Thanks AD, that's very interesting.
I was struggling for a metaphor the other day - I don't have many but the speed of light came to mind, the increasing mass is limiting - more and more energy needed for marginal gains in speed. It has felt that with all of this mass of information, subtly and insights become lost, and so more isn't going to be better, naturally - just worse. I go back to Chomsky when around 2 and half years ago, a few months before his stroke, he said that human-intelligence is about making connections on very limited information. It seems we are designed mostly very well for this, unsurprisingly, though harsh evolution - these models seems the opposite (to me) - fairly simple in design (a big matrix) with an overwhelming amount of information to make judgements upon.
If the model is built on ever increasing mass, we can imagine the struggle. Humans seem a little like this we start out as children with wide "varaince" making weird and unusal connections adults don't, this seems to decline with an ever gorwing mass of information. It is interesting that a reported trait of several geniuses were that they reamained playful. Sometimes we see experts switch fields and make major breakthroughs, they know less but are able to make unusual connections and solve hard problems.
At the end, I was a little surprised by this:
"In the context of large language models, research found that training LLMs on predecessor-generated text — language models are trained on the synthetic data produced by previous models — causes a consistent decrease in the lexical, syntactic, and semantic diversity of the model outputs through successive iterations, notably remarkable for tasks demanding high levels of creativity"
This seems like a bad design idea, the models don't know reality, and then external world stops being a stimulus - the information they received were it seemed, analagous to the shadows in Plato's cave, anyhow. Once it starts using itself, it's own answers, as learning, then there may become fairground-mirror images of those shadows - and eventually lost through interations of these images. Beliefs in this model seem reinforced by itself - a pseudo stimulus, rather than the reality it is trying to understand, it would seem.
I must admit there does seem a blandness to AI at times. It both seems creative but missing it at the same time.
Again thanks, that was very interersting - it does seem in line with the experience that these models haven't developed as hoped. I didn't run through the maths, that skill is presently flat-packed in the ivory tower loft! Another one of our useful evolutionary-adaptations it seems - it's been years since I rode a bicycle too!
(I was rushed in the previous post, there a few errors which completely inverted the meaning of the sentence, but hopefully the general direction and context of the post made this fairly clear!)
Edited by ambivalent, 03 June 2025 - 03:27 PM.
