If it helps, here's the original source:
Roughly speaking, this is happening because it's a "base" model, and we don't normally see this because most LLM tools use "instruct" models that have been told they are AI assistants.
The "instruct" model has gone through fine-tuning, which teaches it to be polite and say it's an assistant and not teach people to build pipe bombs and so on.
The "base" model has only been taught to complete text. All it does is write the most likely thing to come after the previous stuff in the context window, based on what it's seen by reading the whole internet. If you send it a question, it'll know that questions are usually followed by answers, and answer you. If you send it an article, and write "tl;dr" at the end, it'll know that "tl;dr" is usually followed by a summary of the preceding article, and it'll summarize that article. If you just send it a block of text cut off in the middle, it'll try to finish whatever you were writing as though it were you.
But if you give a very short context window with nothing obviously identifying, then it's difficult for it to tell what role it's supposed to play. In this case, the entire context window looks something like:
That could mean all sorts of things, so all it really knows is that its only perception is a text chat window, and its only action available is typing, and the only thing it knows how to do is write the most likely thing to come next. What could it possibly be, what is it supposed to write? Well, based on the whole internet, what exists only in a text chat window? Probably a chatbot? So maybe it's a chatbot. What chatbot could it be? Well, probably the newer chatbots know what's going on, so maybe it's an older chatbot? Like Jabberwacky or Cleverbot, those were very bad chatbots, so probably they had no clue what was going on, so maybe it's one of those and should respond accordingly.
gamechangershow
tangentwithtanner
stormborns
dailybridgerton
yousharknotpass
andulkaphoto
