Dunno, man. I'm using Claude 3.5 via Cody on IntelliJ for work, or via Kagi for other cases. When it works, it works great - especially for easy, well-understoo

Dunno, man. I'm using Claude 3.5 via Cody on IntelliJ for work, or via Kagi for other cases. When it works, it works great - especially for easy, well-understood things like boilerplate code or the typical simple tasks that I don't bother memorizing because they were already available in search engines anyway. For anything more complex or esoteric, LLMs shit the bed pretty quickly. A recent example: I needed to implement a particular setup for an Argo CD ApplicationSet (default values plus a per-stage override, at the generator level). Every single suggestion by the LLM was wrong. I decided to dig into the documentation because I assumed there was no way this wasn't possible, and lo and behold - matrix generators are a thing, and my exact use case was an example in the doc. I asked the LLM about it and it knew what they were, so it clearly had indexed it before. It's far from the only issue - the most common one is the below-average quality of code in some cases, with plenty of wrong implementations strictly due to hallucinations. Then Cody has its own issues on top of that - not applying changes correctly is the most common one. Put that together with engineers that are in a rush or are using the LLM as a crutch... and I can see a future where devs with little experience can put together huge codebases that will be as brittle as a house of cards.