An artificial intelligence (AI) program by Google has shocked technology and gaming communities worldwide, figuring out how to collect coveted diamonds without prior training in the popular video game, Minecraft.
Unlike previous AI models that needed to be trained with hours of gameplay footage, Dreamer, the name of this AI system, learned how to navigate Minecraft without prior exposure. Creators of Dreamer say this event marks a huge step forward for all AI, as these machines can now generalize knowledge they’ve learned to new situations.
Dreamer was not given explicit instructions when launched into the game, but was encouraged to explore. In addition, every 30 minutes, researchers would reset the game terrain, forcing Dreamer to adapt to a completely new, randomly generated map. This made it hard for the AI to memorize a specific strategy.
Despite these constant changes, Dreamer continued to improve at an impressive rate. At the nine day mark, the AI system could navigate the game world and mine diamonds in under 30 minutes, matching the rate of human players.
“Dreamer marks a significant step towards general AI systems” says Danijar Haferner, a Google DeepMind researcher. The key to Dreamer’s success, says Hafner, is its ability to build a model of its surroundings, utilizing this model to imagine new scenarios that would guide decision-making. “The world model really equips the AI system with the ability to imagine the future,” says Hafner.
Just as Dreamer was able to build on existing knowledge to develop new skills and concepts, future AI models may soon be able to apply these same principles to tackling real-life challenges requiring skills in problem solving and innovation.