My own guess as to why GPT overuses delve is that it’s an artefact of RLHF. When you’re training a model to respond as a chatbot rather than the next most probable token you’ve people marking responses as good/bad. You’ve got some other criteria as well, like what sounds good.
What’s probably happened is that the “delve” responses sound better to the people doing RLHF, so they’re disproportionally included in the output.
It’s not just delve, there would be a whole list of overused words that you could find by comparing a large corpus of GPT output (or any LLM) to a large corpus of human-written text. You could use that as heuristic for an AI detector, only problem being that you’d need a different corpus for each LLM.
What’s probably happened is that the “delve” responses sound better to the people doing RLHF, so they’re disproportionally included in the output.
It’s not just delve, there would be a whole list of overused words that you could find by comparing a large corpus of GPT output (or any LLM) to a large corpus of human-written text. You could use that as heuristic for an AI detector, only problem being that you’d need a different corpus for each LLM.
reply