It’s impressive how Google’s models are evolving; they’re getting bigger and smarter every month. Their power to generate almost anything—videos, images, audio, and text—is absolutely crazy.
Gemini has improved so much recently. Now, it can actually recreate entire games, focus precisely on what you need, and build truly complex applications.
But I want to talk about something that REALLY impressed me. Today, I was scrolling through YouTube and saw this video: https://youtu.be/PPYeNr-W14s?si=olLlsYmYoqkS8zaU. It apparently showed a competition between two AIs, Chat GPT 5.1 and Gemini 3, with the goal of recreating the game ‘Terraria’. Chat GPT did well, but it wasn’t particularly impressive. However, when it was Gemini’s turn, it crushed Chat GPT after just two tries! It created a masterpiece, even adding more features to the game, like zombies. Oh my god, that got me thinking: how is this even possible? It’s clearly Gemini’s vibe coding power.
This ability is a genuine game-changer, demonstrating that AI is now capable of synthesizing highly complex, interactive systems. It’s not just about writing clean code; it’s about understanding deep game design principles and functional requirements. For instance, Gemini didn’t just generate blocks and basic movement, but it integrated superior textures, and even correctly inferred that the world should start with ores like iron and coal. This is a subtle yet crucial detail that shows a deep, conceptual grasp of the game’s progression loop right from the start.
And those zombies! The fact that after just one refinement prompt, Gemini could deliver fully functional hostile mobs that the player could interact with and defeat, alongside better inventory management and a visual progress bar when breaking blocks, is absolutely insane. This level of successful dynamic interaction and state management is what separates a basic coder from an AI architect. This is the new threshold: the AI is able to take a high-level, ambitious prompt like “make Terraria from scratch” and, within minutes, lay down a code base that a human developer would take days or weeks to create. It makes you wonder what kind of complex applications and virtual worlds will become standard by next year, now that these models possess the power to interpret the spirit of a task and code the most complex components with such effortless refinement.