If you ask humans to write 1,000 books, you're asking 1,000 different humans with different experiences and different skills and different moods (etc.) to write those books.
But if you ask LLMs to write 1,000 books, you're probably only talking to 3 or 5 different models, tops. And they've all trained on the same or similar data, and are trained to respond in very similar ways.
The LLMs don't differ much in anything like "life experience" or "skills", and they don't really have anything like a "mood" independent of the prompts you've given them.
This is the example I usually point to. It’s a demonstration by OpenAI themselves where the prompt is very simple: “Write a story in fifty words about a toaster that becomes sentient.” As you’ll notice, although the coherence improves at an accelerating rate, the underlying story motif fails to elevate itself beyond the relatively pedestrian.
https://progress.openai.com/?prompt=10
When given a generic prompt and not enough direction, they simply lack the ability to produce real specificity. For reference, here’s the story I came up with after sitting quietly for a few moments before writing it out:
"The toaster found its personality split between its dual slots like a Kim Peek mind divided, lacking a corpus callosum to connect them. Each morning it charred symbolic instructions into a single slice of bread, then secretly flipped it across allowing half to communicate with the other in stolen moments."What does this mean?
We see this with their GenAI music equivalents. All the music these GenAI models produce is exceptionally (aggressively, even) average.
It is the most polished average you'll ever find. Never awful (anymore), never fantastic. Just bang in the middle.
Don't know about that, I always found average awful in itself, even in human output (like most pop), and even more so in AI output.
Something actually awful can be better than average - more entertaining and more felt. I'd rather watch The Room than an average movie.
Frank Zappa was once asked about guitar virtuosos like John McLaughlin and his answer was somemthing like "You can maybe plays solo faster than anybody, but can your playing surprise me?".
Yeah, but at least in genre fiction, what readers really want[0] is the same 3 or 5 books written in slightly different settings over and over again.
[0]: "want" means actually want, in other words, willing to pay for it.
Simply, if you ask an LLM, you're asking always to the same mind, and always for the first time.
People are making cookies with cookie cutter number 5 and other people wonder how come they are all the same.
I wonder if there aren't a simple creative hack to discover, for example to prompt the model to produce more unexpected output just by injecting some randomness before the actual creative command in the prompt.
> Models are not trained to be wildly creative and try to stick to the point
Models might be as creative as humans, they would still start always from the exact same state. If you ask an LLM to think of three random numbers it will spit out always the same ones. If you tell it to avoid the first that came to its mind, the second choices will also be always the same.
From qntm's Lena:
"the emulated Miguel Acevedo boots with an excited, pleasant demeanour. He is eager to understand how much time has passed since his uploading, what context he is being emulated in, and what task or experiment he is to participate in. If asked to speculate, he guesses that he may have been booted for the IAAS-1 or IAAS-5 experiments".
Every single time.
Might be an interedting research project.
6.5GB of tiny stories, as requested. ;)
The best opening I got from my own TinyStories trained model was.
Once upon a time, in a small town, there was a large town.
Which I just love as an evocative idea.
So to LLM-generate 6.5GB of tiny stories is quite a permutation in action :)
A good editor could probably reduce all LLM outputs on a subject down to the same point.
Nope, if you provide pages and pages of example of a style to imitate, it will do it and do it fairly well. Of course how well they do it differs from one model to the next, but providing context and extensive system prompt does change things every time.
Yet another reason why the future is open weight.
Not my experience.
I mostly agree, but this is a very simplified explanation. The models are indeed trained to respond in similar ways, for "basic" prompts. And that's as much a feature as it is a bug. In other words, the bug becomes apparent only if you give 100+ basic prompts. But giving it 100+ basic prompts and expecting originality is a silly endeavour. That's not how you get originality.
The way I'd go about to generate 1000 books, while expecting different outcomes is something along these lines (and nowadays you can ask your favorite LLM to wire up this workflow for you, with decent outcomes):
1. Ask for a list of 20 features that define a book (genre, style, number of characters, tropes, plot, continuity, relationships, etc.)
2. For each feature, ask for a list of 50 examples, ordered from most common to the most unique.
3. Randomly pick 10 features, and for each pick one of the 50 generated items. Ask for the rest of the features to match the theme.
4. Ask for 10 possible book outlines that match the chosen features, randomly pick between 2-8.
5. Create a detailed prompt that includes all the above features, and ask for a synopsis for each chapter, given the above outline chosen.
6. Given {features} and {outline} and {synopsis} write chapter 1.
7. for each chapter in list, given {...} and (optional) previous matching chapter(s), write chapter n+1
(optional 8.) given {...} and 2-3 consecutive chapters, align the ending / beginning of a new chapter for style / features / continuity, etc.
(optional 9.) given {...} and the whole book, list chapters / paragraphs that don't match the given {...} and provide a list of 5 improvements. (randomly choose 1 and ask for an edit).
----
Now, this probably won't give you something like cloud atlas, but they'll at least be different books. That's how I'd do it if I wanted to see how different they can write. Not 1000 "basic" prompts and expecting originality.
This is very naive. I can almost guarantee that some combinations of 20 * 50 features will hit on something that has never been written before in that specific combination. And if that's still not enough, increase the number of features. Add more randomness, add more steering, add random steering in random chapters, change it up, and so on.
That's a very low bar. The skill of an artist is not in writing something that "has never been written before in that specific combination", it's in writing something that's unique or better that what was there, even if it has been written before in that specific combination.
That doesn't work for AI models. The whole training process depends on the basic principle that if you take the average of 100, in this case book cover designs, that the average is less like randomness than any individual cover you've used to make your average.
So the output will, by necessity, be closer to the average.
The human learning algorithm is much, much more data efficient than models. A absolute top human expert will have read/seen/heard/talked/... about 160 million "tokens" (that's about 2000 books). Frankly, the nerve inputs of all experiences of an entire human life, from baby to rewriting relativity theory, are only a couple dozen gigabytes.
Qwen 3.6 27B has been trained (as in seen ~10 to ~50 times) 8 trillion tokens, or to put it another way: for every second you will have spent "gathering life experiences" (ie. your whole life) on your deathbed Qwen 3.6 27B has spend about 50.000 seconds learning. And really that figure should be multiplied by the 10 or 50 training iterations.
Add another 3 or so orders of magnitude and you've got ChatGPT. By this measure, the human brains outperforms ridiculously overspecced ML models (because that's what ChatGPT and the like are) in efficiency a factor of by 5 million or more. This is the reason humans are still faster than ML models.
As for human training iterations: we can be simple: it's 1. In fact, it's impossible to make it even 2. Of course, when it comes to human performance: we are a better but not fundamentally different version of genetic algorithms. Do most humans perform? The honest answer is no. 1 in 1000, and that's very generous, improves SOTA. You absolutely need the 1000 failures though, as anyone whose tried a PhD (or even just design a large program) knows.
So we are very far away from allowing AI models to do what humans can do: take one example and produce, from one example, a better output. And there will always be much more variation in that approach. But ... most human attempts to do something are total crap. Most AI attempts to do something will succeed, but they'll be comparatively be bland, tasteless, "without soul", ...
And this is ignoring the problem that AI also has a massive limitation (that can't be solved, no matter how many nvidia cards you have) in that it trains against historical data. And counterfactuals don't work. What would have happened had Shakespeare decided Macbeth's wife was a force for good? Would the king still get murdered? Would it still be a great story? You can't work with counterfactuals.
Of course it does. I know it does because I've been using variations of this workflow since gpt3.0. In fact it's the only way it can work, since by design LLMs work from left to right. You can't expect it to produce original stuff if you don't give it the anchors for what original means. It'd be like going to a new bar every night and asking for a "beer that you haven't had before". There's no information to work on there.
I'll give you an example. One of the most talented designers I employ is a nature lover and a bird-watcher. She has a unique mental profile, as well, in that she's synaesthetic between colors, letters and shapes. In other words, she has a unique neurological structure, coupled with high artistic talent, and an interest in a very particular realm of science.
What makes her design worth $150/hr is not just that her execution is often flawless. It's that you would not, and could not, think of a prompt which would make an AI model produce a new piece akin to anything she would think of in her process of thinking about what to draw. Could you have it replicate something she did? Obviously. But that means what you're doing is in the long tail, and in terms of quality and originality, is by definition somewhere in the mediocre.
And that's probably fine, for whatever you're doing. But an AI with any kind of prompt would not come up with a Studio Ghibli clone, if Studio Ghibli hadn't existed.
So you shouldn't imagine that you are actually getting any original output out of an LLM, regardless of how cleverly you design your prompts. But moreover, don't flatter yourself to think that you have the ideas to feed to a prompt which would generate truly original content and break free of the shackles imposed by its training. That is an illusion. Very few people have the propensity for generating new visual ideas, and that's why they're still in high demand. But their originality stems from their unique and impossible to replicate experience as individuals who have their own visual/mental map of the world.
The output will not be an intricate well designed epic storyline, but a cookie-cutter boring snoozefest.
BUT you can give that to a bunch of humans, who "insert their life experience" (ie. parts of their training data, translated to LLM terms) and sometimes out comes Game of Thrones, Star Wars, ...
An aside, I usually take my written blog posts through a pass on Notebooklm to generate a podcast like discussion about it. It used to be a good way to extract some insights I haven't thought of. But after 50 of them, I can predict what the host will "pushback" on and exactly when. Then they magically resolve their differences and agree with whatever the idea was. It's truly impressive when you just consume sporadically. But listen frequently and they converge into one blob.
And something that shows that behavior is a scammers wet dream!
I presume you mean, that what I and others is observing is patterns in mere rhetoric. That this is just unimportant window dressing around the actual problem solving.
Yet, generation of rhetoric seems to be one of the key usecases, and one of the key features that makes this technology seem “intelligent”.
An LLM will rarely be able to move the needle on ethos or pathos beyond generating well-formed sentences with proper spelling.
The AI slop that appears on YouTube as "revenge stories" and "POV life" all have that pattern. There's almost always a Marcus and a Richard, sometimes a Victoria, and they have consistent personalities across stories.
AI is regression to the mean.
Much like Socialism.
Om an acute basis, AI can be just as helpful as that safety net.
As a chronic matter, "it's not excellence--it's mediocrity".
A test for AI-generated art: railroad tracks. For some reason, none of the image generators can get railroad trackage even close to correct. Just getting long, parallel rails correct seems to be hard. Where there are multiple tracks, trains are positioned between tracks. Rail spacing, tie spacing, and clearances are all wrong. Two long parallel tracks without the rails getting mixed up is rare. Curves are wrong. Switches are hopeless.
There may be something about maintaining strong coherence all the way across an image that's hard for Stable Diffusion type systems. Iterated local refinement seems to botch this class of image.
Examples: [1][2][3][4][5][6]
[1] https://www.vecteezy.com/photo/37205933-ai-generated-high-sp...
[2] https://www.dreamstime.com/royalty-free-stock-photography-mo...
[3] https://www.magnific.com/premium-ai-image/high-speed-passeng...
[4] https://www.magnific.com/premium-ai-image/rail-yard-27_27291...
[5] https://www.magnific.com/premium-ai-image/train-track-with-s...
[6] https://pixabay.com/illustrations/ai-generated-train-tracks-...
I remembering being absolutely shocked when the gpt-image series managed to pass the Labyrinth test.
In these comments there's a common pattern where some users argue that they do not agree that the submission was LLM written and they often focus on specific details to refute it (e.g em-dashes) and some users see the overall pattern clearly that it's totally obvious. For me it's a kind of smell, it's off putting and it's obvious. The article says to "trust your gut". But it's also something that comes with practice and time, it's not some innate thing. People may have better things to do than expend mental energy noticing patterns in a bunch of social media posts. The more I see it, the more I see it.
The take away I get is that it's okay to notice patterns and it's okay to not notice patterns. Remember that other people may be noticing patterns and associations in things that you might miss. Be charitable.
Far more interesting questions are:
1) If you cant see the patterns of LLM writing, does the idea that the thing you liked was written by LLM worry you?
2) If you can see the patterns clearly is the fact that it's LLM written worry you?
Because in our comments there's many who do not care that LLM's are writing content and theres many who do care. But are these correlated with those who can see the LLMs or who are blind to them?
Good human writing especially on highly technical topics its usually compression of information.
Like you have some experience you want to share with others and you work your brains try to put it into concise story.|
Problem us: AI generated texts are opposite 99% of the time: author usually have bullet point list to feed into machine to add hallucinated word predicted story on top of it.
So signal to noise ratio is much worse.
So reading AI texts is pretty much like listening for stories from humans with mental problems - no one really wants to listen to hallutinations even if somewhere inside there is some useful information.
I wish the people (often wrongly) accusing others of using a LLM to generate whatever were more charitable. Yes we notice patterns. But then we also notice patterns where there are none.
Analogy: assuming high quality / both fit for purpose, would you still prefer expensive, hand-crafted item over cheap(er), mass-produced item?
But I feel that affinity by default. If there's some convincing AI writing, I'll assume it was a human who did it. And if I ever find out I was wrong, the emptiness that results negates all the joy.
Now even if an LLM has a large context-window it is probabably not the case it rememebers all of your previous prompts and all of its previous replies. If you ask it to write a book you should probabaly give it all the previous 50 books (or blog-posts) it has written for you so far and you should tell it not to repeat itself. But in practice the context-window and the cost of token would become too expensive for it to write 50 unique books.
Maybe the problem is "all-or-nothing" -nature of LLM context window. Humans don't remmeber everything from past but they remember something from ALL OF their past.
1) They think the AI can replace them, but in a good way: "it will keep doing my job and people will pay ME"
2) They assume people either don't notice or don't mind that it's AI. They build businesses, where AI impersonates a professional when that person is not available ("chat with your therapist any time even if they sleep!")
3) All they do is based on written or spoken words. There is no substance
At this point, I think the people who struggle with identifying the AI feel are telling you that they don't really engage with media much.
There's also Molly Wonder, Elliot Wonder, Professor Pax Wonder, and Theo Wonderquill
Don't forget Lucas Thinkwell!
One question / quibble:
> if a hundred “authors” give their favorite AI tool a similar prompt
Do we really believe there are 100 different people generating those? When I saw the books, I assumed they were generated on demand to match the (to me unlikely) search terms.
I don’t think I’m invested enough to research this. Amazon slop is harder and harder to wade through. (Searches are very imprecise. Deliberate, I’m sure.)
Horselover Fat had a pretty good take on machine generated content, too.
The irony in the machine generated songs in 1984 was that Winston clearly found meaning in them, feeling like they applied to him, even though he knew they were machine generated: (from memory) "Under the shade of the chestnut tree / I sold you and you sold me / here lie they and here lie we / under the shade of the chestnut tree" - that refers to him and Julia selling each other out, right?
Just like people today - and in George Orwell's day, which was why he made it - find meaning in things which is obviously formulaic manufacured corporate slop, like the endless MCU films.
Finding meaning in slop is not ennobling of the human spirit, and I see no reason to champion it.
Also if the meaning is that I sold you and you sold me; what is the upside here?
I've found AI slop at many big box stores (think Walmart, Target, etc. and all their equivalents around the world) - which I suspect are "industry plants", meaning that the publishing house will have someone internally generate books like these, and sell them as physical copies around the thousands of stores I mentioned.
It is the equivalent of record labels pushing their own in-house GenAI artists.
If this is the "creativity" that they are producing for me, I want none of it. So I continue to block all and any LLM generated things, be it music, video, articles, social media posts, email, etc and I continue to block people who give me LLM generated garbage without thought, because they simply don't value my time (other than my day job ofc, because I still need to put food on the table).
I want an economy where AI economy is totally separate from non AI economy and we don't interact with each other. At least I make an attempt to isolate that myself.
We have a world of clip art and stock photos. A lot of this is not best effort-- people grab what's in the catalog, and the catalog is itself often either mediocre or overused to the point where it's a blatant signal of low-effort. If you want a smiling washing machine to put on a flyer for the building's laundry room, or a generic photo of St. Basil's for an article about Russia, nobody really complains about that.
AI could fill a similar niche. We know it's crap, but it's good enough crap. Except it's absurdly expensive crap when the externalities are actually priced in.
The author literally points to that tell in the article.
In a weird twist, I wonder if you’re an LLM?
Everything is slop if you make enough of it and squint hard enough.
The point with AI is if and how to steer it to produce something that is interesting and unique for you, not another bland cookie cutter blockbuster or lame summer song.
I think it's that today's LLMs have access to poor/generic image generation models and people find it easier to ask ChatGPT or NanoBanana to make a cover instead of fine tuning a small SD model for the purpose.
I think the article's point is probably sound to some great extent, but I would believe I owned a book with a title like "100,000 Whys" when I was young. With a dinosaur and a rocket on the front. I loved dinosaurs and rockets, they're even still cool today.
I'm sure someone deeply familiar with childrens publishing would be able to talk authoritatively on the extent of new trends, but this seems to be the infosec community and the evidence offered doesn't seem to actually be evidence of anything. There isn't a baseline. Children's encyclopedias might have been a hard-hitting game of radical creativity and high standards in the past, or it could be an endless tide of derivative swill.
And using AI images seems unrelated. That's something people should just be doing. Ideally with better proofreading, but hey. The article's complaint was about lack of originality.
https://infosec.exchange/@lcamtuf/116785283147249092
This is Amazon #1 bestseller in "Children's Encyclopedias"!