> This might sound callous, but I wonder if people saying this themselves have very limited brains more akin to stochastic parrots rather the average homo sapiens.
I have a different theory.
Aside from a few exceptions like Blake Lemoine few people seem to really act as if they believe A.I. is doing the same thing the human mind is doing.
My theory is people are for some reason role-playing as people who believe human thought is equivalent to A.I. for undisclosed reasons they themselves may or may not understand. They do not actually believe their own arguments.
I just asked like I said, give me plot summary until chapter 14, don't spoil the rest of the book. And of course when I told it what it just did it was like oh I'm sorry, here's a summary without the spoilers for the ending. So clearly it could do it without additional context.
>>Do they even have direct access to published works to use as reference material?
I mean, clearly, given that it did answer my question eventually. Also wasn't it a whole thing that these models got trained on entire book libraries(without necessarily paying for that).
>>I wouldn't expect any LLM to be able to respect such a request
Why though? They seem to know everything about everything, why not this specifically. You can ask it to tell you the plot of pretty much any book/film/game made in the last 100 years and it will tell you. Maybe asking about specific chapters was too much, but Neuromancer exists in free copies all over the internet and it's been discussed to death, if it was a book that came out last year then ok, fair enough, but LLMs had 40 years of discussions about Neuromancer to train on.
But besides, regardless of everything else - if I say "don't spoil the rest of the book" and your response includes "in the last chapter character X dies" then you just failed at basic comprehension? Whether an LLM has any knowledge of the book or not, whether that is even true or not, that should be an unacceptable outcome.
Why though? They seem to know everything about everything, why not this specifically.
The problem with this line of reasoning is that it is unscientific. "They seem to" is not good enough for an operational understanding of how LLMs work. The whole point of training is to forget details in order to form general capability, so it is not surprising if they forget things about books if the system deemed other properties as more important to remember.
>> if they forget things about books if the system deemed other properties as more important to remember.
I will repeat for the 3rd time that it's not a problem with the system forgetting the details, quite the opposite.
>>The problem with this line of reasoning is that it is unscientific.
How do you scientifically figure out if the LLM knows something before actually asking the question, in case of a publicly accessible model like Gemini?
Just to be clear - I would be about 1000000x less upset if it just said "I don't know" or "I can't do that". But these models are fundamentally incapable of realizing their own limits, but that alone is forgivable - them literally ignoring instructions is not.
I wouldn't expect an AI to know exactly what happens in every chapter of a book.
Knowing the plot of Neuromancer isn't the same as being able to recite a chapter by chapter summary.
I tried this Neuromancer query a few times and results greatly vary with each regeneration but "do not include spoilers" seems to make Gemuni give more spoilers, not less.
Not really- if you had examined the output closely you probably would have seen noticed it conflated chapter 13 and 14 or 14 and 15. Or you got very lucky on a generation. It definitely doesn't exactly know what happens in each chapter unless it has a reference to check.
I'm not the person you replied to but I'm wondering which Google AI product you are referring to that you use for search which is so excellent that you need someone to find for you an example of it failing?
I think Google has several ai products with search features?
Which one in your experience "seems correct"?
I'm fascinated because I've never found any LLM to be particularly error free at search.
Google.com with the AI overview or whatever they call it now. It seems to source web page information for grounding so it's reasonably correct and doesn't hallucinate recently at least.
I played around with it and its better than it used to be but if you ask it something like
"Whats the name of the third book in the peripheral trilogy going to be" it just regurgitates some dumb reddit comment by someone who seems to be making things up.
There's no actual title that has been announced and the reddit post was not a reasonable bit of speculation.
The problem with these LLMs is they rarely say "the search results were not credible no response can be provided."
These days, Google AI overviews regularly add a qualifier to the effect of "... according to this comment on Reddit <link>"
That's basically a UX trick to entirely sidestep being held accountable for the results, but seems sufficient to notify the user about the provenance of the answer to adjust their grains of salt.
By that logic a Markov chain is better on average just for the fact that it was trained on a large corpus of human knowledge, including psychology, therapy and study material.
I've definitely posted to the same subreddit with two different accounts by accident without being banned.
The android reddit app annoyingly doesn't check for account matches. If you click a browser notification link on Account A it can open a reply form on App account B.
reply