[flagged]
GPT LLM algorithms use a probabilistic language model to generate text. It is trained on a large corpus of text data, and it estimates the probability distribution of the next word given the previous words in the sequence.
The algorithm tokenizes the input into a sequence of tokens and then generates the next token(s) in the sequence based on the probabilities learned during training. These probabilities are based on the frequency and context of words in the training corpus. You can ask ChatGPT/etc yourself, and it'll tell you something like this.
This is not remotely like what human brains do. Your ideas cohere from network connections between the neurons in your brain, and then you come up with words to match your idea, not your previous words or the frequency that the words appear in your brain.
This is just a description of the input/output boundary of the system. The question is what goes on in-between and how to best characterize it? My input/output is also "tokens" of a sort: word units, sound patterns, color patches, etc.
>Your ideas cohere from network connections between the neurons in your brain, and then you come up with words to match your idea, not your previous words or the frequency that the words appear in your brain.
At some high level of description, yes. At a low level it's just integrating action potentials. What's to say there isn't a similar high level of description that characterizes the process by which ChatGPT decides on its continuation?
I'm pretty confident that that isn't all the human brain does, but we certainly do that in many situations. Lots of daily conversation seems scripted to me. Join a Zoom call early on a Monday morning:
Person 1: Good Morning!
Person 2: Good Morning!
Person 3: Did anyone do anything interesting this weekend?
Person 1: Nah, just the usual chores around the house.
etc.