The tests you picked have fixed answers and are mostly about remembering the correct information that needs to be applied.
Things like chatgpt and computer are excellent at storing and being able to access information so then trick is getting the program to be able to “read” the question, find the right information, and write a sensible answer.
What these programs are doing is taking all the data they have stored and computing how likely words and topics are related and how to put a readable sentence. So the words Heart and cardiovascular are going to put together, stroke and cardiovascular as well, stroke and numbness etc. It also does this with how sentences are supposed to written. The rules say it’s white (adjective) dog (noun) not dog white.
And like how we teach babies words/associations by repetition the ai gets lots and lots of input that repeat the same relationships. This makes the ai better both at understanding what the question is asking and how to put then sentence together since it sees how all the data we gave it did it, and then used math to make sure it puts words in the right order.
Because it cheats. In a very dumb way too.
It takes all the answers it finds “on the internet”, shuffles the words around and hope for the best.
Chat gpt is very good a doing this kind of game, specifically tuned to minimize errors. But it still only copy-pastes stuff in a very fancy way.
It is NOT able to come up with original solutions, but it is pretty good at giving out known solutions in a way that looks organic.
GPT is not AI in the sense most people think, the AI people imagine is in fact AGI which has not been invented yet.
The AI that is GPT is in fact just a very well “trained” pattern matcher, ie a program that records patterns (sequences of things that repeat) has had an incredible amount of data fed into it and stored.
A significant part of that will be from legal texts, medical texts, and more. As a result when gpt sees a question such as “what is the legal meaning of the word ‘Some’?” It will look through its recorded sequences of patterns find something that has keywords of that question, then find patterns of responses (needs to be multiple so it can have an assurance of accuracy) and then takes key words from THAT, applies to a sentence generator (not hard in english) and boom…you have an answer to a test question.
Especially since most legal and medical exams have previous tests posted online, it’s not had for gpt to do this at all since things like that is perfect foodder for training an “ai”.
Latest Answers