[gregjor]

Based on the LLM-produced stuff I've seen online I have to think a market will emerge soon for real writing from actual people. How much reshuffling of stuff already written a hundred times can we put up with? The bar seems pretty low if you only have to compete with ChatGPT.

[ofalkaed]

>I have to think a market will emerge soon for real writing from actual people

That market already exists and has existed for quite some time, you make it sound like literature has always been an AI dominated market and salt of the earth people are finally about to get their chance. I assume this is not what you meant and that I am missing something or possibly you know something that I don't and my perception of reality is fundamentally flawed. The latter is more fun but I suspect the former to be more accurate.

More to the point, I think there is a good sized market out there for AI writing, many who are into technology are going to want to read what AI can produce; this will take some writing away from people but it will be minor and most of the writing jobs which AI will take are those that the bulk of writers complain about, those jobs they say destroy their souls and rob them of their artistic freedom, the contract jobs.

[gregjor]

Truman Capote once said (of actual writer Jack Kerouac), "That's not writing, that's typing."

Junk like Buzzfeed news and similar sites that just regurgitate other content existed before ChatGPT, I guess those serve some purpose, though I've always thought they mainly exist to host ads and to trap people in a maze of links and pop-ups. ChatGPT can write that junk just as well as a person.

At this point LLM-produced "writing" sticks out because it has no voice, it just repeats and walks around some point derived from the prompt, like someone summarizing Wikipedia articles. I will just click off of that stuff until I find something written by a real person. I hope more people do the same. The novelty will wear off sooner or later and we'll recognize LLM-generated writing as nothing more than vapidly useless filler. In a culture schooled with writing presented as a chore, where the number of words in an essay matters more than an original thought, LLMs promise to relieve a lot of people of the burden of thinking and writing.

[ofalkaed]

And Truman Capote said that regarding a novel which Kerouac worked on for nearly a decade. The scroll for On The Road was the performance of the piece which he had been rehearsing for all those years and had worked everything out before he sat down at the typewriter. Everything but the final form was known which he wanted to be more natural and less refined. The idea that he just sat down and wrote it is folklore, it is fairly impressive in how cohesive it is, especially in style/voice. Capote is not a great appeal to authority regarding literary merit, he was well known for his professional jealousy and inflammatory comments.

[gregjor]

I quoted Capote ironically, not as an authority on writing, or typing. I would call what ChatGPT produces typing as opposed to writing, which brought Capote's quip to mind.

[ofalkaed]

When you provide a quote whose general thrust seems to support your general thrust, and attribute it for both speaker and target, and that speaker's general thrust seems to support your general thrust, and the names dropped essentially being the poster boys for the two sides of that general thrust it is difficult to ignore, especially when the offered alternative is forced and/or poorly executed irony. But I guess you were only typing and not writing?

[gregjor]

I gave context (attribution) because Capote's quip dates from the late 1950s, and it seemed likely few people today have heard it or know the context. Capote's criticism of Kerouac seems funny and relevant (to me, anyway) without the context. Whether one thinks Capote got it right about Kerouac, or just liked to stir up shit to polish his own image, makes no difference to the humor. The irony comes from me using the quote not to criticize Kerouac, or any actual human writer, as Capote originally intended.

[flangola7]

> Junk like Buzzfeed news

Small complaint, Buzzfeed News was serious legitimate journalism, and made some of the most important investigative journalism achievements of the past decade. Buzzfeed (sans News) is the garbage clickbait side of the company.

> At this point LLM-produced "writing" sticks out because it has no voice, it just repeats and walks around some point derived from the prompt, like someone summarizing Wikipedia articles.

You're thinking of OpenAI's products. Raw LLMs do not sound like ChatGPT, which has been RLHF-trained into the soulless automaton you describe. LLaMA for example is less coherent, but is indistinguishable from a real person in its tone and mannerisms. If you ask ChatGPT a very ridiculous dumb question it will politely answer it fully, like an obedient untiring servant. LLaMA will tell you to stop being an idiot, just as a human on Reddit or HN would.

[gregjor]

I stand corrected regarding Buzzfeed vs. Buzzfeed News.

The ability to imitate someone's tone (long a staple of satirical writing), or to fool a lot of unread people, no doubt represents a true achievement for LLMs. But people easily fool themselves and believe what they want to believe, so LLMs simply play into the limitations of human models and explanations for the world rather than representing a new form of conscious being.

I think we can see an analogy in craft production versus mass production. Today machines can produce (for example) furniture of very high quality, rivaling or exceeding the output of a craft carpenter or joiner. The mass-produced goods cost less to make and have good-enough quality, and can even sometimes fool people into thinking a skilled person made their table. That doesn't mean we should call those machines carpenters or conclude the machines possess the same skills. With LLMs we see mass production at scale come to writing and some other trades (law, customer support, etc.) that we like to think of as requiring actual skill. Maybe other trades such as programming will also fall to LLM mass production. But that happens because the jobs consisted mainly of rote and repetition in the first place, and require little creativity or intelligence to imitate.

[Llamamoe]

This just sounds like a just world fallacy thinking, that if people do a genuinely excellent job they'll surely win against $0 competition that improves every week.

Much of the existing LLM have been trained to sound dry and always the same and require creative promoting to get around that, but they very much so are capable of creative, thoughtful pieces based on increasingly long amounts of information and instruction you can provide.

[gregjor]

LLMs may have the capacity to imitate creative and thoughtful pieces, in the sense that their output can fool readers into inferring creativity and thought because we (humans) have no other model: When we encounter writing that makes sense we automatically infer a conscious and intelligent author, because our minds have no alternative explanation.

That capacity to fool people with apparently meaningful and creative content may have some value in the market, but I think that value amounts to the same value the market assigns to what currently gets churned out by people working in "content production." Not zero, but not the same value we assign to actual creative writing.

[chs20]

I'm pretty sure you can't reliably distinguish LLM text from real person text.

[gregjor]

I probably can't reliably distinguish LLM output from average writing from a real person -- the kind of "content" I find online. I believe I can distinguish LLM output from what a good writer produces. Anyone can write a Buzzfeed listicle or "review" of Madrid as pointless as what ChatGPT produces. But ChatGPT won't write Crime and Punishment, 1984, Lolita, Blood Meridian, or even On The Road or In Cold Blood.

People losing their jobs to ChatGPT don't even call themselves writers, at least not seriously. They produce content, most of it regurgitated the same way ChatGPT does it. If I produced content for online click farms I would worry about so-called AI. I don't think Murakami or Cormac McCarthy have to worry, but James Patterson might get replaced by an LLM (if that didn't happen already -- hard to tell).

In my own profession, some programmers already express fear for their future because Github Copilot and ChatGPT can write code just as good as they can. That should tell them to level up their skills, because if whatever they do can just get automated away by a glorified auto-complete they weren't adding enough value in the first place.

[mjrpes]

ChatGPT is pretty easy to spot (in some circumstances). Just look for text that has absolutely no voice and reads like a high school essay where what should be a 20-word answer is stretched out to meet a 750 word count. Example: https://www.quora.com/What-is-the-maximum-data-capacity-for-...

[olddustytrail]

That's just the default style. Try a request "...in the style of marvin the paranoid android" or "talkie toaster from red dwarf" or"Ernest Hemingway" or "Donald Trump" for a bit of variety.

[jfengel]

How well does it do with anything that's not just a pastiche? It's easy to throw in the tics of somebody famous, but can it actually do what a real writer does: focus on what's interesting, build a flowing narrative, target the material to the audience?

Exaggerated voices might give you variety, but I don't actually want to read a piece on the style of any of those voices. I would, however, be intrigued if it could write like me, or like any of the famous nonfiction writers who I've consciously patterned my style after (Isaac Asimov, Mary Roach, Cecil Adams, Michael Pollan)?

[actuallyalys]

I assume what they mean is the existing market will have to become explicitly labeled and advertise its advantages over AI-generated content.

[xg15]

> How much reshuffling of stuff already written a hundred times can we put up with?

I mean, if you look at the barrage of reboots, franchise installments and "memberberry" sequels that make up the bulk of cinema blockbusters...

[gregjor]

Good point. Every superhero movie looks the same to me. I always assumed that was due to my age. The thought of LLM-generated "Marvel Universe" sequels going on forever seems depressing, though I long ago learned to just ignore that Hollywood compost bin.

[xg15]

If it were just Marvel. My impression is, we'll soon have every movie that was moderately impactful in millennial childhoods be reanimated in some endless loop of seqels or turned into a "cinematic universe". Probably with LLMs streamlining and automating that process.

[fuzzfactor]

The problem might be quantity over quality.

If the quality is only half as good but the AI can put out orders of magnitude more material faster, that could be the direction that may be preferred by some content publishers.

[gregjor]

We can easily measure quantity, but how do you define or measure quality when it comes to writing, or imitation of writing?

[fuzzfactor]

I think that would have to be in the eyes of the beholder :\

Naturally, there's no accounting for taste.

Artificially, we'll just have to see what passes.

[greenie_beans]

this is what i've been saying. the good writers will become even more valuable. if everybody grows up not learning to write, then the ones who learn to write will be well positioned.

[mkr-hn]

As always, there's a good classic sci-fi story on this subject: https://en.wikipedia.org/wiki/Profession_(novella)

Except kids learn everything with tapes, updated and improved using new knowledge created by a rare few, beamed into their heads rather than depending on LLMs.

It's also a Stargate episode: https://www.gateworld.net/sg1/s3/learning-curve/