|
Welcome back! Catherine and Aaron here. Without video games, the artificial intelligence boom may not have happened: the Nvidia chips that powered recent AI breakthroughs were originally designed to process video game graphics. Now, game studios are increasingly turning to generative AI to automate the work of developing their games. Game developers have been publicly musing about the possibility of using AI to generate parts of video games for years, and studios such as Microsoft’s Xbox have published hypothetical research about one day using generative AI to develop games. But AI models themselves have only recently gotten good enough to be reliably used in game development, according to developers and AI founders. A raft of startups is selling AI tools for game development to those studios and promising they can save money that would have previously been spent on humans to design the games from scratch. Some video game firms are becoming receptive to that pitch. For instance, Genvid Technologies typically hired game studios to design the cutscene animation—the scenes that play in between the interactive parts of a game—which Genvid would then use to create interactive worlds on behalf of film and television streamers and other tech companies (clients have included Meta and Twitch). Most of Genvid’s titles are choose-your-own-adventure games where users decide what actions characters take in between short animated scenes, and the animation was always the most expensive part of production, said CEO Jacob Navok. But now, video-generation models from Google, MiniMax and Kling have reduced the costs of producing a 20 minute video to around $1,500 from $2 million, he said. The AI has also saved tons of time—it took a deputy of Navok 6 hours to make three minutes of a prototype version of a game, a task which previously would have taken six months, he said. These time and cost savings have allowed Genvid to take on more projects than it used to. It wasn’t always like this. Since ChatGPT sparked the latest AI boom in late 2022, Genvid had been experimenting with ways to use generative AI to create graphics but the tech wasn’t good enough until this year, Navok said. In fact, when gamers in late 2023 suggested that a Genvid-produced game, Silent Hill: Ascension, was powered by AI, Navok publicly denied the allegations, saying AI wasn’t good enough to do the job. Two years later, many game studios are now using generative AI models to generate specific features like collectible items or clothes within a game. Square Enix, the studio behind the popular Final Fantasy franchise, invested in a Vienna-based startup Atlas and is now testing out the generative AI startup’s technology to automatically generate assets in its games. Atlas is also selling its technology to the game studio Parallel, which has released a limited preview of a game that lets users create AI-generated armor for their characters. Other firms are starting to test using AI to generate entire interactive worlds that a user could explore, similar to a game with a first-person point of view, such as Call of Duty. AI startup Decart, which has trained these so-called world models, has been in talks with several big game studios, according to founder Dean Lietersdorf. Other AI firms like Google and World Labs, a startup founded by Stanford AI professor Fei Fei Li, have debuted similar models capable of creating interactive worlds. Lietersdorf thinks the cost of models still needs to come down before we’ll see a fully AI-generated game. While Google’s Veo 3 model costs about $1,000 to generate an hour of video, Lietersdorf said that Decart has trained smaller models that can generate video for less than 25 cents an hour, although the video quality is lower. Once AI models can reliably generate high-quality video at less than 10 cents an hour, game studios can offer such games on a subscription basis for users, Lietersdorf said. “The entire gaming industry will change when there‘s a single game that’s completely AI developed, and that game goes very viral, and I think we’ll probably see that in the next three to six months,” Lietersdorf said. Generative AI for video games isn’t as easy as flipping a switch. Navok said Genvid had to create its own tech to synthesize different AI models for graphics creation, partly because different models had different strengths and weaknesses in their animation abilities. Also, the video models struggled to generate videos long enough for a whole game, so Genvid’s internal tool could help make sure clips created separately remained visually consistent. And Navok said that for certain parts of video game creation, like voice acting, music and editing, humans still have the edge. Still, those tasks cost much less than the animation itself. The cost savings Genvid gets from using AI are high enough that the company will only consider partnering with studios that are comfortable with its use, which can be a touchy topic among some creatives, Navok said. “In the same way that digital filmmaking replaced people painting on cels, this is going to happen,” Navok said.
| 
