When AI hallucinations are a Good Thing

« previous post | next post »

Locally consistent hallucinations, anyhow… Zoë Hannah, "We pushed this ChatGPT game to the limits, but playing it the right way is more fun", Polygon 7/30/2024:

Apparently, we all like playing god, and we all like doing it badly. I bet none of us thought that removing the ladder from our Sims’ pools was such a universal experience until it became a pretty popular meme, and it’s no secret that lots of mods are centered on adding, uh, explicit elements to games. So, naturally, when I started playing around with DeepGame, Utile Labs’ ChatGPT-based choose-your-own-text-adventure game, I put my best sicko foot forward.

The game, which runs on ChatGPT and is available to anyone with an account, generates stories in a variety of genres. You start off with a command like “Play a romantasy story” or “Surprise me” and let the GPT do its thing — and despite my desire to break the game, I found it much more enjoyable when I took it just a little more seriously.

DeepGame’s first response almost always begins with scene-setting followed by introducing you, the protagonist, as well as a few side characters and a clearly stated challenge or adventure. Then the game asks, “What do you do next?”

Answering the question is titillating, to say the least. There are no prompts or choices to pick from — you can go in any direction at all, and the game keeps up, spitting out several paragraphs to move the story along after each of your responses.

What’s more, if you have the paid version of ChatGPT, you can use the command “visualize” to generate an image of the current scene using Dall-E. It’s a feature that’s easy to forget about if you treat the game like reading a novel — at least for me, since I typically create an image in my head as I read — but you shouldn’t ignore it, because whatever parameters the devs put on the image generator make for some truly delightful interpretations. It’s the more imperfect side of DeepGame, which is part of why it makes me so giddy — and you can always regenerate the image if it decides to throw in some random characters or elements that don’t match up with your narrative.

As Wikipedia explains,

Adventure games were initially developed in the 1970s and early 1980s as text-based interactive stories, using text parsers to translate the player's commands into actions. As personal computers became more powerful with better graphics, the graphic adventure-game format became popular, initially by augmenting player's text commands with graphics, but soon moving towards point-and-click interfaces. Further computer advances led to adventure games with more immersive graphics using real-time or pre-rendered three-dimensional scenes or full-motion video taken from the first- or third-person perspective. Currently, a large number of adventure games are available as a combination of different genres with adventure elements.

Exchanging old-fashioned parsers and story-generation rules for an LLM is an obvious development, but this is the first example I've seen. The transcript for one of Ms. Hannah's DeepGame adventures is here.

In any case, this is an application where real-world facts are not relevant — though keeping track of the state of the fantasy world still matters, and forgetting what's happened, what's been found and what hasn't, etc., would definitely be bad.

Update — in real-world trip planning, an AI-based system "has no concept of time or space or what a human being might find interesting" (Natasha Bernal & Amanda Hoover, "We Asked AI to Take Us On a Tour of Our Cities. It Was Chaos", Wired 7/18/2024). Users will be more tolerant of chaos in DeepGame's in-adventure travel plans, but I wonder how much chaos they'll need to tolerate.

 



1 Comment »

  1. Haamu said,

    August 2, 2024 @ 4:36 pm

    I feel like this post deserves at least one comment, so I had the game generate the following. Rather than providing the prompt that I used, I think it's more entertaining to let you imagine it.

    Prologue

    It's a crisp autumn afternoon in Philadelphia. Dr. Samuel Grey, a distinguished professor of linguistics at the University of Pennsylvania, sits in his cozy, book-lined office, surrounded by stacks of papers and glowing computer screens. His renowned blog, "The Linguist's Lens," has just reached its twentieth anniversary. Known for its insightful commentary on language trends, historical linguistics, and the quirks of modern usage, the blog attracts a diverse readership. Today, Dr. Grey is about to post on a highly controversial topic: the influence of social media on language evolution.

    Scene 1: The Controversial Post

    Dr. Grey types the final words of his new blog post, "Hashtag Linguistics: The Impact of Social Media on Language Evolution," and clicks "Publish." Almost immediately, comments start to flood in. As usual, the regulars are there:

    Professor Elaine Walker, a syntax expert with a penchant for long-winded but informative comments.
    Dr. Julian Thompson, a computational linguist who often shares insightful data visualizations.
    Maxine Weller, a high school English teacher with strong prescriptivist views.
    "GrammarManiac99", a persistent troll who constantly argues that language change is inherently bad.

    However, today there's a new commenter: "Linguist_Sleuth." This mysterious individual, who has no previous posts or profile information, posts a cryptic message:

    "Dr. Grey, your post scratches the surface. The real story lies deeper. Follow the clues, and you'll uncover a linguistic conspiracy that spans centuries."

    Scene 2: The Discussion Heats Up

    Dr. Grey leans back in his chair, intrigued by the new commenter. The regulars quickly react:

    Professor Walker: "What nonsense! Conspiracy theories have no place in serious linguistic discussions."

    Dr. Thompson: "Hold on, Elaine. Let's see where this goes. It could be an interesting angle."

    Maxine Weller: "More crackpots. Great. Just what we need."

    GrammarManiac99: "Finally, someone who gets it! Language is being manipulated!"

    The debate grows heated, with each comment sparking a flurry of replies. Meanwhile, "Linguist_Sleuth" remains silent, watching the chaos unfold.

    Scene 3: The Mysterious Clues

    As the evening progresses, Dr. Grey receives a private message from "Linguist_Sleuth":

    "If you're interested in the truth, meet me at the Franklin Institute at midnight. The answer lies in the language of the stars."

    Dr. Grey's curiosity is piqued. He decides to follow this enigmatic lead, feeling a mix of excitement and trepidation.

    Scene 4: The Franklin Institute

    At midnight, Dr. Grey arrives at the Franklin Institute. The museum, closed to the public, is eerily quiet. He spots a shadowy figure near the planetarium entrance. It's "Linguist_Sleuth," a tall figure in a hooded cloak.

    "Linguist_Sleuth" speaks in a low, controlled voice, "Dr. Grey, our language has been influenced by forces beyond academia. Ancient scripts, hidden messages in star charts, and codes passed down through secret societies. Follow me, and I'll show you."

    Scene 5: The Revelation

    Dr. Grey is led into the planetarium. The star projector lights up, casting constellations across the dome. "Linguist_Sleuth" begins to explain the connections between the stars and ancient languages, showing how certain constellations align with cryptic symbols found in historical texts. Dr. Grey is astonished by the depth and intricacy of the information.

    The planetarium glows with constellations and ancient symbols, casting an otherworldly light on Dr. Grey and the hooded figure. "Linguist_Sleuth" points to a constellation, explaining its connection to an ancient script.

    "Linguist_Sleuth" turns to Dr. Grey. "The secrets of our language are written in the stars, hidden in plain sight. But there's more. These symbols are part of a code used by a secret society throughout history. It's a society that has influenced the evolution of language in ways we never imagined."

    Dr. Grey's mind races as he processes the information.

    What do you do next?

    I don't know how to post images, or I'd include the visualization, which is quite nice. "Linguist_Sleuth" looks a little like the Ghost of Christmas Yet to Come.

    If anyone wants to suggest a next action, I'll post the next chapter.

RSS feed for comments on this post · TrackBack URI

Leave a Comment