The TLDR is that neither of the algorithms actually has access to the dataset it was trained on (or any data for that matter), and therefore can't do "lookup" to find facts for you. Based on the fuck-ton of data it was trained on, when presented with a statement, it interprets what you're saying and makes an educated guess about what words you are probably looking for in a response, and puts those together into a sentence with a whole bunch of math.
For stuff that is somewhat common knowledge, it will often guess fairly accurately as the correct response was most likely part of its training data and it was a common example that it can reiterate with confidence. For more obscure stuff, or problem solving (that is often unique, like your question about the age of a person), it does not immediately remember this (because it wasn't part of its dataset) and can't give an accurate answer.
Think about it like this – you are ChatGPT, and this is the basic process you've gone through. You've taken a university course on *everything in the world.* Over time you've been reading, then gotten tested, and reviewing over and over again millions of time until the knowledge starts to solidify in your mind. Then you finish the course, you get your degree, and every time people ask you a question it is like taking a big exam. You have no access to the material you learned from, and now you're forced to give answers straight out of your memory. Some stuff was part of your course curriculum very many times, and these things you know and can reiterate correctly every time. But some stuff only appeared a few times or not at all, and you make an educated guess based on what you think sounds sensical. But, crucially, you have no way of reasoning your way to see if it's correct or not.
It's a very nifty tool for language manipulation, but shouldn't be used for information lookup.
2023-04-06 08:59