Regarding access, OpenAI aren't very open in revealing the training dataset, but from the evidence Google found it seems to be extensive. One might assume from its responses that at minimum it would contain the entirety of Wikipedia, and countless literary works and text books. If you give me a question and 1 hour I can probably give you a similar answer to ChatGPT (especially if I can just patently copy and paste as ChatGPT often does).
On the topic of recall...it's a computer model...you cannot equate its ability to recall a conversation to the ability of a human to recall a string of numbers in much the same way that calculator can't pass a maths exam. Ergo my opinion that IQ scores of ChatGPT are invalid, much like all the other multiple 'tests' that ChatGPT 'passed'. In academic circles this evidence hasn't been viewed as proof of AI awesomeness but rather the laziness of our current testing systems.
On comparing AI training to the learning that occurs in a small child; if you have small children you know that there is a gulf of difference in their learning styles. Not only is a large proportion of child development devoted to acquiring coordination and gross/fine motor control (something a stationary computer doesn't require), small children will often attempt to immediately apply newly learned material to novel settings (i.e., generalise). Here we see, even in infancy, the ability for human intelligence to generalise learning from one context to another - something notably missing from our current AI models hence the OP.
As pointed out by other responses, and indeed yourself when you speak of weights, our current AI tools are effectively statistical prediction tools. ChatGPT doesn't 'understand' your question in the same way a human counterpart might, but rather it responds with a string of words taken from the average of its training data on the given topic. This is extremely evident when ChatGPT is probed on topics of which you have an advanced understanding.
I used ChatGPT 4 extensively last year, and now only use the free version sparingly. During that time I routinely experienced hallucinations in responses, erroneous information and downright mistakes in its coding responses. The extent of its utility to myself now is that it can restructure a paragraph in a few different ways (that still needs editing to adjust the verbosity), or provide 3 wrong and possibly one right answer to a coding problem. Essentially it saves me spending time on editing my reports and searching StackOverflow/Exchange...apart from those tasks it's essentially your buddy who thinks he knows a lot about a lot but ends up only knowing a little, and even then it's mostly inaccurate.
What is more interesting, and relevant to the OP topic, is the speed with which humans have become accustomed to using AI tools to augment their abilities, flexing their ability to generalise learning across contents, almost like some kind of generalised intelligence