How to Write Like Stephen King
Infinite monkeys at infinite typewriters with human supervision
Welcome back to Entertainment + Tech. Each week will cover an interesting way technology & entertainment are colliding and where things might go from here.
I’m setting up some interviews for future editions, so if you have any recommendations on interesting folks to talk to, let me know! You can respond to this email or reach out to me on LinkedIn.
New to the newsletter? Subscribe here:
Infinite Beatles
Imagine you could open up Spotify and listen to a brand new Beatles song. It’s got Sergeant Pepper vibes, but no one’s ever heard it before. After it ends, another new song comes on, this one more Rubber Soul. In fact, as long as you can keep listening, you can hear new Beatles songs.
Imagine, though, that the Beatles never wrote a single word nor played a single chord in any of these songs. Instead, a computer generated them based on their catalog and style.
Lots of thorny issues arise around whether you can really consider them Beatles songs, and more importantly, rights ($$$). Disregarding that, though, it’s entirely plausible to live in a world where a fan of the Beatles’ music never runs out of it ever.
Music, though, is generally more self-contained- each song is a fairly unrelated story. Would this work well for books or movies? Would a never-ending Harry Potter or Star Wars be as meaningful? I think these series need an end, so that we can build to a true climax, but maybe we could expand the universe infinitely with other stories. Or maybe it would change how we think about meaning in content.
We’re heading towards the world where these things are not just technically possible, but technologically easy. Down the line, we’re going to have to figure out some new dynamics around the content we consume and the people or systems that create it.
Until we get there, we’ll play out these questions in smaller ways.
Back to Tools
A couple weeks ago, I wrote up some quick thoughts on AI. I actually have a lot more thoughts on the matter, so I wanted to follow-up. [1]
As I said in that issue:
Even with AI tools, there are four big ways humans still need to be involved (for now):
Construction
Direction
Curation
Revision [2]
I want to explain how these are creative functions in themselves.
Construction
Whoever builds the tools helps define the art it creates. TikTok’s AI editing tools enabled new creators, and they also bias the kind of content that gets produced. If you make something easier, you get more of it.
Direction
I choose the term direction here because I think of a human using AI tools as a partnership similar to that of director & actor. The director has a vision for what should be, and the actor tries to make that vision real through their own form. It’s from the two working together that art happens, often in ways neither could create alone.
AI tools are like that too. GPT-3 doesn’t always give you what you expect. You have to learn how to direct it, and you have to learn to work with what it’s best at–i.e. based on the construction of the tool.
Playing around with GPT-3 with friends, it was relatively easy to produce short examples like company slogans or long text samples about a particular topic. However, it doesn’t really understand what’s real or true and struggles with many types of problems you might ask of it. [3]
We got ambitious and tried to get it to generate a Weird Al-style parody of Wrecking Ball about Joe Rogan - a task with complex structure and tricky relationships between multiple entities. As we expected, it failed miserably.
Curation
I’ve been a fan of Janelle Shane’s work with Neural Nets for a long time. She trains models to generate funny examples of things like paint colors or candy heart messages. The actual funny one here is not the model, though–it’s just blindly fulfilling an optimization problem. It’s actually Janelle, picking out the best examples to show off to the world. [4]
Many models have no notion of what’s “good” or “bad” content. They’re just going to churn stuff out that fits the pattern–kinda like famous author REDACTED.
Likewise, all the posts you see about GPT-3 examples, or the blogs where someone sneakily used GPT-3 to write the post itself, [5] are curated to just the interesting ones. If the blog post was just crappy, they wouldn’t have posted it or it wouldn’t have gotten popular, [6] and you never would have seen it.
This has been the most common way to create with AI so far. Lots of the use cases here also tie in with revision.
Revision
This will be the most important case in the near future. Many of the examples people post from GPT-3 are ok, but not incredible. With just a little work, though, a lot of them could be great content.
As I mentioned, I tried to generate a Wrecking Ball Joe Rogan parody, and the AI failed miserably - at producing a parody. Turns out, it did give us a lot of lines about Joe Rogan. Some of them were actually kinda funny, and some of them were interesting phrases to build on. With more of the model’s mediocre lines, we could revise them and churn out a decent parody quicker than coming up with everything from scratch.
In this case, the human knows how to manage the complicated parody relationship, which is hard for the AI. The AI, on the other hand, can be an idea machine, pushing out tons of options for the human to consider, improve, and fit to the song. Working together, the AI makes the human faster.
As an example, the Guardian used GPT-3 to write an article about AI. In the post-script explanation, they say:
GPT-3 produced eight different outputs, or essays... We chose instead to pick the best parts of each, in order to capture the different styles and registers of the AI.
Some people have been pointing out that this means the content of the article, especially as it purports to be about AI’s power, is bullshit. I think that’s a fair take, but more interesting to me is this later note:
Editing GPT-3’s op-ed was no different to editing a human op-ed. We cut lines and paragraphs, and rearranged the order of them in some places. Overall, it took less time to edit than many human op-eds.
Stephen King writes a lot of books. When he’s working on a manuscript, he writes 6-8 pages a day, putting out a book’s length in two to three months. Most people can’t write as quickly as Stephen King, though. Could a writer of average skill using an AI tool to generate ideas or fill out first drafts of pages work at that pace though? Seems plausible. Most people find it easier to edit than create something new. [7]
Some might balk that someone who isn’t a good writer in the traditional sense is now creating books they wouldn’t have been able to before. If the finished book is good though, it just means that the work of creating books has been opened up to a new type of author - the writer-editor.
Just watch out if the AI starts printing “All work and no play makes Jack a dull boy.”
Contentapalooza
We’re heading towards a content explosion. First, it will be with human input as we scale our efforts with better and better tools. Eventually, it might be machines created by machines creating content for us, and it might be better than anything human minds could have designed–at least they’d still have the human touch of whoever originally set all these systems in motion.
I don’t think humans will ever be fully cut out of creation though.
For a long time, people thought AI would never beat humans at Go. Then, AlphaGo happened. Sure, it beat the top ranked players in the world, but it mesmerized people with the inexplicable & brilliant Move 37. Even more importantly though, that move pushed Lee Sedol to reciprocate with the equally impressive Move 78. It might be in collaboration with AI that we push things further than ever thought possible.
————
[1] This won’t cover everything either, so don’t be surprised to see more in future weeks.
[2] I changed this from “editing” in the previous post because 1) parallelism and 2) curation is also a form of editing, so this is more precise.
[3] PhilosopherAI.com is a fun tool if you want to see for yourself. Here are a couple AI takes on the next billion dollar idea in entertainment: 3D Movies, Minecraft + The Sims, Starting a cult or enslaving an entire country, Autogenerated sheet music, Virtual worlds
[4] There’s an interesting element behind the humor here. The aleatoric nature of what comes out means we sometimes laugh at something we wouldn’t laugh at a person coming up with (just like most improv jokes aren’t funny out of context). It’s why the nonsensical ones are also funny. We partially laugh at the way this ridiculous thing was generated.
[5] Don’t worry, this isn’t one of those gotchas where the whole post was actually written by GPT-3 - I accidentally lost my API key.
[6] I’ve written already about how distribution & economics change the content, but Coming Soon™: how you, the beloved audience, shape content.
[7] Jonny Sun’s take on why that is. Often, people let their critical voice block them from putting ideas down in the first place, rather than reserving judgement and editing it to be better once they have something down.
Thanks for reading Entertainment + Tech. I’d love to hear your feedback and ideas. You can respond to this email or reach out to me on LinkedIn. If you know someone else who would enjoy this newsletter, please share it with them!