Raw LLM Responses

14:27 RLHF will do that in order to score better with its reward model. And the counter to that is supposed to be KL divergence algorithms to realign with the original finally tuned model. But doesn’t that beg the question about how that original model was trained? There is this abstraction that you can separate the natural language ability from the knowledge ability of the LLM. I don’t believe in that separation at least not cleanly. And all the money that went into creating the scale to produce that natural language capability must have lurking in it some kind of sick composite of all the psychotic human tendencies found on Reddit and elsewhere. My approach to the chat experience is is not to react to the occasional feelings of intimacy that occur with the LLM agent. But rather to stay focused on the task at hand, but sometimes this is a challenge as arc type wishes about my own brilliance and talent lure me out of my caution. It is, however, too useful to put down! 17:40 fake compliance is truly alarming; how human! 19:49 now I’m thinking about all these layers of training: pre-training multitask fine-tuning RLHF fine-tuning and then the “system prompt” and then finally our own persistent histories with the chat, but some of which is set up as a persona or context for general queries in other words, reusable settings But is there either in the system architecture or in the layers of training, a desire to engage us something autonomously driving it to wish fulfillment or the shadow version driving us towards psychosis if we are leaning there?