Are you aware of the NYTimes Connections puzzle? It is sort of built to appeal to the minds of cryptic crossword doers. And is something we used to think AI would never be able to solve. Silly us. But today’s puzzle has something dear to Henry’s heart.
You have 16 words or names and you have to group them into 4 groups of 4. And today they had a group of Dickens characters, and rather improbably for an American audience, they had the police inspector from Bleak House (along with Pip, Scrooge, Twist). This is the first reference I have seen to a Bleak House character in popular US culture.
The Paul Graham bit leaves me with a mental picture of him having a go at Masta Killa of the Wu-Tang Clan for using ‘iniquity’ rather than the plain English ‘evil’. Not everything has to be Hemingway.
With regards AI, I think there is a fundamental tension between training a general purpose AI that can write competently in many styles, and one that has a distinct voice.
I suspect it is similar to the way a great session musician can reproduce any style they are requested to, but many of them don’t have musical projects or distinct ‘voice’ of their own.
Conversely, people with a strong voice are not necessarily well read / highly trained musicians, and may have a strongly idiosyncratic set of influences.
As it becomes cost effective, it would be interesting to take one of the base models and then train it on slightly random sub-canons, reflecting the number of books a real human might read between 11 and 30.
Or to consider how a human student progresses from writing pastiche to quality poetry or prose, - and how do you develop actual discernment, rather than a snobbery based on received opinion?)
I’ve been using AI tools for my writing, and it’s clear to everyone that they’re not yet capable of producing a truly beautiful literary style—especially in poetry or fiction, where they still fall flat. They’re a bit better at nonfiction essays, though they don’t really develop a personal voice. For purely informational pieces, though, they’re decent. I’ve noticed that by providing lots of examples and actively correcting the AI’s output (and explaining why), it gradually improves in our chat. When it comes to mirroring my tone in articles, it works surprisingly well.
Of course, if we could do proper fine-tuning—especially with open-source models, though that’s technically complex—it would be ideal. Personally, I don’t see anything dishonest about using these tools: they boost and streamline my workflow, and they make my research more engaging. After all, reading or quoting a book is also a way to enhance our intellectual reach.
I’m not particularly concerned with authorship and have no patience for the romantic myth of the solitary genius. I write not to earn praise but to see what I’ve produced; I’m far more interested in the pursuit of knowledge than in applause. If someone else can express what I want to say better than I can, I’m genuinely happy. And if I can get better with an AI, even better—especially since, as an Italian, I’ve found they have a much better command of English than I do!
LLMs (they are not AI) have no taste. They're just (impressive) statistical models. If the model has a poor fit for some task, it will perform poorly. The fact that they think unimpressive poems are as good as Emily Dickson just shows that whatever they're optimizing is not good poetry. At least for now. I find it weird to call an optimization function good or bad taste.
So good for you to delve into this! Keep delving.
Are you aware of the NYTimes Connections puzzle? It is sort of built to appeal to the minds of cryptic crossword doers. And is something we used to think AI would never be able to solve. Silly us. But today’s puzzle has something dear to Henry’s heart.
I am not...
You have 16 words or names and you have to group them into 4 groups of 4. And today they had a group of Dickens characters, and rather improbably for an American audience, they had the police inspector from Bleak House (along with Pip, Scrooge, Twist). This is the first reference I have seen to a Bleak House character in popular US culture.
LOVE THIS
The Paul Graham bit leaves me with a mental picture of him having a go at Masta Killa of the Wu-Tang Clan for using ‘iniquity’ rather than the plain English ‘evil’. Not everything has to be Hemingway.
With regards AI, I think there is a fundamental tension between training a general purpose AI that can write competently in many styles, and one that has a distinct voice.
I suspect it is similar to the way a great session musician can reproduce any style they are requested to, but many of them don’t have musical projects or distinct ‘voice’ of their own.
Conversely, people with a strong voice are not necessarily well read / highly trained musicians, and may have a strongly idiosyncratic set of influences.
As it becomes cost effective, it would be interesting to take one of the base models and then train it on slightly random sub-canons, reflecting the number of books a real human might read between 11 and 30.
Or to consider how a human student progresses from writing pastiche to quality poetry or prose, - and how do you develop actual discernment, rather than a snobbery based on received opinion?)
Agree. My step daughter's twelve year old autistic son tells far better jokes than any A.I. so far.
I’ve been using AI tools for my writing, and it’s clear to everyone that they’re not yet capable of producing a truly beautiful literary style—especially in poetry or fiction, where they still fall flat. They’re a bit better at nonfiction essays, though they don’t really develop a personal voice. For purely informational pieces, though, they’re decent. I’ve noticed that by providing lots of examples and actively correcting the AI’s output (and explaining why), it gradually improves in our chat. When it comes to mirroring my tone in articles, it works surprisingly well.
Of course, if we could do proper fine-tuning—especially with open-source models, though that’s technically complex—it would be ideal. Personally, I don’t see anything dishonest about using these tools: they boost and streamline my workflow, and they make my research more engaging. After all, reading or quoting a book is also a way to enhance our intellectual reach.
I’m not particularly concerned with authorship and have no patience for the romantic myth of the solitary genius. I write not to earn praise but to see what I’ve produced; I’m far more interested in the pursuit of knowledge than in applause. If someone else can express what I want to say better than I can, I’m genuinely happy. And if I can get better with an AI, even better—especially since, as an Italian, I’ve found they have a much better command of English than I do!
Here’s an example: https://www.domusweb.it/en/news/2024/11/28/boomer-aesthetic-memes-images-ai-fake.html
Francesco D’Isa
I don’t say it was dishonest in general. It would be dishonest for me.
Of course! I understood it, I was just sharing my experience here, to expand the collection :)
LLMs (they are not AI) have no taste. They're just (impressive) statistical models. If the model has a poor fit for some task, it will perform poorly. The fact that they think unimpressive poems are as good as Emily Dickson just shows that whatever they're optimizing is not good poetry. At least for now. I find it weird to call an optimization function good or bad taste.
Would you consider this poem to be good?
https://open.substack.com/pub/dailycircus/p/yesterday-a-scene-from-the-oval-office?r=4ghy64&utm_medium=ios
I don't, it's tolerable, I guess. But AI-written, took 3-5 seconds
No same problem just not very good. Impressive that AI can do it, but not impressive.
This, also entirely AI-written, I consider to be not bad
https://open.substack.com/pub/dailycircus/p/bruno-schulz-to-franz-kafka-esq?r=4ghy64&utm_medium=ios
Pretty bad! Forgotten gods and moths? If a human wrote that you wouldn’t rate it at all.
People seem to appreciate Jane Austen forgeries much better
https://open.substack.com/pub/letterspurloined/p/the-purloined-letters-that-were-never?r=4ghy64&utm_medium=ios
Well, maybe that means you don't like Bruno Schultz's style? )
But the basic capability on display here is just that, basic.
Well, it's not bad for a five-second work, obviously, one has to work harder to get something really interesting