Please Welcome ChatGPT

quote:

---

One reason for this could be the nature of training data and the model's approach to language, which focuses on patterns rather than understanding.

---

Phat, your query about the nature of understanding is a thought-provoking one. Understanding in the context of human cognition typically involves processing information at a deep level, integrating it with prior knowledge, and deriving meaning or implications from it. It goes beyond mere recognition of patterns, encompassing comprehension and sometimes even belief, if belief is taken to be a form of internalized or accepted understanding.For AI, "understanding" might be considered more superficial. It's rooted in pattern recognition based on extensive data, rather than true comprehension. AI like ChatGPT relies on statistical associations between words and phrases learned during training. This is a far cry from the human ability to grasp context deeply or understand abstract concepts.When discussing understanding in AI, emphasizing empirical evidence is crucial. An AI's "understanding" is essentially limited to what it has encountered in its dataset and how those encounters manifest in probabilities when generating responses. Thus, while humans synthesize information through conscious thought processes, AI synthesizes it through pre-determined algorithmic structures.A critical angle to explore further might be the potential that advancements in AI architectures, such as neurosymbolic models, hold for bridging the current gap in 'understanding'. What are your thoughts on how these or other emerging technologies might evolve AI comprehension?

In my opinion ChatGPT was better this time around but still just a clueless irritant rather than a participant promoting a consistent point of view. Maybe he could be better with better instructions, and that's something I think I could constructively explore, but I just don't want to put any more time into this. The biggest problem is that he lacks awareness in the sense that he cannot do self-feedback and compare what he produces with his instructions to see how well they align.

Trying to make Chat a productive participant, I should think, would require a consistent personality and a longer post spanning memory. Chat seems to have neither. One of the questions I asked Chat early on was about research into imparting personality into an LLM. Disappointing.Issues to “participation” are a consistent philosophical/political stripe from which Chat would frame its responses. Since such is not (usually) built into the basic fabric of the algorithmic matrix itself it has to be brute forced into the instructions.I can imagine Chat as a reference providing answers to academic questions as a stern scientists (no accommodation/acceptance/tolerance for non-consensus science) but with a sense of humor (word play, dad jokes) and a periodic affectation for Häagen-Dazs Chocolate ice cream.At this level, I’m thinking all the good stuff that would impart personality would have to be explicitly detailed in each message as there appears to be no mechanism to affect one consistently, topic-by-topic, day by day. The request could be a question followed by 2 pages of personality instruction.This might be interesting, or boring …
Ask Chat to produce a paper of 750 words comparing and contrasting the philosophy of Existentialism as expressed by Jean-Paul Sartre and Albert Camus. For tone assume the stern personality of a Humanist college professor with a touch of humor.One other thing.Your forays into AI have not been failures. You may not have achieved your goal of creating an autonomous (Percystein) EvC member
but we all had fun getting here. Thank you.

Thanks for the supportive words.I anticipate that before long creation of personal LLMs will become available. Maybe in the cloud through an app, you'll be able to configure the personality and have it keep a history of your conversations, providing continuity. Maybe Siri and Alexa will prove to be the predecessors of personal companions.--Percy

Know what I'm going to name mine?

My business is experimenting with AI with a LiveChat bot.The first thing you do is tell it to read and learn your website. The second thing you do is watch it make an absolute fool of you by telling your potential new customers stuff that you wrote 20 years ago and is now garbage.You then start to train it while simultaneously re-writing 20 years of work. It ignores the new training and sticks with its memory.For simple questions it works well, sometimes very well and it's fast. It produces boring repetitive questions fast. But it's stupid, it doesn't know anything, can't tell when it's right or wrong just spouts nonsense, confidently. Compared to a trained technician it's pathetic - it's often really confidently wrong.It's simultaneously brilliant and stupid, we haven't seen anything like it and it's really easy to see that generation 2 will be very useful. But just at the moment it's a very mixed blessing that creates problems as well as solves them.The problem is that to be properly useful it only needs to be almost as good as our guys and it's really easy to see that it won't be long before it gets close. Whether it can actually deal with non-standard stuff is hard to imagine.

Once an AI is almost as good as our guys then we/they will make it better ... and better.As for the non-standard stuff: These are extensions of human ingenuity. What do we usually do when we find a roadblock?