successfully authenticate and parse the file and folder structure, but didn’t implement any of the conversion and linking logic correctly. I know gpt 5 is not geared towards coding, but it’s the only thing my company has (copilot) and I had to spoon-feed it on details for the auth mechanism and parsing for example so much so that it would have need easier and faster to just do it by hand.

I relate to this article. On the bright(er) side, I looked at the latest release notes for Cursor; they're still adding in features that I:

1. Would call basic
2. Felt the need for after spending just a few days writing my own cli chat/agent (talking to/with a local model run using ollama)

Similarity, I watched some of a recent week-long "vibe-coding game dev stream" (read: Cursor + Omarchy product placement/sponsored ad). Two cursor employees were present for around half a day to basically do some on-the-ground reconnaissance (read: talk to some actual users and get concrete feedback). I was astonished at how they seemed to discover in real time some of the sharp edges and missing functionality in their own product. Do they not use Cursor (or any of their competitors) to develop Cursor? Even if neither employee was a dev, surely they hear from their coworkers and their customers on a regular basis!

So, if these LLM-based agents are supposed to replace us, why are their makers so slow at building them out? Have all the "skilled" and/or experienced developers decided to shun working for them? Have they assumed up till now that most of their users would not be using them to develop software?

I know how hard it is to produce good estimations in software development. It just doesn't seem like the companies working on them actually understand what the grunt, moment-to-moment work of software development consists of.

If that's true, then programmers like this article's author should be fine if they can weather the current LLM craze.