Google’s AI-focused subsidiary, DeepMind, recently announced SIMA, its new “instructable game-playing AI agent.”
SIMA, which stands for Scalable, Instructable, Multiworld Agent, is currently still in its research phase and is being trained to learn a broad range of gaming skills across a variety of scenarios — instead of just destroying humans at StarCraft II.
Through partnerships with video game developers Hello Games, Embracer, Tuxedo Labs, Coffee Stain, and others, SIMA is learning how games work and how to apply what it learns to games it’s never seen before. DeepMind’s eventual aim with SIMA, other than furthering natural language AI model research, is for it to be a devoted member of your party that does what it’s told and doesn’t take all the good loot.
Does this SIMA good idea?
“SIMA isn’t trained to win a game; it’s trained to run it and do what it’s told,” said Google DeepMind researcher and SIMA co-lead Tim Harley, according to The Verge.
SIMA researchers have focused on games that involve open-world play, rather than linear or story-driven titles, so the agent can learn to follow instructions. To achieve this, SIMA was trained by watching pairs of humans play a game — where one watched and gave instructions while the other carried them out. In a different scenario, players played freely while DeepMind researchers recorded instructions that would’ve resulted in what the player did.
We’ll admit this sounds rather appealing. If you’ve ever played an online co-op game that drops in randoms, you’ll know how risky that can be. There’s a good chance of them ruining your game, whether through incompetence or toxicity.
Having an AI party member who follows instructions means you won’t have to worry about watching your back or your hard-earned loot. Don’t feel like spending hours collecting resources? Tell SIMA to do it while you handle more important tasks.
Read More: DeepMind is back at it again, this time teaching AI how to play football
However, as appealing as this might sound, it’s worth remembering how training AI models on human behaviour — especially when online human interaction is involved — has gone in the past. TayTweets, anyone?
This probably isn’t a problem in a controlled research environment but, should SIMA ever be trained on average human-based online gameplay, we doubt it will take long before the griefing starts.
