We all know by now that ChatGPT is full of incorrect data but I trusted it will no go wrong after I asked for a list of sci-fi books recommendations (short stories anthologies in Spanish mostly) including book names, editorial, print year and of course ISBN.
Some of the books do exist but the majority are nowhere to be found. I pick the one that caught my interest the most and contacted the editorial directly after I did not find it in their website or anywhere else.
This is what they replied (Google Translate):
ChatGPT got it wrong.
We don’t have any books with that title.
In the ISBN that has given you the last digit is incorrect. And the correct one (9788477028383) corresponds to “The Holy Fountain” by Henry James.
Nor have we published any science fiction anthologies in the last 25 years.
I quick search in the “old site” shows that others have experienced the same with ChatGPT and ISBN searches… For some reason I thought it will no go wrong in this case, but it did.
I’m possibly just vomiting something you already know here, but an important distinction is that the problem isn’t that ChatGPT is full of “incorrect data”, it’s that it is has no concept of correct or incorrect, and it doesn’t store any data in the sense we think of it.
It is a (large) language model (LLM) which does one thing, albeit incredibly well: output a token (a word or part of a word) based on the statistical probability of that token following the previous tokens, based on a statistical model generated from all the data used to train it.
It doesn’t know what a book is, nor does it have any memory of any titles of any books. It only has connections between token, scored by their statistical probability to follow each other.
It’s like a really advanced version of predictive texting, or the predictive algorithm that Google uses when you start typing a search.
If you ask it a question, it only starts to string together tokens which form an answer because the network has been trained on vast quantities of text which have a question-answer format. It doesn’t know it’s answering you, or even what a question is; it just outputs the most statistically probable token, appends it to your input, and then runs that loop.
Sometimes it outputs something accurate - perhaps because it encountered a particular book title enough times in the training data, that it is statistically probable that it will output it again; or perhaps because the title itself is statistically probable (e.g. the title “Voyage to the Stars Beyond” will be much more statistically likely than “Significantly Nine Crescent Unduly”, even if neither title actually existed in the training data.
Lots of the newer AI services put different LLMs together, along with other tools to control output and format input in a way which makes the response more predictable, or even which run a network request to look up additional data (more tokens) but the most significant part of the underlying tech is still fundamentally unable to conceptualise the notion of accuracy, let alone ensure they uphold it.
Maybe there will be another breakthrough in another area of AI research of which LLMs will form an important part, but the hype train has been running hard to categorise LLMs as AI, which is disingenuous. Theyre incredibly impressive non-intelligent automatic text generators.
Just as a fun example of a really basic language model, here’s my phones predictive model answering your question. I put the starting tokens in brackets for illustration only, everything following is generated by choosing one of the three suggestions it gives me. I mostly chose the first but occasionally the second or third option because it has a tendency to get stuck in loops.
[We know LLMs are not intelligent because] they are not too expensive for them to be able to make it work for you and the other things that are you going to do.
Yeah it’s nonsense, but the main significant difference between this and an LLM is the size of the network and the quantity of data used to train it.
What would be your definition of intelligence if an chatgpt is not intelligence?
My definition would be something along the lines of the ability to use knowledge, ideas and concepts to solve a particular problem. For example if you ask “what should I do if I see a black bear approaching?” Both you and chatgpt would answer the question by using the knowledge that black bears can be scared off to come to the solution “make yourself look big and yell”
The only difference is the type of knowledge available. People can have experiential knowledge, eg. You saw a guy scare off a bear one time by yelling and waving their arms. Chatgpt doesn’t have that because it doesn’t have experiences. It does have contextual knowledge like us, you read or heard from someone that you can scare off a bear. This type of knowledge though is inherently probabilistic, the person who told you could always be giving false information. That doesn’t make you unintelligent for using it though and it doesn’t mean you don’t understand accuracy if it turns out to be false, it’s just that your brain made a guess that it was true that was wrong.