Yesterday, Wednesday 13 March, Google DeepMind has published research in the context of Scalable Instructable Multiworld Agent (SIMA). We are talking about an agent who, apparently, is used by Google to train theArtificial intelligence through video games.
According to what was stated by Tim Harley, an engineer at DeepMind involved in the project, the interest in the gaming sector should not be surprising. This, in fact, would represent a good “training ground” for AI.
Since the launch of ChatGPT, which occurred during November 2022, the market has seen a flourishing of different AI-related tools. The technology giants, in fact, have dived headlong into this new and unexplored world, with theGenerative AI exploited first for simple text and then to create realistic images, music, videos and even video games.
SIMA, AI put to the test with video games and surprising results
As described by Harley, SIMA is able to interact with games, executing instructions. This does not mean that AI is unbeatable, on the contrary.
The researchers involved in the project wondered whether the AI could maintain the skills obtained by trying a title, also using them on a game it had never tried before.
To carry out this interesting test, DeepMind researchers recorded Images e keyboard input e mouse of human players and exploited imitation learning techniques to transmit instructions to SIMA on how to play. The titles examined were No Man’s Sky, Eco, Teardown e Goat Simulator.
The interactions proposed to SIMA have been cataloged in 600 categoriesamong which they are included movements, interactions with objects e using the menus. The results obtained by the AI were surprising.
In fact, SIMA has proven to be a better AI than “flesh and blood” gamers. For Frederic Besseresearch engineer at DeepMind, SIMA “He is able to take advantage of shared concepts between games, gain better skills, and learn to better follow those instructions” compared to a human gamer. All this, however, does not mean that AI has proven infallible.
The weakness of AI compared to video games
As Harley pointed out, SIMA has shown some undeniable weaknesses. These are small imperfections that the engineer himself does not want to define as hallucinations.
Per Harly “Often what we see when the agent fails… I wouldn’t call them hallucinations, his behavior seems intentional most of the time, but he fails to perform the behavior necessary for a certain action“.
Despite fallibility, SIMA’s experience is valuable to DeepMind. In fact, it is a solid basis on which to work to make AI even more advanced and useful even outside the gaming context.