The Companion looks so much better on our free app. Customize your feeds, get notifications on new content, and stream podcasts with our in-app player.
Download it from App Store for iPhone or Google Play for Android now – and don’t forget to rate and review!
In my first diary entry, I shared my anxiety around the A.I. experiment of these proportions, and in sharing my thoughts, I realized a lot of it came from my computer science studies 20 years ago!
Right, so how do you begin to tackle a problem that’s overwhelming? You start by doing laundry.
And now that I have extraordinarily clean whites hanging on my drying rack, I go into research mode. A good place to start is to understand the basics so check out the piece our Editor James Hoare and A.I. expert Janelle Shane wrote on “How Does Artificial Intelligence Write New Stargate?”
Here’s my TL;DR version:
- AI isn’t smart or sentient the way it’s portrayed in films, it’s about prediction.
- How it predicts the next words are based on the input you give it.
- And it can only respond to inputs based on parameters you feed it.
For example, Let’s start with a word association game.
If I said a word like “Teal’c,” what word is most likely to come up next? It might be “Indeed”!
Or if I said, “Apophis,” you might reply, “Shol’va!”
(Of course, you also need to consider articles because the most likely words might actually be: the, be, to, of… but let’s take that out for now.)
The next thing to do is to have a play with it. I mean, I have no idea what’s going to happen. Unlike the A.I. experts, I haven’t really dabbled with this since 2002. The fastest way I did this is with GPT-2 to give you some examples:
Here’s my first input:
Sam and Jack share ice cream in the mess hall.
And here’s the output >>>
Then Sam tells him that she has to take care of something, but that she’ll be back in a little while.
When she returns, she tells Jack that the other Runners have gone crazy; one has tried to kill them all and the rest are in the mental ward.
Sam asks Jack what he thinks about things.
Jack says that they have to stop at their cabin for the night.
When she wonders if he really means it, he assures her that he does.
Oh my, we’ve just ventured into fanfic heaven.
Despite my urge to continue down this storyline, I thought I’d try the exact same input again:
Sam and Jack share ice cream in the mess hall.
And here’s the second output >>>>
Jack continues to say that his mom is sick but what does he do, he leaves her and this mess.
Jack then tells Sam that it was his job to take care of his mom and his role now is to help find him.
We see Jack become more focused on his mission to help find his dad.
Now, Jack is not doing well with the first contact.
Jack’s mother tries to contact Jack and after several failed attempts, she finally has a successful communication with Jack.
In an interesting way, we’re kind of able to piece moments together and if you squint hard enough, particularly in our ‘shipping fantasies, you can start to believe this to be a story.
If you haven’t read Janelle’s very funny ‘Starcake SG-Bun’ script based on GPT-3 (the next iteration from my test), it’s a good example.
However, like Janelle and James explain, to get even a short script like ‘Starcake SG-Bun’ to be created, it required a “labor-intensive and heavily edited process” in order for it to not turn into gibberish quickly. So in theory, we might be able to brute force this project through by clicking “Generate Another” a hundred times, a thousand times, a million times to try and find something semi-coherent…
Ra – yes
Sponge Cake – no
Atlantis – yes
Decorations – no
…but we’d like to go a little bit further.
GPT-3 is an amazing language model. It’s so powerful with the full version having a capacity of 175 billion machine learning parameters.
TL;DR: GPT-3’s basically read the internet.
And that’s why it’s able to return something so interesting that includes a bit of Stargate, a dash of cake recipes, and a sprinkle of decoration classes, but after brainstorming with Brad Wright, it’s not exactly the Stargate script we’re hoping for.
Let’s find the best expert we can.
Going back to my last chess A.I. class, here’s the problem: I only know what I know and I only knew what I knew back then, and it wasn’t a lot. Back then, I had to ask my friend Ryan and so now we need to enlist someone.
To tackle this unique challenge, we brought on board Laurence Moroney who Leads A.I. Developer Relations at Google and teaches at Harvard. Not only is he an expert in this field, but he’s a huge sci-fi fan and published SF author. More importantly, he’s a massive Stargate fan. He created the Silverlight/Photosynth 3D-tour of Destiny ahead of the launch of SGU, and wrote the Stargate Universe: Icarus comic-book.
Unsurprisingly, he had a lot of answers:
- An A.I. can only spit out what’s been fed in.
- GPT-3’s power is that it’s been fed a lot, but that makes it prone to going off message.
- If we are trying to conduct a very specific experiment – can an A.I. spit out a Stargate script –then let’s avoid feeding it anything that’s not related. By feeding it only Stargate scripts, then the only output can be Stargate scripts.
- It’s a totally different approach. Instead of feeding our A.I. 175 billion inputs, we may be feeding millions, but those inputs are very specific to our project.
So our next step is to track down all of the Stargate screenplays and ingest them to create a Stargate bot specifically for this project.
As this sinks in, Laurence, Brad, and I wonder what this means for the future of creative industries like screenwriting. For reasons explored here and in our previous articles, A.I. is unlikely to reach the level of sophistication needed to generate anything truly new. There won’t be new new Stargate because A.I. can only live in the past, but what it might deliver is a remix of old Stargate that begins to look and feel new, with a familiar melody.
Like in our little GPT-2 experiment, what this project could very well turn into is a tool to help generate thought starters. We already have examples of a love story, a story in search of her father, and a potential fantasy. Creativity doesn’t take place in a vacuum – we’re all influenced by things we’ve read, things we’ve experienced, dreams, or whatever.
Maybe what we’re doing is creating a program to help creatives move forward. Something that can generate direction when you grind to a halt or hit the wall of writers’ block. Even if it’s seemingly unhelpful in itself, it might give you a bad idea to cast off against with such force that the momentum carries you into something better. It might surface aspects of a relationship between characters that you’ve overlooked, and you can either run with that or have a serious think about whether what you’ve already written – the text that the A.I. is riffing on – has unintended interpretations. It might bruise the ego a little, but it could draw your attention to phrases or tics that you lean on too heavily, or idiosyncrasies of your own that have been propelled into numerous mouths.
A.I. is highly unlikely to displace human creativity, but maybe, just maybe, it can act as fuel for that creativity, as encouragement, and perhaps chastisement.
If you’re interested, go have a play yourself and copy and paste some A.I.-generated stories below in the comments, it’ll be super fun.
Stargate | How the GPT-3 A.I. Wrote its First ‘Starcake’ Script
Ask Robert Picardo A Question
Lawrence Kao, is the Co-Founder of The Companion. He’s been working in the technology and media space since 2003.
Follow The Companion on Twitter @TheCompanionApp