Google is making a video game called Unbounded which is inspired by infinitely playable games like Tamagotchi and The Sims and was built on the Gemma-2B LLM. The prototype of Unbounded features a wizard character controlled using natural language inputs. If the wizard gets hungry you can feed it, and if it gets bored you can send it to study.

One major innovation with Google Unbounded is real-time image generation. This is done using LCM LoRA to achieve a refresh rate of close to one second. One of the biggest challenges Google faced while building Unbounded was characters or landscapes changing unexpectedly to align more closely with the prompt given by players to the LLM built on Gemma-2B. To fix this, Google introduced a regional IP-Adapter for improved environment and character consistency within the Unbounded game.

The LLM used for this game was built on Gemma-2B, and distilled using 5,000 user-simulator interaction samples collected from GPT-4o.

These samples from GPT-4o included five characters (a dog, cat, panda, witch, and wizard), 100 diverse environments, and 1,000 text prompts. The creators of Unbounded also concluded that their model, which was built using Gemma-2B, achieved comparable performance to GPT-4o.