Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
A short sci-fi story written with GPT-3 and illustrated with DALL-E 2
125 points by andreyk on Aug 30, 2022 | hide | past | favorite | 36 comments
Hi there HN,

Disclaimer: I've submitted a Show HN as well as the link for this general project before, but particularly like this one short story so want to submit one for it specifically. Hope it's not considered spammy!

I and a collaborator who writes sci-fi just released the short story "The Great Filter Button" - https://storiesby.ai/p/the-great-filter-button

Here's why it's relevant to HN: most of the text for it was generated by GPT-3 (with human curation, using SudoWrite) and it was entirely illustrated using DALL-E 2 and MidJourney, and a bit of DreamStudio aka Stable Diffusion (of course with human selection of prompts) AND it narrated using neural voice synthesis (via BeyondWords). And I think it came out very well!

To my mind it is a pretty good example of how the newest commercial tools by powered by learned media synthesis models can be leveraged by humans to make art. It also shows some of the limits: DALLE-2, MidJourney, and Stable Diffusion all have trouble with more complex prompts and don't follow various aspects of them, the voice synthesis is still pretty robot-y, and the GPT-3 written parts are heavily guided by human text (and the whole story is quite short).

We plan to keep exploring these realm of human-AI creative collaboration by releasing weekly short stories with this newsletter, and would love feedback, suggestions, or even entire submissions of your own creative work done using AI. Feel free to just comment here on HN, email us at contact@storiesby.ai, or comment here - https://storiesby.ai/p/submit-your-stories-ideas

Last thought: even with AI doing the "heavy lifting" of writing and illustration, a great deal of creative decision making is still left up to us with respect to subject matter, style, formatting, etc. I hypothesize that sturgeon's law will remain true in the age of AI-generated text/images (most of everything will be crap), and the job of literary agents, producers, etc. will just become far more involved. Sort of like A24 is mostly a distributor (and to some extent producer) yet have made a huge name for themselves - this may become the norm.

Edit: wow thanks for feedback HN! To the comments saying this is at best a mediocre story/outline, totally agree. Since we want to put something out weekly, these stories are quickly generated with the intent to be a neat example of human-AI collaboration, rather than with the intent to reach the bar of published sci short stories. Maybe one day...



Clickable links

https://storiesby.ai/p/the-great-filter-button

https://storiesby.ai/p/submit-your-stories-ideas


Capabilities like the textual inversion post from yesterday is going to really push story illustration into high gear.

https://github.com/hlky/sd-enable-textual-inversion

https://news.ycombinator.com/item?id=32643564


Knowing it's an AI that wrote it, it seems impressive. (Although I'm not sure how much human guidance it has).

The premise is I think actually pretty good -- but maybe the premise and basic plot really comes from the human-written parts?

The actual writing? If I didn't know an AI had created it and thought a human had, I'd say the writing is pretty terrible. At the sentance level and at the story level, it's just pretty poorly written and hard to follow, with weird choices.

But... sure, maybe it's too much to expect an AI to write a well-written story at this point, the fact that it seems to be a somewhat coherent story at all is the point? OK if so, but yeah.


This kind of summarizes my impression. I'd add that the story has a style I've come to expect from GPT-3. Basically, sequence of superlative statement alternating with fairly vacuous "softeners" - "I was in the lab the day that the greatest discovery in the history of science was made (though I acknowledge that I may be biased)." ... "But life did develop on one planet, which evolved one of the most diverse biospheres we had ever seen..." If it was a human writer, I'd say that the breathless tone makes it exhausting to read. And the lack of simple, "fleshing out" sentences means that, for example, it talks a lot of being excited by testing "their theories" but doesn't ever say that those theories are. And lots of human-written stories also do that. Many science fiction writers don't know modern physics in any detail and don't work out their science-variations in great detail. But naturally this system doesn't know anything but associations between words and makes up for this lack by using those associations quite effectively.


> it talks a lot of being excited by testing "their theories" but doesn't ever say that those theories are. And lots of human-written stories also do that

Good news, in a few years you can open your favourite book in GPT-3's successor and ask it to flesh out those theories. When AI will be able to write stories, all our pet peeves can be tried out, in any context.


The first sentence you quoted was part of the prompt

(But yes, I thought the human wrote like an AI too. If nothing else, it achieved stylistic consistency)


wrt "(Although I'm not sure how much human guidance it has)" - at the top it says "Text Formatting: Human-written text is italic, AI-generated text is normal" so that kind of demonstrates the ratio [and the human selects the AI completions among a few options each time]

And yeah, generally agree - as we say on our About page, these stories are sort of explorations/demonstrations of what AI can do, as much as they are attempts at entertaining writing. So in a sense we want to keep the "weird choices" AI makes to let the story go in weird directions. Since we release weekly, the intent is in general not to go for a lot of polish but rather for a fun concept and just an entertaining read.


Fair. But i'm not talking as much about the story "going in weird directions", as I am about the way the sentences and paragraphs are constructed.

Or apparent contradictions like "We found that planets with life on them were rare" followed by "But life did develop on one planet" -- so is that "rare" or just one? "which evolved one of the most diverse biospheres we had ever seen" -- wait, so how many "biospheres" (implying life) have you seen, again? "one of the most... we've seen" is a weird thing to say if you've seen only one, or even only a few! It's just like, wait, what are you trying to tell me here, what parts of these sentances serve the story? Is the feedback I'd give if this were a writing workshop,

I'm not talking about the way the plot developed, I'm talking about the way the words are constructed into sentances and sentances into paragraphs. It just seems... sloppy, like someone wasn't trying very hard just kind of throwing words on a page without thinking very much and not editing. But it is all understandable more-or-less grammatical English which is a coherent (if poorly-written) narrative, so, sure, that's a thing.


To your last paragraph, yeah that's kind of the case. We want to use the AI generated as-is more or less to showcase what it produces, so editing those sentences you mention to be more consistent with each other would mess with that. Plus, trying to finish a short story once per week while working a full time job (we currently have one main author on the team) kind of necessitates rushing things and not polishing that much.

So the intent is not to write a particularly well-written narrative so much as a fun/entertaining read that showcases what the AI can bring to writing today (even if it's flawed, as you say). Perhaps we'll aim for more polished writing or more human editing in the future, we are still figuring things out as we go!


Well this reminds me of the priming that is in AI Test Kitchen https://aitestkitchen.withgoogle.com/how-lamda-works

I mean, the example with "I'm at a planet made of marshmallows" and how the generated text changes from the default I guess style to a bombastic description.

I don't have an example of that in GPT-3 playground right away, but I suppose it works similarly and could be used to nudge the AI writing style in some way.

This is also why I'm not holding any expectations with offering you a story idea; the AI can come up with a story idea, in my experience, once I did not start with "You wake up in a cart in Skyrim" but with mere "You wake up in a carriage" and the AI started to improv right away, "A man in black robe enters the carriage", and I let it continue, chiefly on its own. Though I haven't saved that story, I will remember it well - the mysterious man in black robe locked up the protagonist with a book containing every knowledge of the universe, telling that after a year he'll come back to ask questions; and he came back after a year and asked Deep Philosophical Questions (TM), and the protagonist had not to shabby answers to them; among them stoically accepted death. Subsequently, the mysterious man took the book and left, leaving the room no longer locked. The protagonist traveled to the nearest village and settled to a work as a teacher. Fast forward twenty years later, meeting the black robed figure again, this time it was revealed to be Azrael (the angel of death); only then the protagonist felt truly the fear of death and that changed at least one of the answers to the Deep Philosophical Questions (TM) from the earlier part of the story.

Feel free to use that as a prompt; I doubt the AI will recreate that story with a good writing style with just that, but AI are capricious and one never knows until tries.


I am not critisizing you, I am giving my own review of what the AI you are using has or hasn't accomplished in this case!

I agree that you manually editing to be better written wouldn't have much purpose. I think the point of this exersize is probably "See what an AI can accomplish?" I was reviewing that.


ah I see, I did not get that it seems. Cool, yeah I think your analysis of the AI's output is quite solid.


> maybe it's too much to expect an AI to write a well-written story at this point

You just gave a great example of feedback. This kind of data can be used to train the model to generate self evaluation, maybe a few thousand samples would suffice to fine tune a large model. Then we could condition on good feedback.


I am not sure what you mean. You are talking about literally taking my comment as written verbatim and somehow feeding it back into the model? Can you say more about how this works?


I feel it's the typical HN reaction to be negative about generated content. I enjoyed it. I think it could have used another pass on editing but I felt the idea was interesting and I wanted to hear it through.


I did a more modest version of this as a handful of simple children book-like stories [0,1 2]. I had just gotten access to Dalle and was so excited about how thr future could look like.

I can't wait for a future where MMORPGs have characters, scenarios and landscapes filled in by AI. The users and the developers can give the skeleton of the in game lore, and then the AI tools can fill in with beautiful detail.

[0] story of a seamstress, https://twitter.com/RamonDarioIT/status/1552152241959669760?...

[1] story of a bnb host, https://twitter.com/RamonDarioIT/status/1552162769176109056?...

[2] story of playing chess against Magnus, https://twitter.com/RamonDarioIT/status/1552164214462025728?...


Don't take this the wrong way, but in my opinion, the story would get a form rejection letter from a SF magazine, but with a handwritten note saying, interesting pictures!


True, but then I get form rejection letters from SF magazines too. I have a nice collection of them! That bar is pretty high.


It's been done before, multiple times, Rick and Morty even did a take on it ( https://rickandmorty.fandom.com/wiki/The_Ricks_Must_Be_Crazy )

I guess that, these ML approaches are trained on lots of existing text, and therefor derivative works are to be expected?


I have never had so much fun and so much interest in machine learning until now. My new favorite thing to do when feeling down or stressed is to fire up Dall•E and go nuts. I buy credits like they are farmville coins.


Reads like a pretty decent outline. It's a sufficiently interesting premise for a story. But no named characters or even characters at all seems like a shortcoming. Understandable, as I don't believe GPT-3 is anywhere close to being able to create consistent, believable fictional people and relationships that can exist across paragraphs just yet, but I'm sure some future version will get there eventually.

That is, of course, the hardest part for human authors as well.


I understand people comparing this to some human written content but let's not forget how far this has come and at least in my estimation, how far it could, and probably will, go.


Without the human written parts it's an incoherent word salad. As an intelectual worker, I feel relieved - for a few more years at least.


Fear not, GPT-4 will soon be out ;)


This isn't a valid Show HN. Please read the rules: https://news.ycombinator.com/showhn.html.

We've taken "Show HN" out of the title now.


Got it, my bad!

This project was partially inspired by this Show HN "I had some time yesterday so I made a GPT3 podcast to help you sleep" https://news.ycombinator.com/item?id=29428910

So it seemed like the same sort of thing and therefore valid.


Our A.I stories and illustration generation service is live and installed at Computer History Museum in Silicon Valley:

https://www.youtube.com/watch?v=tH_qLoWMgKY

Contact me if interested in investing/customers: https://www.linkedin.com/in/kashlik/


Seems like AI generated art will give everyone access to decent art. GPT-3 and such I see as only useful to create spam on the internet by flooding it with tons of useless content.

Theres a huge gap between these two technologies. I guess our mind helps the art because it doesn't need to make sense, but a story has to be consistent with itself and I just don't see how that is possible without actual understanding.


It was pretty impressive especially AI that completed the plot in intresting direction ,but image illustration was still not at that level ,I bet a human would do better here , images were too vague and out of context


I thought it was amazing. Especially considering it's the equivalent of a day old baby crafting it. This is truly exciting, terrifying and intriguing in equal measures. It's all moving so darn fast! :)


How much text was edited away to leave this piece? The images are nice, but both the images and text have only partial coherence (which is to be expected, of course).


https://livebookai.com/category - tons of AI generated stories


This is horrifying. It's like the AI is hinting at what it will do once we run a few more experiments with it.


Yeah but they're pushing it to do this. The plot is human-generated. And in any case it passes the buck onto the inside universe to come up with the excursion that escapes endlessly outward.

It's a little like Rick and Morty. I think they have an episode about this with energy, infinite recursion then something happens then you get excursion.


Think of GPT-3 raw output as a markov chain operating on paragraphs and contexts instead of tokens. You can't write a coherent short story with it without your own "inpainting".

Check out https://novelai.net/ if you're looking for ways to preserve continuity.


[flagged]


Be kind.




Consider applying for YC's Summer 2026 batch! Applications are open till May 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: