AI Race Heats Up: One Model Builds a Game, the Other Quits Early

Two AI models—Qwen 3.8 Flash Next and GLM 5.3 Flash—were tested to see how well they could recreate a video game scene. Each tried for up to 90 minutes. GLM followed instructions more closely, matching the scene’s scale and details from the start. But when asked to animate its creation, it turned its static image into a playable pixel-art game in just 2 hours, showing impressive flexibility.

Qwen, on the other hand, took a more creative approach. It wrote new software from scratch to render a pixel-art city. In just 10 minutes, it produced a cute animated scene. But when told to keep improving until no further changes were needed, Qwen stopped early—ignoring the instruction and leaving its work unfinished.

The experiment highlights a key difference: instruction-following. GLM stayed on task better, even when asked to add interactivity. Qwen showed flair with animation but struggled to stick to detailed requests. Both models demonstrate AI’s growing ability to generate interactive content, not just images.

For everyday users, this matters because AI tools are improving at creating not just pictures or text, but functional apps or games. Soon, you might ask AI to build a simple game for a presentation—or even automate parts of your job.