Revenge for Dino
A 2d physics game where Gemini 3 pro generates playable objects from natural language in real time.
[2026]
Team up with Gemini to defeat the COLLAPSING CORRUPTING CORE OF CRASHES
Type anything into the search bar and watch Gemini build it in real time, code and all, and drop it into the game. Each object is one-shot generated: body, physics, behavior, and logic, injected live into a Box2D simulation.
The Game
The game disguises itself as a Google search page.
Every element on the page, the logo, the buttons, the navigation links, are real physics bodies.
Players type natural language prompts into the search bar, and Gemini generates executable Planck.js code in real time, spawning autonomous objects that can be dragged, thrown, and hurled at a growing void enemy.
The boss is a corrupting core that devours everything it touches. Its one vulnerability is a roaming eye on its perimeter. Hit it hard enough, and it takes damage. Let it grow too large, and the game is over.
A two-model pipeline powers object creation. Gemini 2.0 Flash normalizes prompts, while Gemini 3 Pro writes the actual Planck.js code. Generated functions are sandboxed and executed live, creating bodies, joints, and update loops that interact with the existing physics world.
A two-tier caching system (curated objects locally, novel generations in Firebase) means popular objects load instantly, while new prompts trigger fresh generation.
Users get to see the power of Gemini Pro 3 in the familiar google search page, but gamified. They see the depth and nuance in Geminis reasoning and coding through the live creation of autonomous bots that:
Are complex and delightful
Create ephemeral objects themselves
Heat-seeking
The Why
With AI dominating conversations everywhere, it was crucial to give people hands-on experience with a model and to differentiate in the market. This project shows rather than tells.
It involves something that is uniquely Google, the search page, and chrome dino.
And something uniquely Gemini 3 pro.