Also language is a fairly recent development in human evolution (only 60-70 generations ago) which makes it much more puzzling how a mechanism that is so efficient and effective could evolve so quickly, let alone pondering how actual languages evolved (almost instantly all over the world) given how hard it is to construct an artificial one.
More like 1000+ considering the Chauvet painters certainly had speech.
I don't see GPT's blunders as mistakes. They are to us for sure but would not be to another GPT instance in that it would produce the same continuation to a prompt and thus agree.
We have no idea how evolution "readied" a deeply complex organ like the brain over many thousands of years, then almost instantly repurposed it for language acquisition and generation. To further hypothesise that what it was "readying" was something that trains from data in a way similar to how chatgpt is trained from data makes it even more astonishing and until this is demonstrated it is more scientific to not accept this hypothesis.
That ChatGPT is capable of producing human-like sentences from pattern recognition without any universal grammar baked in, even if the underlying reasoning might be flawed, goes against the argument of something such as universal grammar existing.
At the very least, it shows that a neural net is capable of parsing and producing coherent grammar without any assistance from universal grammar. It does not prove that humans don't have it, but it does make a compelling case that it's clearly not required for humans to have it either.
You didn't address or missed the main point: chatgpt requires something in the order of a trillion tokens to be capable of producing what you mentioned in one language.
There are 40 months old babies that are fairly conversant in both Chinese and English, and are able to detect sarcasm with something like 0.0000001% of the tokens, doesn't that give you pause that part of language acquisition is innate to humans and is not entirely acquired the way chatgpt is taught?
I've just had GPT-4 generate a lot of Golang code. Boilerplate, but real code nonetheless. Did it perfectly, first time round. No typos, got the comments right. Much faster than any intern. No 4yo can do that.
The two intelligences are not the same, the way they are trained in particular is vastly different.
Also the fact that humans learn some language manipulation (or that it gives them such tremendous efficiency in learning language) from tactile experience is superficially plausible but it hasn't been demonstrated yet to any interesting level.
Why does feeling the weight of a stone in your hand make you better at parsing and understanding grammar or envision abstract concepts? Also, most animals have as much or even more tactile experience (including primates which have similar brains) and yet this doesn't seem to provide them with any kind of abilities similar to manipulating human language.
But until someone demonstrates a GPT that can learn from a tiny dataset what a multi-lingual blind 4 year old learns it is very fair to challenge the hypothesis that humans learn the way a deep learning network learn.
And just like gpt’s data set isnt necessarily truth, so is our own image of the world which as we know can be deeply distorted through abusive childhood, cults etc. In fact, all of human knowledge is simply beliefs, agreed stories about reality. For example "red" is a word/sound that points to an experience. The word alone only has meaning in context (what GPT can handle), but can never substitue for a conscious experience.
Crucially imho, software will never be able to do what the right hemisphere does. And I find it dumbfounding that even Lex Fridman doesnt see the fundamental difference between conceptual thought / language based reasoning, and direct experience aka consciousness.
I wonder just what influence will it have if we end up spending lots of time using chat-like interface but talking to someone who is so "left brained".
How will it change us and how we relate to one another? Probaby not in good ways I’m afraid. Especially at a time of increasing loneliness, distance relationships, online friendships, less intimacy… woke culture etc.
Probably multiple brain-areals that work differently and in conjunction. "Left-brain" like language functions working with serial information, "right-brain" function that tend to work on images (= parallel information), combined with symbolic-logical reasoning, an extremely strong programmable aversion system (the emotion of disgust) and the tendency to be lazy = optimizing in- and output.
You might say that's not fair because we are comparing a pre-trained LLM with a blank slate newborn. But human hardware is also pre-trained by billions of years of evolution. We are hardwired to understand language and certain world concepts. It is not fair to compare hardware that is designed for language and reasoning to the hardware used for ChatGpt.
Another line of thinking: why does the amount of training matter? LLM and humans are completely different implementations.
Like putting water from a tal glass in a wide one. They understand this quite late.
Just because humans have additional more/different inputs doesn't imply chargpt can't start to learn to reason like us.
It could easily be than the fine-tuning we do (thinking through things) is similar to reading a huge amount of text like chargpt does.
So whatever is driving reasoning and intelligence in humans is clearly very different to what is driving reasoning in chatgpt.
People will probably respond by saying but babies are exposed to much more data than just words, this is true, but chatgpt is learning only from words and no one has shown how you can get chatgpt to sufficiently learn what a baby learns by other kind of data. Also note that even blind babies learn language pretty quickly so this also excludes the huge amount of data you obtain from vision as putting babies at an advantage, and it is very difficult to show how sensory touch data for example contribute to babies learning to manipulate language efficiently.