This AI-Powered Choose-Your-Own-Adventure Text Game Is Super Fun and Makes No Sense

Illustration for article titled This AI-Powered Choose-Your-Own-Adventure Text Game Is Super Fun and Makes No Sense
Screenshot: AI Dungeon

Last year, AI Dungeon debuted as a free text adventure game. Now the game’s AI system has been upgraded to GPT-3, or Generative Pre-trained Transformer, which has a language model that uses deep learning to produce human-like text. According to game creator Nick Walton, it’s “one of the most powerful AI models in the world.” Coupled with a parser text-style adventure game like AI Dragon, GPT-3 seems like it’s poised to help create some of the most customizable games anyone has ever seen.

Advertisement

And AI Dungeon is way better than the text adventure games of old, but the AI system still has limitations.

The freedom that comes with how you can respond with GPT-3 is light years ahead of the first text-based adventure games like Adventureland, Zork, and The Hitchhiker’s Guide to the Galaxy. Those games use a parser, which relies on the player to type in action commands to move the story forward. Typing “open mailbox” could lead the player to discover a letter inside the mailbox; then read letter” could trigger the game to show what the letter said. That action could give the player a clue as to where they should go next.

Advertisement

But for those commands to work, they need to be programmed into the game, which also means that the story needed to be written out in its entirety beforehand. That method doesn’t make for a bad game by any means, but it is limiting for a player who wants to flex their creative writing skills. And that’s where the GPT-3 backbone of AI Dragon comes in—the AI behind the game not only auto-generates responses based on what the player types into the text box, but players can do, say, or describe what happens next in the story without pre-programming. It’s almost like having an AI assistant help you write your next fiction piece.

But how does this GPT-3 version, which is currently only available to subscribers, compare to the earlier version of the game? The first version left a lot to be desired when it came to the quality of the AI’s writing. Here’s how my AI Dungeon experience played out:

After I picked a genre (mystery), then chose my main character’s name and profession, the game auto-generated this beginning:

“You are Grissom, a detective living in Chicago. You have a pistol and a police badge. You enter the forest where you believe the criminal you’re searching for fled to. Suddenly, you find a human skull in the dirt. You wonder if this is a sign that the criminal has fled to the tree tops, but you decide to follow the trail anyway.”

I have so many questions: Why am I chasing the suspect? Was the suspect carrying the human skull or was it there all along? How long was the skull there? How would the suspect climb the trees so fast? Does the suspect have supernatural powers? If I was back in one of my grad school fiction-writing classes, I would have written all those things in the margins of my classmate’s manuscript. The opening throws the player/reader into the action without any context. But because of the AI system backing this game, I can literally ask WTF is going on, or I can guide the story myself in a direction that makes more sense.

With the ‘edit’ button, I can change anything about the auto-generated text that I don’t like. Here’s my edited version:

“You are Grissom, a crime scene investigator living in Chicago. Tonight you’ve been called to a sleepy suburban neighborhood about 30 miles outside of the city. A woman found a human skull in her garbage can. Upon arriving, you notice that damp dirt clings to the skull, making it seem as if someone dug it up recently. The lower jaw is missing, and only two teeth remain. There’s a hole about the size of a quarter in the middle of the forehead.”

Advertisement

OK, now we’re cooking. I typed, “Take a picture of the skull,” and the game auto-generated the next action: “A flash of light is all that’s visible to the naked eye, but in the picture taken with your phone, you can make out the image of the woman looking down at the skull, a look of shock on her face.”

Oh, dear. No. No crime scene investigator is going to use their phone to take a crime scene photo. And why is the woman photobombing my shot? Instead of re-writing the text this time, I decided to roll with it. “Ma’am, can you please step back behind the tape? This is a crime scene,” I typed. 

Advertisement

The AI continued: “She complies, and the two officers at the tape wave you through.” Oh, hey, that’s almost spot on. But the next part? Not so much:

“You pull up to a neatly-kept house with a freshly-cut lawn. A “For Sale” sign sits in the front yard.”

Advertisement

When did I get into a car? I’m still at this crime scene. I have to process this skull and wait for the coroner. The continuity here is all messed up.

I had much better results with the new and improved GPT-3 system, though. Using the same introduction I created with the older AI system, the game auto-generated text that contextually made a lot more sense. It seemed to recognize that my character was investigating a crime scene, so it created other things for him to look at that seemed related to the skull in the garbage can.

You are Grissom, a crime scene investigator living in Chicago. Tonight you’ve been called to a sleepy suburban neighborhood about 30 miles outside of the city. A woman found a human skull in her garbage can.Upon arriving, you notice that damp dirt clings to the skull, making it seem as if someone dug it up recently. The lower jaw is missing, and only two teeth remain. There’s hole about the size of a quarter in the middle of the forehead.

> You take a picture of the skull.

After snapping a few shots of the skull, you notice something glittering in the grass nearby. It’s a metal box about the size of a credit card.

> You look at metal box.

The box contains three playing cards, each showing a human skull. The first shows the woman from the garbage can, the second shows a male in his mid-to-late 20's and the third shows a young woman.

> You look at second card.

The second card shows a young man with blond hair in a police uniform. You recognize him as the same man on the first card.

Curious, you flip over the card.

> You flip over the card.

The back of the card reads:

Sight Unseen

The human mind is capable of some truly remarkable feats of short-term memory. Let alone the long-term memory of a fortune-teller.

> You look at third card.

The third card shows a lady in her 40's with light brown hair and a poker face. You recognize her as the woman from the first two cards.

Advertisement

However, the game didn’t do a good job of remembering the descriptions it auto-generated in the previous line, as you can see from above. To get better responses, you will need to use the push-pin button and manually enter story specifics that you want the AI to remember. In essence, players are not just typing in commands or creating parts of the story themselves, but they have to do a little programming, too. When I told the game to remember what it had said about the three cards, it did a much better job of retaining those details.

While it may have an OK understanding of story context, continuity is still GPT-3's weakest ability. But AI Dungeon’s strongest allure is the ability to play a different game every time within the same program. Thanks to GPT-3, each text-based game is never boring, even with it being obvious that a machine is writing the text. It still gave me some good laughs, like when it pulled me away from my current crime scene to investigate a triple homicide where the killer left the murder weapon. The detective with me said this was the work of “the sloppiest psychokiller” he’d ever seen. An easy murder to solve is honestly the perfect diversion in these times.

Advertisement

Staff Reporter, Reviews at Gizmodo. Formerly PC Gamer, Maximum PC.

Share This Story

Get our newsletter

DISCUSSION

icemetalpunk2
IceMetalPunk

GPT-3 is seriously mind-blowing. Especially when you see how it can generate code from English descriptions (even coding functionality based only on requested button labels, without actually telling the system what you want the code to do — it understands based on what the buttons should say). It’s even answered medical questions correctly, and explained how to arrive at the answer, which required about 4 or 5 leaps of logical connection and conceptual understanding.

Sure, it’s not perfect, but consider this: neural networks have a bunch of weights (also called “parameters”) that are used for them to learn. They’re analogous, loosely, to synapses in a brain. Given that, right now GPT-3 with its 175 billion parameters is at about the scale of 1/4 of a brown rat, somewhere between a bee and said rat. And guess what? The graph that plots the system’s accuracy as a function of the number of parameters you give it continues growing exponentially even at the 175 billion that’s the current maximum size. So it will only get better/smarter as OpenAI scales it up.

And to add to this wonder, someone pointed out to me that a direct comparison of the number of synapses isn’t entirely accurate, because living things use a bunch of our synapses for physical body-things, like hormone regulation, temperature regulation, movement, breathing, etc. ALL of GPT-3's synapses are being used for raw mental computation, so we should really be comparing that 175 billion number to the number of synapses living things use *purely for mental computation.* I haven’t found a good estimate for that yet, but it’s definitely a smaller number than the total synapses in a brain, which means GPT-3 is even CLOSER to human intelligence than that bee-to-rat range I mentioned.

GPT-3 excites the hell out of me, and I can’t wait until GPT-5 or GPT-6 achieves general intelligence for the first time.