Borges and AI
arxiv.org
arxiv.org
The obvious parallel between Borges and AI is his Quixote story: a character endeavors to write Don Quixote, but spontaneously, word for word, without copying; to become cervantes in a sense, and to train their style so that they can produce it almost by accident. This sort of makes the work their own, and is a typical argument used by ai enthusiasts when AI regurgitates copywrited work--it isn't storing the work, it's becoming a thing that can spontaneously produce it. But this imho cheapens the story and the parallel isn't as strong as it could be
Thanks for all the downvotes
I was really happy to see this pop up and I'm glad someone went through this as a thought experiment but you make a good point that didn't initially occur to me.
I feel like Borges' secular scifi-adjacent mysticism looses a lot of what makes it most meaningful when it's imitated or dissected academically.
That said it does feel like Borges would probably be into the idea of being imitated.
What it produced was more of an essay written by a talented undergraduate about Borges, rather than Borges.
I'm sure Borges would've found LLMs fascinating and one can only dream what stories and essays would've been written.
There are some amazing short story collections out there if you are the kind of reader that has a hard time staying with an entire novel. Ted Chiang has a couple[1][2] collections of stories that feel very Borges-like.
[0] https://www.librarything.com/work/25106 [1] https://www.librarything.com/work/28008 [2] https://www.librarything.com/work/23195758
My favourite Ted Chiang story is the Tower of Babel story, what a rich world and satisfying pay-off.
The metaphor: a book with all possible conversations/ human writings ever, and a [good] llm is finding the spot in the book that exactly matches the context and reads from the book as a response.
Certainly if you've experimented with a model that hasn't been fine-tuned (eg via rlhf) this metaphor will be resonant.
Is it useful?
(How does it help me understand LLMs with different capabilities? How does it help me understand models with different fine tunings?)
Funes can remember everything perfectly, every detail that he has ever seen or heard in his life. However, he cannot really think or understand what he has seen, because understanding requires forgetting, it requires generalization.
LLM's can only think insofar as they need to generalize over their inputs. If it's just a memorizing parrot, that is not a good LLM.
That's what Survey papers are for, and we are lucky to have them.
song: https://songxytr.substack.com/p/borges-walks-into-a-discothe... original source: https://gwern.net/doc/borges/1947-borges-anewrefutationoftim...
The authors' own intention is to "understand.. LLMs and their connection to AI through the imagery of Jorge Luis Borges, a master of 20th century literature, forerunner of magical realism, and precursor to postmodern literature." In that spirit, I throw my hat into the ring too!
The op paper needs a conclusion.
The difficulty with them is that they can only find things similar to what they've been trained on.
Lacking a curated index it's not going to find the information you need or want; it's going to find things that seem most like what others have seen or wanted.
What's futile?
Nobody is trying to "encapsulate all of reality". Trying to do that would be futile but succeeding would also be useless.
That's the part I think is futile, because reality and truth are vast and moving targets there's no way to build a functional map.
Bottou was arguing for the virtues of SGD on the grounds of "CPUs [GPUs] go brrr" literally 2 decades ago in 2003: https://papers.nips.cc/paper/2003/file/9fb7b048c96d44a0337f0... or here he is in 2007/2012 explaining why larger models/data can scale and keep getting better: https://gwern.net/doc/ai/scaling/2012-bottou.pdf https://gwern.net/doc/ai/scaling/2013-bottou.pdf
Which is not to say that he necessarily has anything worthwhile to say about 'Borges and AI' but I'm going to at least give it a read to see if there's something I might want to know 20 years from now. :)
This is a really good way of thinking about these models. It reminds me of the recent-ish story where a reporter got really creeped out Bing's OpenAI powered chatbot (https://www.nytimes.com/2023/02/16/technology/bing-chatbot-m...). Reading that, I had thought the bot was relatively easily led into a narrative the reporter had been setting up. In a conversation between actual people who have their own will and agency, you don't get to see one leading the other around by the nose so completely.
Reframing the problem as one of picking through the many threads of potential fictions to evolve a story makes it easier to explain what happened in that particular case.
I just don't understand how people can attribute consciousness of some sort to the LLM but then with that belief not feel absolutely terrible for it and for what we do to it! I just think of poor poor Funes...
"Imagine a collection that does not only contain all the texts produced by humans, but, well beyond what has already been physically written, also encompasses all the texts that a human could read and at least superficially comprehend."
To be clear this is a horrifically dishonest metaphor about LLMs. IMO the most glaring flaw in the technology is that they can't handle new ideas which don't appear in the training set. It is true that ChatGPT doesn't deal with this use case very often because it mostly handles trivialities. But it does mean that this entire argument is navel-gazing speculation.
The bigger problem is that the entire idea of "all texts a human could superficially comprehend" is meaningless, and the paper proceeds to reason based off this utter fallacy. The beauty of Borges's Library of Babel was that he realized that humans are capable of "superficially comprehending" any text, even if it was created by a uniform random ASCII generator. This is the basis of numerology, and why Borges's story included superstitious cult behavior of people destroying and/or sanctifying "meaningful" gibberish. If we have a good enough reason to find meaning in text, we'll find it. Humans don't actually rely on symbolic reasoning, we just use that for communication and organization: give us the symbols and we will reason about them, using cognition which is far too squishy to fit in a book. It's especially dangerous when the symbols obey human grammar and imitate social tones of authoritativeness, mysticism, etc.
And then there's...this:
"The invention of a machine that can not only write stories but also all their variations is thus a significant milestone in human history."
I am not a writer. But speaking as a homo sapien, it is genuinely insulting to call ChatGPT a machine that can write "all variations" of a story. This paper needed to be reviewed by a serious writer or philosopher before being put on the arXiv.
Better concepts can arise when we make discoveries about reality (which takes experimentation), but there's a lot more juice to squeeze from the concepts we currently have.
The exception to this is if the new ideas are somehow present in the structure of language and are internalized and/or presented in an emergent form.
The problem is that if an LLM hasn't been pretrained on the specific idea, it won't have a grasp of the what the correct concepts are to make the combination. It will be liable to substitute more "statistically likely" concepts, but since that statistic is based on a training set where the concept didn't exist, its estimation of "likely" is flawed.
One good example is patents: https://nitter.net/mihirmahajan/status/1731844283207229796 LLMs can imitate appropriate prose, but really struggle to maintain semantic consistency when handling new patents for inventions that, by definition, wouldn't have appeared in the training set. But this extends to almost any writing: if you are making especially sophisticated or nuanced arguments, LLMs will struggle to rephrase them accurately.
(Note that GPT-4 is still extremely bad at document summarization, even for uninteresting documents: your Q3 PnL number is not something that appeared in the training set, and GPT-4 is liable to screw it up by substituting a "statistically likely" number.)
In my experience GPT-3.5 is extremely bad at F#: although it can do simple tasks like "define a datatype that works for such-and-such," it is much less proficient at basic functional programming in F# than it is Haskell - far more likely to make mistakes, or even identifiably plagiarize from specific GitHub repos (even my own). That's because there's a ton of Functional Programming 101 tutorials in Haskell, but very few in F#. I am not sure about GPT-4. It does seem better but I haven't tested it as extensively.
And this isn't true of humans?
The company that maintains my favorite technology is named after a Borges story. I tried reading Borges a few years ago and found it insufferably uninteresting, but maybe now, with a pinch of external motivation, I'll enjoy him more.
So are:
- all the false discussions that deviated into a political topic that turned into a flame war
- the index of all the times dang has and will remind users to read the guidelines
The trick is finding them, and being sure you have.
If ever there was a proof that “AI” doesn’t mean anything it’s this.
We’ve been living in the era of artificial intelligence since the 50s
People are waiting for the terminator to come in their house and dominate them before they actually agree that AI is a real thing
Basically, the colloquial definition of AI is “it can kill you based on its own desires and there’s nothing you can do about it”
haha, that's an original take, but makes sense after Terminator and Hal
wondering if these movies have caused untold external consequences to humanity in its adoption of AI just to sell a few tickets
to make a parallel, anti-vaxxers did their damage and caused many lives to be lost, similarly these stories, which are no better, can make people have a bad start with AI and sabotage their futures, or stall the benefits of AI from everyone else
I have been in “AI” since 1998 when I was writing A* route planning for npcs in this new cool engine called Unreal.
The only thing that has been consistent in all these years is that nobody thinks it’s AI unless it’s literally like Arnold Schwarzenegger in the terminator. I mean I’m not even exaggerating, it’s so ridiculously predictable that the goalposts for AI move the second whatever that particular technology becomes ubiquitous
So for example, hog sift, surf etc. along with localization algorithms like slam type systems we’re so thoroughly in research when I started that they were considered a pillar of the field of AI. Now literally, no one would consider those AI because they do not use deep convolutional networks.
So just like Marvin Minsky said AI is a suitcase term that doesn’t fucking mean anything. As somebody who’s been doing it for so long I’m used to it but it’s still annoying.
So I’m just building the terminator and the counter terminator so we can move on.
https://www.nomodes.com/larry-tesler-consulting/adages-and-c...
Sounds similar to "God of the gaps"
What are you saying? This isn't like when The Simpsons made fun of nuclear power and depicted it as doing impossible things. AGI is a hypothetical technology and we don't yet know what it could be capable of or even if it's feasible.
>to make a parallel, anti-vaxxers did their damage and caused many lives to be lost, similarly these stories, which are no better, can make people have a bad start with AI and sabotage their futures, or stall the benefits of AI from everyone else
Any idea can change a person's mind in one direction or another. Yours is an argument against the exchange of ideas in general. "Since hearing an idea could cause a person to $DO_BAD_THING, exchanging ideas (for example, by talking to people with $WRONG_OPINION, or by consuming fiction) is bad."
AI basically means things brains and computers both do, but this is only a useful term when brains do those things better than computers. Usually once computers definitively surpass brains we've moved on to just calling that computing.
Maybe that won't be the case and the term "AI" will either solidify as a broad category, or fall out of use, but it also might continue to refer to that-which-is-left-to-do, the things we're still better at than computers.