Technology

as a consultant/freelancer dev whose entire workload for the past year has been cleaning up AI slop, no with dev it hasn't been what I would say a smooth or even good implementation. for my wallet? been a fantastic implementation, for everyone else? not so much.

The thing is as a TOOL it's great depending on the model. As a rubber duck? fantastic. As something that the majority of companies have utilized with vibe coding to build something end to end? no, it's horrible. It can't scale anything, implements exploits left right and center, and unlike junior devs doesn't learn anything. If you don't hold its hand during a build then it'll quickly go off the rails. It'll implement old APIs or libraries or whatever simply because those things have the most documentation attached to it.

An example. a few weeks ago a client wanted to set up a private git instance with Forgejo. They had Claude Code set it up for them. the problem? Claude went with Forgejo 1.20. ForgeJo is currently on 12.0. MASSIVE security hole right there. Why did Claude do that? 1.20 had more documentation as opposed to 12.0. And when I say "documentation" I could simply be referring to blog posts, articles, whatever that talked about it more than the latest version because The LLM's will leverage that stuff when making decisions for builds. You also see it if you want something in Rust+Smithy. Majority of the time the AI will go for a very outdated version of Smithy because that's what a lot of people talked about at one point. So you're generating massive tech debt before even throwing something into production.

Now like I said as a tool? a problem solver for a function you can't figure out? it's great. the issue is like I said companies aren't seeing it as a tool, they're seeing it as a cost saving replacement for a living human being which it is not. It's like replacing construction worker with a hammer attached to a drone and then wondering why your house frame keeps falling over.