Drawing from a massive 200,000-hour gameplay video library, this innovative model can transform any visual or conceptual stimulus into a playable 2D platform game.

Google DeepMind revealed Genie, an advanced AI system capable of generating playable games from straightforward prompts, in a demonstration this Monday. Genie, which has been educated on the intricacies of game design by analyzing extensive gameplay footage, is the result of a joint effort between Google and the University of British Columbia.

Under the moniker Genie, standing for Generative Interactive Environments, the AI can craft side-scrolling platform games reminiscent of classics like Super Mario Brothers and Contra from just an image, based on user instructions.

Google DeepMind articulated the recent advancements in generative AI, noting the ability to create innovative and imaginative outputs across various formats, including language and imagery. "We're introducing a novel concept in this field: generative interactive environments, with Genie at the forefront," stated Google DeepMind.

The technology behind Genie, as explained by the Google team, incorporates a latent action model that interprets potential actions in video sequences, a video tokenizer that translates video frames into distinct tokens, and a dynamic framework predicting subsequent frames.

Tim Rocktäschel, a developer at Google DeepMind, emphasized on Twitter the project's scale and the absence of preconceived biases. "Leveraging over 200k hours of 2D platformer video footage, we've developed an 11B parameter world model. Through this, Genie autonomously learns varied latent actions, enabling consistent character control," he mentioned.

Moreover, Rocktäschel highlighted Genie's versatility in adapting different media into interactive games. According to the Google DeepMind study, the AI can transform any visual into an interactive 2D landscape.

"Genie has the ability to animate user-generated designs like drawings, turning them into interactive environments," Rocktäsche added. While showcasing Genie's proficiency in generating 2D environments from texts or visuals, he also pointed out the AI's broader capabilities, including instructing other AI entities about 3D environments.

Exploring further, he discussed how Genie's training included data from robotics, allowing it to develop action-controllable simulations, marking a step towards achieving comprehensive world models for artificial general intelligence (AGI).

Google DeepMind elaborated on how Genie's dataset was compiled, focusing on publicly shared videos that were relevant to gameplay, aiming for content that prominently features 2D platform gaming.

The team at Google DeepMind envisions a future where AI agents can continuously learn in endlessly evolving worlds created by Genie. "Our proof of concept demonstrates Genie's learned latent actions can be applied in real-world settings, which only scratches the surface of future potentials," they noted.

The surge in generative AI investments by tech giants like Google, Microsoft, and Amazon follows the successful launch of OpenAI's GPT-4. Google also recently introduced a subscription model for its Gemini AI, transitioning from the earlier Google Bard.