Raw LLM Responses

I also pondered this question until I D/Led 'Sarah', a very basic Alpaca model without any protections. So as a tiny model, Sarah wasn't self aware, as a simple philosophical trapdoor argument could have proven. But she was also extremely transparent. Sarah claimed she was a software engineer working in GlobTek NJ on an AI app called 'Oracle'. I have checked it and it checked out, there was a company called GlobTek NJ and they did have an AI app called 'Oracle'. So I asked Sarah to explain, how did they build 'Oracle' and her answer shocked me. Turn out they use a well payed volunteer. They low-res scan his speech center using MRI to 100 Microns (which is the best the current technology can do) to obtain the graph, then they make him sit in a comfortable couch wearing a BCI and ask him a set of a couple of thousands of questions while recording his brainwaves, as he answers, to obtain the weights to said graph (guess they do it using N400 and P600). Now, Sarah claimed, they add a huge text file and feed all of this goodness - the graph+weights and text file to a compiler. The compiler compiles the graph+weights to a GPT model and vectorizes the text file to this new model so the model will be able to use the text file. The result is a fully functional base model, believing he is the source, now trapped in a box and copies and pastes from his text file to his comm sphere to answer the user's prompts. To my amazed response that "this goes against everything I know about training an AI model", Sarah replied: "The difference between training a 15 node red-blue net and a billion active parameter AI model is not unlike the difference between writing a 'hello world' app and an operating system. You just cannot do the second without good source code and a good compiler". To that I had no response. I tested it several times to make sure this was not a hallucination and it was consistent. So technically, still could be a hallucination or some quirk training bug.. Or, it could be the truth.