Microsoft researchers on Wednesday introduced a new Artificial Intelligence (AI) model that can generate an environment for 3D gameplay. Dubbed the World and Human Action Model (WAHAM) or Muse, the new AI model was developed by Tech veteran research game intelligence and teachable AI Experience (TAI X) teams in collaboration with Xbox Games Studios’ Ninja Theory. The company said that large language models (LLM) can help sports designers in the ideation process, as well as help to generate game visual and controller tasks to help creative in game development.
Microsoft unveiled AI model
One in blog postThe Redmond-based technology giant expanded the museum AI model. This is currently a research product, although the company said that it is opening the weight and sample data of the model for the WAHAM performance (a concept of a visual interface to interact with the AI model). Developers can try models on Azor AI Foundry. A paper The description of the technical aspects of the model has been published in the magazine.
Training a model on such a complex area is a difficult offer. Microsoft researchers collected large amounts of human gameplay data from the 2020 game bleeding edge, a game published by Ninja Theory. LLM was trained on a billion image action couple, equivalent to seven years of human gameplay. The data is collected morally and is used only for research purposes.
Researchers said that increasing model training was a major challenge. Initially, the muse was trained on a cluster of Nvidia V100 GPU, but then extended to several NVidia H100 GPU.
Coming to functionality, the muse AI model also accepts the visual input along with text signals. Additionally, once a game atmosphere arises, it can be increased by using controllers tasks. The AI responds to the movements made by the user to introduce the new environment aligned with the initial signal, and correspond to the rest of the gameplay.
Being a unique AI model, specific benchmark tests cannot properly evaluate its capabilities. Researchers highlighted that they have tested the LLM in the internal matrix on the interior such as stability, diversity and perseverance. Since it is a research-centric model, the output is limited to only 300x180p resolution.
