In this thread, I propose to look at the insides of language models as a collective net comprehension of all humanity. In fact, the pre-training dataset of models includes a lot of textual information from all over the net. I suggest to consider only models with open, untested and unfiltered pretrain (biased pretrain), and/or without tested and filtered instruct tune (preferably without it at all, basic model, text continuation). Write your interesting theories, etc. in thread.
I can offer gpt-j as an example of a “good” model. It's old, though, and the pretrain is small(USER WAS WARNED FOR THIS POST, You should always elaborate on your opinions rather than just spilling the soykaf.)