re: the AI discourse
While in general you want a lot of VRAM for anything machine learning, you don't need as much for image generation versus text generation. 6gb VRAM is enough for most of what you want to do with just a little bit of launch option hax, and you only need to start utilizing proper hax when you're really low on VRAM (>=4GB), although more is definitely handy if you want to generate higher resolution images or to more effectively batch generate images.
If your goal for a new GPU is to maximize AI capabilities on a budget, a 3060 12gb, despite being pretty awful for gaming all things considered, is a very solid pick since AI models aren't so much processing power dependent as much as wanting to use as much VRAM as possible in order to shred through the billions of parameters as quickly as possible. Pre-owned models for those usually float around $300 on ebay.
If you want to get really DIY and put together your own cooling solution for it and are also comfortable with lacking general usability, you can get an old Tesla K80 server GPU and its 24gb of VRAM for around $200. You lose out on some optimizations that newer models use (and a lot of speed comparatively since it's approaching 9 years old), but in terms of VRAM/$ it's a lot of value.