Almost Timely News, June 11, 2023: AI, Creative Writing, and Fan Fiction

Warning: this content is older than 365 days. It may be out of date and no longer relevant.

Almost Timely News: AI, Creative Writing, and Fan Fiction (2023-06-11) :: View in Browser

Almost Timely News

👉 Watch my brand new keynote from Chicago, The Intelligence Revolution, all about how generative AI is the end of marketing as we know it »

Content Authenticity Statement

100% of this newsletter was generated by me, the human. When I do use AI-generated content, I will prominently disclose it.

Watch This Newsletter On YouTube 📺

Almost Timely News: AI, Creative Writing, and Fan Fiction (2023-06-11)

Click here for the video 📺 version of this newsletter on YouTube »

Click here for an MP3 audio 🎧 only version »

What’s On My Mind: AI, Creative Writing, and Fan Fiction

Okay, let’s talk about exactly how large language models work, especially in the context of creative writing. I was invited recently to join a group of fanfiction authors in protesting the use of AI to consume people’s works without permission, and they had a lot of good questions that I want to try working through in one big shot, or in fanfic terms, a big one shot. While today’s issue will be specifically about fanfiction, everything we’re going to talk about also applies to any kind of content on the Internet, including your marketing content.

We will put aside the can of worms that is the questionable legality of fanworks (which is an ongoing legal debate, but in general leveraging other people’s intellectual property without permission means you cannot exert intellectual property rights of your own) to talk specifically about how large language models work, because there is a lot of confusion about them, and once you understand them, you’ll be able to make more informed decisions about how to use them.

Let’s take a moment to clear up some terminology. A model is AI terminology for software. Microsoft Word, for example, is software written by humans, for humans to use in writing. A model is AI software, written by AI, for AI to use in writing. What we think of as AI software – tools like ChatGPT – are actually interfaces to AI models. So what are models? They have arcane names like GPT-4, StableLM, LLaMa, Vicuña, MPT, etc.

Why does this matter? Because different models are good at different tasks. Some models, like GPT-4 (the most powerful model that powers the paid version of ChatGPT), are massive models that are good at many things, but excellent at very few. If you’ve ever tried to write some mature fanfiction, you know that GPT-4 and the ChatGPT interface simply won’t let you. It’s like the Disney of writing tools; anything remotely controversial simply isn’t permitted. But that’s not the only game in town. There are other tools like Bard, Bing, etc. that are very consumer friendly, and then there are specialized models and interfaces like GPT4ALL.

For the rest of this piece, we’ll be talking about the models themselves, as opposed to their interfaces, because it’s the models that do all the work, and the models are where controversies lay.

First, let’s talk about how these large language models are built. We will leave out tokenization and embeddings and all the fancy tech talk in favor of framing everything in terms of… pizza.

How do you make a good pizza? There are a gazillion different ways to make it, but generally it’s a dough with tomato sauce and cheese, baked in an oven. Of course there are variations like alfredo sauce pizzas or dessert pizzas, but most of the time, when you say pizza, you’re thinking of flat round dough with tomato sauce, cheese, and toppings.

Now suppose you wanted to teach an AI how to make pizza. How would you do it? You would find as many pizzas as you possibly could and feed them to AI to learn from all those samples. Over time, with enough pizzas and enough time, a machine will learn that generally, pizza is flattened, round dough with tomato sauce and cheese. It will also be severely overweight.

You’ll have a lot of variations, like whether to add sugar to the sauce or not, or what kind of cheese to use. You’ll have big groups of common variations like Chicago deep dish versus New York slice. And what is the machine doing? Taking all those pizzas it ate and boiling down those dining experiences into recipes.

This is a key point to understand about large language models. The source texts themselves are not in large language models, any more than a cookbook of pizza recipes is made of actual pizzas. When a company makes a large language model, what’s being stored are the statistical probabilities of words and their relationship to each other – in essence, very fancy recipes, but not the original texts themselves.

So if source texts are not in large language models, then what is in there? The frequencies and statistical relationships of words to each other. For example, take this passage from a fan fiction piece:

The two women walked through the expansive lobby, decorated with the latest scientific achievements from around the company. Nearly lunchtime, the lobby’s enormous glass walls had rotated open, allowing fresh air into the building and open air dining at the ground floor employee cafe. Lena’s stride skipped a step as a notification pinged on her smartwatch. She held it up, glanced at it quickly, then took Kara’s hand and squeezed it three times. Kara nodded silently as they made their way to the elevator bank. Once the doors closed, Lena swiped her ID against the scanner and the elevator began its descent into the basement levels of LuthorCorp.

Is this exact text in a model like GPT-4? No. What is? The first stage of processing is something that looks a bit like this:

Tokenization

The colored boxes are how an AI identifies the pieces of text to start collecting data on, which leads to converting the text into this:

Vectorization

And then this becomes a series of math equations, shorthand for the recipe needed to replicate similar kinds of language to the source material. That’s what the machine has been fed – not even this series of numbers, but a series of math equations distilled from it. It knows, from this small sample, that the word Kara and the word Lena co-occur a lot AND are related to each other statistically, but the source text isn’t in the model – and critically, the model has absolutely no understanding of the text. Machines can’t read. They can only calculate probabilities.

So, back to our pizza analogy, once the model has a huge index of pizza recipes that define pizza, AI companies then test it repeatedly, asking it to make different kinds of pizza in a process called supervised fine-tuning. When it makes a sphere of dough, model makers tell it no, that doesn’t fit the model of what a pizza is. When it puts pineapple on pizza, model makers argue with each other for hours in the office before grudgingly allowing it. The machines eventually end up rewriting their cookbooks many times over as they receive feedback that what they thought was a good pizza was not, in fact, good pizza.

Supervised fine-tuning helps large language models do less of the wrong things like generate gibberish or problematic text, and more of the right things. This process requires a ton of human supervision and weeks, if not months, of fact-checking over and over again that the machine is doing things correctly or not.

Finally, model makers for the most sophisticated models do what’s called reinforcement learning with human feedback. In pizza parlance, customers give feedback on machine generated pizzas, about which pizza best matched what they asked for and how it tasted. RLHF is a very time and person intensive process, but this eventually leads to pretty good pizzas.

This exact process is how LLMs work. They are fed with huge amount of scraped data to learn what to do, essentially boiling down that text down into statistical probabilities that represent the texts in the same way that recipes represent the foods they’re intended to make. Then the models are fine tuned to make fewer mistakes, and then they’re given even more feedback from real world usage.

This is why things like copyright and sources are so difficult to untangle when it comes to large language models, because not only are the source texts not in the model, but the exact word for word probabilities are not in the model either due to the tuning process. In our pizza example, suppose as a joke you put up a recipe for pizza made entirely out of paper. Supervised fine-tuning would end up disregarding many of the probabilities generated from that terrible recipe because, yuck, and thus what’s in the final model would only incorporate fragments of the original probabilities.

A few folks have said, “But a sentence I wrote came out of ChatGPT exactly. That shouldn’t be possible if my writing isn’t in there.”

There are two reasons this will happen. First, you wrote a sentence that is common enough that the machine would naturally generate it. If you write something like “and they lived happily ever after,” that’s not exactly original text. There’s a very good chance any generic phrase you use will show up in generated text because it’s so common.

Or, second, you wrote a sentence that is so unique that when you prompted it for a completion, your prompt was so specific that the model only had one set of probabilities to draw from, and that was a fragment of your source text. For example, if you wrote a prompt to generate a fanfiction about a rare pairing, and there were only 3 fics on the entire Internet about that pairing, there’s a good chance any output the machine spits out for that are going to have recognizable patterns because there wasn’t that much to learn from.

Now let’s talk about AI and its ingestion of content, particularly from sites like Archive of our Own. Can AI companies scrape copyrighted texts without permission? Well, it depends. What does it depend on? The laws in your locality. For example, the law is not settled in the USA. In the EU, the law is leaning towards prohibiting the scraping of copyrighted works without permission. Japan just ruled that AI scraping does NOT violate copyright, in an attempt to seize a leadership position as a country favorable to AI development. Obviously, this then becomes problematic for jurisdictions; Japan is hoping to attract AI companies by allowing them to build models in Japan (along with the jobs needed to do so), and then those models could be used legally and safely anywhere in the world as long as they originate and run in Japan.

Is it stealing, as some allege, to use copyrighted or original works to train an AI model? Let’s go back to our pizza analogy. Is eating a pizza at someone’s restaurant and then going home and trying to write down what you think the recipe is to make that pizza stealing? It’s hard to say. Is it imitating? Absolutely. Is it plagiarism? Depends on how good the copy is. I can try to paint the Mona Lisa but it will be so bad, it will be a unique work of art. Bad, but unique – and no one will mistake it for the Mona Lisa, I can promise you that. Is it stealing? No, because theft deprives the owner of the use of their work. Intellectual property rights infringement isn’t theft. It’s a different category of law entirely.

Now, let’s talk creativity and voice. Is the writing that AI spits out less creative? It depends on two things: the model and the prompt. Really big models like GPT-4 are jacks of all trades, masters of none. They tend to create more generic sounding text because they have to be everything to everyone. A model like MPT Storywriter is tuned for writing long form works, and thus it will spit out better quality fiction for the same prompt.

The second factor is the prompt itself. Prompts are a form of programming language, except that instead of a formal language like Python or JavaScript, you’re writing in English, Spanish, Japanese, etc. But that also means that your prompts need to be as detailed as creative briefs to get good results. Imagine asking a writer to write a piece on commission, with a catch: you are not allowed to contact them again until the work is done. How detailed and thorough would you need to be to get good output, if you couldn’t ask or answer questions? That’s how detailed you have to be when writing prompts for AI.

Writing a prompt like, “write a love story between Kara Danvers and Lena Luthor” is going to get you completely uncreative, boring prose at best because there’s so little specificity that our large language models – which are probability engines – will simply pick the highest probabilities to assemble a story, and that story will be boring. If you submitted a three page prompt with a detailed outline, character descriptions, plot twists, style guides, and writing examples, you’re going to get a much, much better, more creative story. (download my free prompt guide here, no forms to fill out, no strings attached)

Do you want a machine to sound exactly like you? Then you’ll need to do something called fine tuning, a very technical process where you take all of your writing, prepare it in a format compatible with an existing model, and essentially tell the model to sound more like you. If prompt engineering is telling a model exactly how you want your pizza (and it getting it right to lesser or greater degrees), fine tuning is giving a model a bunch of your pizzas and saying “make pizza exactly like this”. It’s much more precise because there are a lot of little details and unique touches to our writing that aren’t in other people’s writing, and it’s hard to put those details into words. Giving exact samples is a much more reliable way to capture our voices. For now, the process to do so is fairly technical, but that will likely change over time as tools get more friendly and easier to use.

Can machines be truly creative? Six months ago I would have said no, they can’t, because they can only draw on what they’ve been trained on. However, there are two nuances here. First, large language models have effectively perfect memory. The more specific, precise, and focused your prompt is, the more likely it’s going to have creative output, output that hasn’t been seen before. Second, there is new research that just came out in the last month for tools to harness the mistakes made by AI as a form of brainstorming. Sometimes AI tools simply screw up – but researchers now argue that these screwups could have potential as a force of creativity, a way for machines to be creative.

Okay, so this was a LOT to run through, but it answers many of the areas that my author friends raised about how AI is impacting creative writing. Here’s the bottom line: AI is a tool. It’s a very sophisticated tool that most people don’t know the internal workings of, but I’d argue you also don’t know the inner workings of the car you drive or the microwave you use, either. As with all tools, AI is agnostic and amoral. It has no moral compass, it isn’t good or evil, and it can’t do anything on its own. It has no free will, no agency, no license to operate independently of us. Thus, when we talk about the dangers of AI in creative writing, in fan fiction, or in marketing, we’re really talking about the dangers of how humans will wield tool. Give a lumberjack a powerful tool like a chainsaw, and you’ll get powerful results from it. Give Jason Voorhees the same tool, and you’ll get a very different, equally powerful outcome. At the end of the day, it’s the person using the tool that will use it for good or ill, and as we try to set laws, rules, and guidelines for AI, we have to focus on what people will do with the tool, because the tool doesn’t do anything by itself.

Got a Question? Hit Reply

I do actually read the replies.

Share With a Friend or Colleague

If you enjoy this newsletter and want to share it with a friend/colleague, please do. Send this URL to your friend/colleague:

https://www.christopherspenn.com/newsletter

ICYMI: In Case You Missed it

Besides the newly-refreshed Google Analytics 4 course I’m relentlessly promoting (sorry not sorry), I recommend the podcast episode from this week on the 6 things that qualify someone as an expert.

Skill Up With Classes

These are just a few of the classes I have available over at the Trust Insights website that you can take.

Premium

Free

Get Back to Work

Folks who post jobs in the free Analytics for Marketers Slack community may have those jobs shared here, too. If you’re looking for work, check out these recent open positions, and check out the Slack group for the comprehensive list.

Advertisement: Bring My AI Talk To Your Company

📺 Click here to watch this ad in video format on YouTube

I’ve been lecturing a lot on large language models and generative AI (think ChatGPT) lately, and inevitably, there’s far more material than time permits at a regular conference keynote. There’s a lot more value to be unlocked – and that value can be unlocked by bringing me in to speak at your company. In a customized version of my AI keynote talk, delivered either in-person or virtually, we’ll cover all the high points of the talk, but specific to your industry, and critically, offer a ton of time to answer your specific questions that you might not feel comfortable asking in a public forum.

Here’s what one participant said after a working session at one of the world’s biggest consulting firms:

“No kidding, this was the best hour of learning or knowledge-sharing I’ve had in my years at the Firm. Chris’ expertise and context-setting was super-thought provoking and perfectly delivered. I was side-slacking teammates throughout the session to share insights and ideas. Very energizing and highly practical! Thanks so much for putting it together!”

Pricing begins at US$7,500 and will vary significantly based on whether it’s in person or not, and how much time you need to get the most value from the experience.

👉 To book a session, click here! 👈

What I’m Reading: Your Stuff

Let’s look at the most interesting content from around the web on topics you care about, some of which you might have even written.

Social Media Marketing

Media and Content

SEO, Google, and Paid Media

Advertisement: Google Analytics 4

Believe it or not, July 1st, 2023 – and Google’s shutdown of Universal Analytics in favor of Google Analytics 4 – is in less than TWENTY calendar days. This means that in TWENTY days, you will no longer be able to capture data in Universal Analytics – it will just stop collecting data. If you haven’t already switched over, it’s urgent you do so right now. So, let’s get you moving.

👉 Trust Insights can do it for you. Reach out to us if you want support setting up your Google Analytics 4 instance.

👉 You can do it yourself. Take our course, Google Analytics 4 for Marketers, to learn the ins and outs of the new system.

Tools, Machine Learning, and AI

Analytics, Stats, and Data Science

Dealer’s Choice : Random Stuff

Advertisement: Ukraine 🇺🇦 Humanitarian Fund

The war to free Ukraine continues. If you’d like to support humanitarian efforts in Ukraine, the Ukrainian government has set up a special portal, United24, to help make contributing easy. The effort to free Ukraine from Russia’s illegal invasion needs our ongoing support.

👉 Donate today to the Ukraine Humanitarian Relief Fund »

How to Stay in Touch

Let’s make sure we’re connected in the places it suits you best. Here’s where you can find different content:

Events I’ll Be At

Here’s where I’m speaking and attending. Say hi if you’re at an event also:

  • MAICON, Cleveland, July 2023
  • Content Jam, Chicago, September 2023
  • ISBM, Chicago, September 2023
  • Content Marketing World, DC, September 2023
  • Marketing Analytics and Data Science, DC, September 2023
  • MarketingProfs B2B Forum, Boston, October 2023

Events marked with a physical location may become virtual if conditions and safety warrant it.

If you’re an event organizer, let me help your event shine. Visit my speaking page for more details.

Can’t be at an event? Stop by my private Slack group instead, Analytics for Marketers.

Required Disclosures

Events with links have purchased sponsorships in this newsletter and as a result, I receive direct financial compensation for promoting them.

Advertisements in this newsletter have paid to be promoted, and as a result, I receive direct financial compensation for promoting them.

My company, Trust Insights, maintains business partnerships with companies including, but not limited to, IBM, Cisco Systems, Amazon, Talkwalker, MarketingProfs, MarketMuse, Agorapulse, Hubspot, Informa, Demandbase, The Marketing AI Institute, and others. While links shared from partners are not explicit endorsements, nor do they directly financially benefit Trust Insights, a commercial relationship exists for which Trust Insights may receive indirect financial benefit, and thus I may receive indirect financial benefit from them as well.

Thank You

Thanks for subscribing and reading this far. I appreciate it. As always, thank you for your support, your attention, and your kindness.

See you next week,

Christopher S. Penn


You might also enjoy:


Want to read more like this from Christopher Penn? Get updates here:

subscribe to my newsletter here


AI for Marketers Book
Take my Generative AI for Marketers course!

Analytics for Marketers Discussion Group
Join my Analytics for Marketers Slack Group!



Comments

Leave a Reply

Your email address will not be published. Required fields are marked *

Pin It on Pinterest

Shares
Share This