I never use the $20 plan but I access everything via API and i spend a couple of dollars per month.
Although lately I have a home server that can do llama 3.1 8b uncensored and that actually works amazingly well.
It works ok but with a large context it can still run out of memory and also gets a lot slower. With small context it's super snappy and surprisingly good. What it is bad at are facts/knowledge but this is not something a LLM is meant to do anyway. OpenWebUI has really good search engine integration which makes it work like perplexity does. That's a better option for knowledge usecases.
The only thing I see is that it hallucinates a lot when you ask it for knowledge. Which makes sense because 8B is just not a lot to keep detailed information around. But the ability to recite training knowledge is really a misuse of LLMs and only a peculiar side-effect. I combine it with google searches (though OpenWebUI and SearXNG) and it works amazingly well then.
[1] https://www.nytimes.com/2024/09/27/technology/openai-chatgpt...
Maybe we can have a service that's an LLM with shared/cached responses as well as training on its own questions/answers for the easy stuff. Currently my usage is largely as a better search engine as google has gotten so much worse than it ever was, though I imagine I'll find custom things it's useful for as well.
At best, theres a slow march to incremental improvements that look exactly like how human culture developed knowledge.
And all the downsides will remain, the same way people, despite hundreds of good sourxes of info still prefer garbage.
The long-term problem may be access to quality/human-created training data. Especially if the ones that control that data have AI plans of their own. Even then I could see OpenAI providing service to many of them rather than each of them creating their own models.