I thought that that’s the entire point of Redox? Not doing that would defeat the basic idea behind that system.
The main function of a LLM is to generate text that sounds like natural language. The information the text contains is really just coincidental and mainly revolves around a “it sounds good”, not trying to ensure that the final code is correct. This makes LLMs great for keeping people or scrapers busy, but dangerous as LLM-generated code is hard to review. LLMs are good at creating noise and wasting time of reviewers (or in other contexts, make finding information about something happening in the world almost impossible). They are a danger to any kind of information flow, flooding the zone with slop. LLMs (and image generation models) don’t create knowledge, they destroy it by generating a plethora of articles and “documentation” that is at best half-true and at worst complete bs. “AI” (Generative models, would be the correct term) is not the symptom but a weapon against knowledge and I don’t think we should support that.
I have a quite hopeful view of Haiku as a well designed desktop system that is well designed and has diligently written software. Of course you could take a shortcut by using a slop generator, but I’d doubt that the end result would be all that well usable. I don’t think quickly “filling up the gaps” with LLM-generated content is a great idea at all. Both Haiku and the application landscape already have made some huge steps since beta2 imo (That was the time when i started using Haiku again after first having had a look at it in 2013 when I actually still was a child).