Latitude, the company behind the text game AI Dungeon, is expanding into Voyage, a new artificial intelligence-powered game platform.
On Friday, the company announced the closed beta, as well as a waitlist for current AI Dungeon users.
It's the next step for a startup that began as a university hackathon experiment but intends to someday enable people develop their own games using trained AI models.
also known as AI Dungeon 2, was published in 2019 and is driven by OpenAI's GPT-2 and GPT-3 text creation algorithms.
You begin by creating some introductory text or writing your own adventure setup.
Then you may type in whatever command you like, and a virtual game master in the style of Dungeons & Dragons will improvise some prose describing the result.
It's very strange and a lot of fun, but it's lacking in traditional game mechanics, more akin to an interactive fiction engine.
The games in Voyage are more organized.
Medieval Problems is a Reigns-inspired experiment in which you play as the monarch of a kingdom and enter freeform text orders for your advisers, with the results displayed in success ratings.
It's still a lot like AI Dungeon, but with a clearer framework for what you're expected to accomplish and a mechanism for measuring success - however, after playing with the game, that approach appears to be quite forgiving and more than a little arbitrary.
Pixel
Meanwhile, in this party game, one person enters a sentence, the AI constructs a pixelated image of it, and that image gradually grows in quality until another player gets it.
It's a cross between the painting app Dream and a Pictionary-style mechanism.
Latitude CEO Nick Walton sees Voyage as a natural progression for the company.
"AI games are kind of resuming from the beginning," according to the business, with text adventures reminiscent of Zork or Colossal Cave Adventure.
"We're now moving towards 2D graphics with some visuals."
AI Dungeon, which is featured with Voyage, has recently included AI-created images generated using the Pixray image generator.
The ultimate objective is to add game production tools to Voyage, rather than simply games.
"Our long-term aim is to enable artists to produce things that are dynamic and alive in ways that present experiences aren't, as well as to be able to create things that would have previously required studios of a hundred people," Walton adds.
There is no set timeline, but Latitude intends to work on the system throughout the first part of next year.
Voyage may be able to find a long-term business plan with the use of creative tools.
AI Dungeon is now free for a restricted set of GPT-2-powered features but requires a membership to access the higher-quality GPT-3 algorithm.
Following the Voyage beta, Latitude intends to provide a subscription service for it. 
But Voyage's new games lack the adaptability and replayability of AI Dungeon — they're still plainly the output of a business trying to crack machine learning games.
"This technique," Walton adds, "is one of the things that I believe would be incredibly advantageous in terms of being able to iterate and find out what experiences people prefer."
"With traditional games, you can use existing models and create a game that you're confident people will enjoy."
But this space is so distinct that it's difficult to say."
The question is how much individuals will be willing to pay to participate in that process.
Latitude's objective will most likely need prudence with OpenAI's application programming interface as it evolves (API).
GPT-3 initiatives are approved on an individual basis by the organization, and they must comply to content criteria designed to avoid misuse.
Latitude has previously battled with these constraints, as AI Dungeon provides users a lot of latitude to construct their own narrative, leading in some users proposing terrible sexual scenarios that concerned OpenAI.
(It has also addressed security concerns about user instructions.)
The business spent months developing filter algorithms that inadvertently rejected more trivial fictitious content before agreeing to send some user requests to a non-OpenAI algorithm.
Pixel This and Medieval Problems are more closed systems with less visible moderation hazards, but offering creative tools risks undermining OpenAI's management over GPT-3, which may provide its own set of problems.
Latitude plans to transition more of its games to other algorithms in the future, according to Walton.
"We'll have more structure and mechanisms in place so that it's not just consuming the [OpenAI] API in the same manner."
"At the same time," he argues, "I expect the majority of our models will be ones that we host ourselves."
This includes models based on new open source projects, which have struggled to compete with OpenAI's efforts but have progressed since their inception.
"I don't think that gap will be there for long," Walton says.
Many games employ procedural generation, which combines developer-created building pieces to generate massive amounts of material, and the majority of video game "AI" is a very basic set of instructions.
A business like Latitude, on the other hand, employs algorithms that have been taught to create text or images that suit a pattern from a data collection.
(Consider them to be super-advanced autocomplete systems.)
Right now, this may result in incredibly surprising experiences, and their ridiculousness is frequently part of their appeal – outside of gaming, other businesses, such as NovelAI, have used text creation for artistic work.
Latitude, on the other hand, is still learning out how to create mechanisms in which participants can anticipate fair and consistent outcomes.
Text creation algorithms, for example, have no built-in sense of whether an action succeeds or fails, and systems for making those decisions may disagree with normal human intuition.
Image generating algorithms are fantastic for creating bizarre art, but in a game like Pixel This, players can't always anticipate how recognizable a specific image would be.
For the time being, Latitude's approach is to lean into the turmoil.
"If you attempt to develop anything extremely serious using AI where people demand a high level of coherence, it's going to struggle, at least until the technology improves," Walton adds.
"However, if you embrace that side of it and allow it to be crazier and wackier, I think you produce a wonderful experience and people are pleased by those shocks."
Post a Comment