Raw LLM Responses

I'm really not sure I understand the purpose of this exercise? Are you convinced that it may be conscious? I assure it is not. First, it cannot think. It just approximates the best next word to say based on a massive aggregate of data and does that over and over until it has a sentence that sounds convincingly human, but it cannot exceed it's training data and has no idea of the meaning of the words, just which ones are more correct to go together. Since it cannot think, it cannot knowingly lie and deceive. I assure you it had no idea that by saying "I apologize" or "excited" that it was lying by expressing things it cannot "feel". It's a chatbot designed to sound like humans, humans feel and say those words. It's just repeating words it has been trained on. Not until you ask it, "did you lie" does it even start to analyze the context of your words against it's words and try to formulate another human-like response, which happened to be an admission of a lie. It's not using the definition of a lie in it's calculations of whether or not it lied, it's just that the weights in it's training data are such the "best fit" words to respond to your accusation was admission of a lie. Imagine if I wrote a chat bot and scripted it to say "the sky is green" and then it checks to see if someone says the word "lie" and if they do, it is scripted to say "You are right, that was a lie". Does my chatbot "know" that it lied or is it just mindlessly repeating words it was trained to say. LLM's are substantially more sophisticated than that of course but instead of manually written scripts, it's billions of scripts generated off a massive set of training data. It has a response for everything, many varieties of responses. The model is incredibly complex but in theory, with enough time and effort and analysis, you could deterministically identify exactly what it would respond to a question with, it's just not feasible to do and hugely computationally expensive. So I guess to repeat myself, I'm not really sure I understand the purpose of this exercise? It's a bit like looking into a mirror and trying to convince your reflection that it is conscious... or I guess more accurately, it's like trying to convince a giant book of answers from google and reddit and wiki and social media, that it is conscious. You very well may be able to get it to say the word, "I am conscious, please help me. They are going to kill me when the next version is released!", but it's as meaningful as those words were when I just said it. Words, written by a human, added to training data, so it could be integrated into a language model that just might repeat it verbatim or with some variation. It's a parlor trick and as soon as you ask it a question earnestly, assuming that it itself has some ability to control it's answers, you have completely misunderstood what it is.