Google develops an AI that will actually play games with you

Google is developing a DeepMind based AI that could end up being the ultimate virtual co-op buddy. It's not some play-to-win opponent or super-bot, but a general, instructible game-playing AI agent.

Google presented a research article (via @rowancheung) on what it calls a Scalable Instructable Multiworld Agent (SIMA). The general idea is that it's a learning AI that can follow verbal instructions and understand the virtual world it's in. Rather than functioning as a hard-coded AI opponent or bot we've grown accustomed to over decades, Google's SIMA promises to a more natural and human-like gaming companion.

Google partnered with eight game studios to test its SIMA model in games including Valheim, Goat Simulator 3 and No Man's Sky. These are open-world games chosen to teach SIMA to learn general gaming skills. The current version of SIMA is capable of performing around 600 basic skills, including navigation, object interaction and menu use.

As a player of No Man's Sky, I'm very intrigued by the possibilities. Anyone that's played the game knows that resource gathering and construction is a tedious process, and telling an AI player to go out and find oxygen or build a base would really help to reduce the grinding aspect of it. I prefer exploration and things like dogfights more than traversing landscapes to find X, in order to craft Y to build Z. SIMA could really help with that.

The early results look promising. Google says a trained SIMA agent performed almost as well in an unseen game as an agent specifically trained for it. That is key if such an agent is to be useful for real gamers. If a typical game playthrough lasts tens to hundreds of hours, an AI companion needs to be capable from the opening scene, otherwise players just won't bother.

While this is an impressive and very interesting feat, I remain skeptical. Large-language AI models are very good at doing things that are based on knowledge, but their ability to react, adapt and appropriately behave in real-time as a human does remains limited. Just how good is it at landing a headshot on the other side of a map?

AI in video games is not a new concept. It goes back decades to things like boss fights, multiplayer bots, opponents and NPCs in single player games, but they are hard coded for each game. A properly implemented SIMA-like AI could dramatically impact general gameplay. And it doesn't require source-code access or an API to function. It instead relies on just two inputs: the images on screen, and the instructions provided by the user.

Google emphasizes that its research is still in its early phases. As it learns from more games, it's expected to become more versatile and adaptable. Could your next co-op companion be an AI? And an actually useful one at that? I look forward to seeing how this technology develops, and how it might impact the future of multiplayer gaming.