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.
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.
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.
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.
Chas Belov said,
August 2, 2024 @ 11:35 pm
Decode the ancient symbols
Chas Belov said,
August 2, 2024 @ 11:40 pm
Actually, my question would be ¿does AI keep track of the state of the game, what objects have been found, which doorways lead to which rooms, which puzzles have not yet been solved, whether I yet have the information I need to decode the ancient symbols, whether the knowledge necessary to decode the ancient symbols can be acquired, how to acquire it, whether the planetarium has a working restroom, and so on?
Chas Belov said,
August 2, 2024 @ 11:42 pm
It's entirely possible that someone left the ancient code decoder book in the planetarium restroom.
Haamu said,
August 3, 2024 @ 7:35 am
[At this point, I had it provide a visualization of Dr. Grey in the restroom. Really wish I could post the pictures. Each time it visualizes, it adds more text.]
Haamu said,
August 3, 2024 @ 7:40 am
Haamu said,
August 3, 2024 @ 7:43 am
Haamu said,
August 3, 2024 @ 7:47 am
Haamu said,
August 3, 2024 @ 7:55 am
Just curious — are there any lesser known places of "linguistic significance" in the city?
Chas Belov said,
August 3, 2024 @ 2:25 pm
As someone who cut their adventure teeth on Colossal Cave, I find this fascinating. I'm still wondering whether there will eventually be continuity errors. Or an end game; ¿do you just keep playing until you get bored or will the game eventually present the opportunity for a solution? ¿How will you know if the game has a solution?
I found the difference between the player and Dr. Grey to be confusing (also that he's Dr. Grey and not Dr. Gray, but that's off-topic). I trace it to:
where I might have written:
I would have Dr. Grey go for a Philly cheesesteak in case that might offer a clue, but that's just me. Like, ¿what if it's some other cheese than Cheese Whiz? ¿What would that mean to Dr. Grey's quest? ¿Is the cheesesteak shop some kind of secret cabal?
Haamu said,
August 3, 2024 @ 3:33 pm
Note that the way the decoded message is filling in displays clear continuity with the original message several exchanges back.
At this point I decided a picture of them eating their cheesesteaks was mandatory, so I typed in "/visualize" — but I wasn't certain, since the narrative had now returned to the library, that the system would "understand" that the most visually interesting moment to depict would require a minor backtrack. But that's just what it did.
The image does not disappoint. The sandwiches look delicious. Both characters' costumes have been updated: Grey now looks vintage 1950 rather than 1910, while Linguist_Sleuth no longer looks like a Grim Reaper knockoff and instead wears what looks like a leather trenchcoat and hoodie. He also actually has a face, the visible lower half of which sports a well-trimmed beard and moustache.
The neon signs on the walls of Geno's betray DALL-E's inability to deal with text in images properly. A sign for "GENO's [ST]EAKS" (partially obscured by Grey's head) is right next to one for "GENGUET' STEAKS". Another neon sign says "LINGUIST_SLE[UTH]" (i.e., when rendering text, DALL-E seems to draw from anything in the context, whether or not it's appropriate).
The "/visualize" prompt produces not just the picture, but a recap bringing us back up to date:
GH said,
August 5, 2024 @ 3:16 am
AI Dungeon has been running since 2019. When I first played it it was based on GPT-2, and it was one of the first free ways for the general public to access the technology that would become ChatGPT.