"The first ten million years were the worst," said Marvin, "and the second ten million years, they were the worst too. The third ten million years I didn't enjoy at all. After that I went into a bit of a decline."
ChatGPT is an incompetent, unethical mess, and the rare occasions where its stochastic output aligns with reality, it is abusing the generosity of people who expected to be licensing their work to fellow human creatives and not the inanimate avatar of cognitive oblivion.
I think it's intellectually lazy to stick with the stochastic parrot line of thinking now. There's a number of emergent properties that are appearing as LLMs scale that give them abilities beyond that paradigm. Check out the "Sparks of AGI" paper from Microsoft research - or more realistically one of the youtube summaries of it since its quite a big read... Here's one from the horse's mouth: https://m.youtube.com/watch?v=qbIk7-JPB2c
It's probably just added to its system message, which feels a bit bolted on. I wish we had a really competent personal assistant (assuming you could keep your privacy, which seems unlikely in this day and age), which learned about you and your preferences over time. Something with actual long term memory and the ability to be fine tuned by user input.
On the other hand, a lot of cognitive functions in the human brain are pretty "bolted on", so this might not be too terrible if you consider that a benchmark...
You're not wrong. Perhaps the way we think of memory, etc is too rooted in our biology, and machine intelligence will continue to surprise us with powerful capabilities from seemingly simple structure.