GPT-3 writing tool that generates surprisingly compelling characters/plot twists | Hacker News Reader

GPT-3 writing tool that generates surprisingly compelling characters/plot twists

sudowrite.com

51 pointseliotpeper46 comments

This doesn't really show it in action. First off, the automatic scrolling over two separate panels makes it difficult to read the story.

Additionally, the page looks to be a few static snippets, instead of dynamically generated and unique to each user. That makes me suspect that these are probably cherry-picked examples.

Hey there, Sudowrite founder here. We're building on GPT-3, and the OpenAI is still early, so we aren't allowed to give public access just yet. So yes, these are static snippets for now. But the quality is on par with average results from Sudowrite.

I have a problem with procedural games in that whenever I face something new, I know it is not the process of a human creative mind. There wasn’t any intention behind that creation, there were no conscious or subconscious decisions about what that was supposed to mean, there was no human bias, no human tropes. Essentially, for me, there was no story. I lack any more understanding of the world, the humans, or the society for encountering that piece of creation. It is empty for me. Meaningless. A parade of random dice rolls thrown by a machine carefully designed to be misinterpreted as a story.

I feel the same about a story written by GPT-3. Or AI music creation.

To make myself clear, I think there is more depth in any superficial pop song composed by humans with the only purpose of earning money based in tried and tired formulas than it is ever possible to achieve in any music ever created by an AI.

For now.

So, suppose it were possible to simulate a brain. Say someone in the future figures out how brains work, and boils it down to some physical processes to do with neurons and synapses.

Ed Sheeran is then scanned, and a battery of machines produces new music from the patterns in his brain.

What do you think of that?

For me, it seems the creative process must be informational. It's something to do with mixing up experienced life, ideas, cultural tropes, etc. But in being informational, it's also not specific to the first substrate that generated it. In fact you have the luxury of being able to "run the code" on new substrates, and that creates a new economics of what is considered good.

I'm also fascinated by the application of AI like GPT-3 to generating video game content. The best content will come from game developers working with GPT-3, where GPT-3 comes up with characters and subplots without exhaustion, and the game developer filters the generated leads for ideas that stand on their own and ideas that invite more exploration.

It seems reasonable for game developers to be able to generate an order of magnitude more NPCs using this method. The game director can pour their heart and soul into the main storyline, and the assistants can create hundreds of hours of less original but still sufficient content. I imagine being able to create DLC after the initial release at a rate that exceeds any players ability to finish the latest release before the next comes out.

I'd argue there is also depth within the "random dice rolls thrown by a [carefully designed] machine", specifically in the design and choices that went into said machine's implementation. It's certainly a bit meta, but I enjoy this more than human created superficial pop songs.

Would it make a difference if you didn't know it was procedural? Games like Dark Souls have a story, but the majority of it isn't spelled out, it's up to you to fill in the gaps based on the environment and tidbits from characters you meet.

I used to agree with this (and I still hate 'procedural content' because of how absolutely soulless it is), but GPT-3 gets close enough for me (though it requires cherry picking). It's at least miles above 'procedural content'.

Like, read this story for example: https://read-the-samples.netlify.app/sample_99

It still has issues, but at its core, it has a lot of character and is quite funny to me. I'm not sure if GPT-3 truly has an understanding of its 'meaning', but it definitely can at least somewhat emulate such an understanding it.

It can be really cool when humans and programs collaborate on creative ventures. For example, the youtuber Kruggsmash [1] plays Dwarf Fortress and illustrates parts of the story as he goes along.

[1] https://www.youtube.com/channel/UCaifrB5IrvGNPJmPeVOcqBA

I know how superficially contradictory this may seem, but I disagree at an emotional level, however, I rationally see where you're coming from, and perhaps even agree to an extent.

For me, the 'process' of creation of a game, or piece of art, is interesting in and of itself. Whether a song was created for profit only, or inspired by the loss of someone close, is interesting on its own but it does not enrich the work, if that makes sense. The work in itself produces feelings, and those feelings would be there regardless of origin. In fact, I've found myself deeply moved by music, ascribing to it meaning and intent, that turns out is completely different from that intended by the author. That is what I mean when I say I emotionally disagree with you. If a work of art can produce feelings in us that the author never felt nor intended, can we say that the author matters in any special way? After all, we're post-Death of the Author.

However, for the rational part of me, the source matters. If for no other reason than philosophically: Why does art inspire feelings in us, even those unintended by the artist? Is it empathy? Even if misguided? If so, we could say art is devalued by a non-feeling creator: after all, what's there to empathise with? Or is it the recognition of something within ourselves? A (guided) self-introspection, if you will? Then the source doesn't inherently devalue the product; unless you're somehow opposed to introspecting based on the output of a machine.

I feel as if some of us have been so worried about sentient/super-intelligent AI outperforming us creatively, we haven't worried enough about what it would mean if non-sentient/'intelligent' AI outperformed us at our own creative process. Or even reached parity - that on its own would be monumental.

At first glance, I agree with you, but upon further thought, I remember a literature class we had in college and we were reading something and had dissected it to bits. We had extracted and applied so much meaning to a few passages and a few of us pointed out that all this nuance was almost certainly unintentional and that we were reading far too deeply into it.

But the professor's question was "does it really matter that this wasn't the intent?"

And that sort of changed the way I approached the literature. When you are reading, are you reading and studying it to merely learn about what the author thought and experienced and intended? Or are you reading to better explore your own thoughts, perceptions, and interpretations?

I would say, porque no los dos?

See, for me that supports OP's point. The reason I developed a distaste for this kind of analysis is precisely because it became clear to me that the connection with the author was tenuous.

I don't feel this kind of stuff is useless, it's just that I don't find 'death of the author' stuff remotely as interesting as actually figuring out /why/ the author wrote what they did, and /what/ they intended to convey, and /how/ they decided to do it the way they did.

I suspect GPT-3 type stuff will never be able to convincingly replace an 'author'. Games might show up that finally have interesting, generated side-missions, but people like myself will likely still prefer human-created versions of them. Not because the generated versions are in themselves inferior, but because the ones that involve a human will necessarily involve said human's choices on a level above the missions themselves.

(Of course, in reality things are much less black and white. In more than one game I've felt like the designer(s) used tools to generate stuff and then curated the results.)

How would you interpret the work of Brian Eno, or other algorithmic/generative artists? Steve Reich’s compositions are algorithmic (but not stochastic), and are widely considered to be masterpieces. Another example: Eno’s Reflection, which exists as both a 54 minute version and an endless generative ambient composition.

> The creation of a piece of music like this falls into three stages: the first is the selection of sonic materials and a musical mode – a constellation of musical relationships. These are then patterned and explored by a system of algorithms which vary and permutate the initial elements I feed into them, resulting in a constantly morphing stream (or river) of music. The third stage is listening. Once I have the system up and running I spend a long time – many days and weeks in fact – seeing what it does and fine-tuning the materials and sets of rules that run the algorithms. It’s a lot like gardening: you plant the seeds and then you keep tending to them until you get a garden you like.

> Moving the composition into software allowed an extra opportunity; the rules themselves could change with the time of day. The harmony is brighter in the morning, transitioning gradually over the afternoon to reach the original key by evening. As the early hours draw in, newly introduced conditions thin the notes out and slow everything down. [https://pitchfork.com/news/70448-brian-eno-details-generativ...]

To me, crafting a system that itself crafts endlessly novel music or art requires significant level of intention, understanding, and meaning.

> To me, crafting a system that itself crafts endlessly novel music or art requires significant level of intention, understanding, and meaning

Then the system is the art, not its output

The Great Automatic Grammatizator, by Roald Dahl, one of the best sci fi short stories ever. This post is the story come to life.

https://www.google.com/url?sa=t&source=web&rct=j&url=http://...

This app doesn't adequately disclose that it's using GPT-3/OpenAI API, therefore making the "our magical writing AI" subheader patently false.

The lack of transparency is going to be a continual problem with these kinds of apps, one I've tried to encourage with my own text-generating apps with only mild success.

Well "OpenAI" is itself a bait-and-switch of a name (nothing open about it). So, it fits.

Agreed, this is impressive but the base results from GPT-3 when given a creative prompt are super impressive on their own. Here's an example with zero additional prompt design using one of the same introductory sentences: https://gist.github.com/dylanpyle/b9ed1d7e26ed6e6247a0bf9290...

Hey Max!

Sudowrite co-founder here. We didn't leave out GPT-3 to deceive and will probably include it in the FAQ we're putting together.

Our reasoning to leave it out initially was more about getting the right audience for our beta test:

- Most people outside the tech world (our intended audience is fiction writers at this point) haven't yet heard of GPT-3, so we didn't want to confuse them.

- There's been a problem where people who want to try out GPT-3 try to get access to it through apps that use it. AI Dungeon deals with this a lot. OpenAI discourages direct user access to the prompt, and we've scaled back some of our features already to prevent unfettered access.

I agree it's a little confusing; I'd want to know what was powering the tool myself. Hopefully as OpenAI opens up access we can be a little more open as well.

> There's been a problem where people who want to try out GPT-3 try to get access to it through apps that use it. AI Dungeon deals with this a lot. OpenAI discourages direct user access to the prompt, and we've scaled back some of our features already to prevent unfettered access.

I find this incredibly frustrating as a user (of AI Dungeon at least). I'm always left wondering if my experience is being crippled by OpenAI's draconian requirement. Hopefully they relax that in the future once the API is out of beta.

This example speaks for itself (please leave a comment if your jaw drops).

> Rohit took a tentative step out of Indira Gandhi International just as a drone sliced through the space above his head, dispersing an aromatic mist. The summer night was hot, thick, and wet.

As always with AI-generated text, be aware that any curated demos are very likely cherry-picked.

Hey there, Sudowrite cofounder here.

Just as clarification, this example is human written text (I wrote it!)

When you click "Open Wormhole", you'll see the output from Sudowrite offering different directions you could go from this start.

When arriving at the site I thought what I saw had been written by sudowrite - you really should make clear what has been written by a human.

That's a prompt written by a human. But 'through the wormhole' I found this snippet that reads like a Black Mirror episode guest-written by Thomas Friedman.[0]

> The drone broke into four units, zipped around him, and whirred away. "He was escorted by a drone, a robot, the whole way," his father told me a few months later. "It was very Indian. It was very on-the-ground. That's the vision of India."

[0] - https://delong.typepad.com/egregious_moderation/2009/01/matt...

I don't think "generate a writing prompt" is a problem that people have. Great for exercises perhaps, but if you can't generate your own ideas, picking one at random isn't going to help. Mad Libs are fun though and they tend to be funny because they're surprising and often absurd.

These kinds of things are great for people who fancy they might become a writer one day. It's the same as obsessing over your editor or your outline method. Nothing's going to write your story for you but a lot of "tools" are available to grant yourself the illusion of progress.

This actually makes me wonder... has anyone actually tried applying deep learning to fiction plots, whether book, film or TV?

I assume not, since it would presumably require a kind of manual tagging of plot elements. ("A thinks B did C. B discovers D did E. A finds out about E...")

But I can't help but wonder if automatic generation of plot, then characters, then dialog... could actually produce fairly realistic screenplays, at least. (Screenplays, being so compact, would be much easier than novels with their lengthy prose.)

GPT-3 (and other Transformers-based models) do a surprisingly-good job at maintaining entity relationships over a long horizon, thanks to attention mechanisms.

I did something sort of similar, but really most similar to the "manual tagging" you described... I wrote an "Excuse Generator" that takes the form of "I wanted to do ___ but didn't because __ and I feel terrible about that because ___". The actual code takes a few more randomized turns in the structure than that, but ends of making some stuff that amazed me. Code: https://github.com/unoti/gpt2-excuse-generator

There isn't a symbolic representation of plot semantics. No one cares that A thinks B, they care why A thinks B - and they care even more about what it might mean for what happens next.

Plots aren't a series of events, they're a series of motivational challenges for the characters. The heart of a plot is how they respond to the challenges, and a good plot has a consistent theme which develops a coherent series of related challenges through superficially different events.

The mechanical specifics of the challenges are almost secondary.

I've done a fair bit of screenwriting, including taking classes in it, and that's really not the case at all. You're confusing plot with other things.

Plot literally does boil down telling a story in beats where each beat is essentially a character wants something, tries to get it, and either succeeds or fails. String together 80 of these and you've got a film.

Now obviously there is much more to writing, like character (which is what you're describing), tests of character, etc. Which is why I talked about plot and character and dialog.

But plot really is just "a series of events" that needs to be sequenced very carefully. Which is an important lesson for writers to learn, because beginning writers often focus entirely on character and motivation, and neglect plot mechanics, which makes for a boring, meandering, not-compelling-at-all screenplay.

80 beats of that will give you McMovie or McNovel, and they're popular enough. You can even buy beat by beat formulaic breakdowns of when specific plot twists and other events are supposed to happen.

But no one will care unless there's a coherent motivation linking the plot points together. To give an extreme example a character who tries to buy a spoon, followed by a car, followed by a laptop, followed by a kite, and so on for another 70-odd beats, isn't going to make for much of a story. It won't matter if they succeed or not - unless there's a motivation for all of those events that explains the action and puts it into a relatable context, which illustrates character through challenges and conflict.

But not too much, because readers/viewers need a mystery.

The mistake beginning writers make is trying to illustrate character and motivation through internal monologuing and introspection, flashbacks and infodumps, and emotional impressions rather than through plot actions. This goes too far the other way and you end up in lit fic where nothing much happens for pages on end, but it does it very beautifully.

Or - more often - self-indulgently and rather boringly.

Meanwhile novice genre writers follow the template with a call to action, refusal, and all the other stock steps, but it's still hollow and uninteresting because the character doesn't have any character, passion, or inner tension, and that makes them boring in a different way.

This is exciting to see. In the start AIs will be used as assistants to human writers, but I imagine that in the future they'll progress to the point where they're able to write on their own with little to no human intervention. If programs can author books/screenplays/etc. for free, that's going to revolutionize the world of fiction. I don't think we need to get to the point of AGI to see bots that can hold their own against human authors.

To get a real story you need AGI that has some understanding of what it's writing about. What GPT-3 churns out is really just pleasant prose about a theme, but it's meaningless/directionless and won't make sense beyond individual paragraphs. It doesn't really make sense within paragraph either, but we're able to read meaning into any string of five or six sentences so it doesn't jump out as much.

I'm not claiming that this is something that GPT-3 could ever accomplish, but I do believe there exists a middle ground between GPT-3 and full AGI that can produce creative works such as fiction novels.

Writing a novel that holds together requires a coherent model of how the world works, and if you've got that you've basically got AGI. An AI author that can write a compelling account of a person or institution's behavior over a non-trivial period of time could also serve as a real person's manager and have a conversation with a real person.

There's a scene in Robert Heinlein's Number of the Beast where his self-similar author character, Jubal Harshaw, dictates the first line of a story to a secretary, who happens to be an AI. While he's having a conversation, she finishes the story -- in his voice.

Heinlein can't be the only author to have that dream. Sudowrite seems like a step toward fulfilling it.

What I hope/expect to see in the next 10 or so years are triple A game titles similar to GTAV, that start off with a set of characters and a situation/problem, but let things develop organically from there. Not to mention realistic dialogue / realistic reactions from characters to dynamically unfolding story.

At one point I would think this is insane, but if they can get the "context" stuff working a little better I feel this would be insane... and somewhat plausible. I suppose you'd have to tie 3d models to the story though, that might be difficult.

Does anyone have a good estimate what the costs for this would be? I feel like affordability might be the make or break, since the compute cost can't be that cheap, but I think it's doubtful whether the results would be worth paying a premium.

I believe Dwarf Fortress history generation algorithm is able to create fantasy history records that rival authored ones.

Can we start it off with "Call me Ishmael" and see if it mentions typewriters in the next 700 pages?

> We believe AI can unleash our creative potential.

Seems a little ironic given that the rest of the page is about using the creative potential of AI to lead the human in a direction it didn’t see.

I think most human creativity is built around a seed of inspiration from outside sources.

In my experience serendipity and happy accidents are exactly what leads to the most creative and interesting outcomes.

This is a system that’s able to generate these inspirations/prompts for you in a way that’s more focused on what you’ve already written. I don’t see it as ironic because I don’t think it’s hijacking your creative potential, just augmenting it.

This strongly falls under the realm of "so preoccupied with whether or not they could, they didn’t stop to think if they should" for me.

At a time when writing is being undervalued and turned into "content" grist for the mill, the last thing I want to see is writing guided by an "AI" down the same old pathways.

Reply on news.ycombinator.com