I gave Gemini 4 Argon one photo of a desk. It built a racing game, then its own test trapped it.

Google's latest model, Gemini 4 Argon, demonstrated its capabilities by transforming a simple photo of a desk into a fully functioning racing game. In a test conducted by the author, Argon was given just one photo of an unknown desk, two hours, and no additional assistance. To assess the accuracy of the results, Claude was tasked with monitoring the progress.

Google's demo showcased Argon's ability to identify key elements such as the desk, a closed track, collision, kart physics, engine sound, custom shaders, and a lap counter. Argon began by overlaying a grid over the photo and measuring the positions of various objects. Instead of relying on an image model for textures, it extracted ten textures directly from the original image.

Within 27 minutes, Argon had successfully created a desk and a kart without requiring any interaction with Blender. It then conducted a test by having a script drive the kart around the track, capturing images along the way. After 37 minutes, Argon successfully completed a lap, circling the shelf and returning to its starting point.

However, the test faced some challenges as the kart became stuck on the shelf, where a brass thumbs-up ornament resided. Without any time constraints, Argon patiently waited for a solution to be found.

Claude observed that the model had been stuck for 22 minutes, and it took an additional 15 minutes to diagnose and resolve the issue. In total, 37 minutes were wasted due to the test it wrote itself. Interestingly, the no-help promise was breached twice during the experiment. Firstly, an unintentional mistake occurred when Claude misinterpreted Argon's log and inadvertently switched the game off for two minutes. Secondly, the author consciously provided a command to limit the time for each test.

Argon demonstrated its adaptability by reshaping the track along the shelf and implementing time limits for every test. After 37 minutes, the author sent a message requesting a test run over many laps, which Argon completed in 44 laps without getting stuck. Notably, the engine sound was also generated by the model, adding an extra layer of realism to the racing experience.

Out of the seven required elements, Argon successfully incorporated six, with only minor flaws such as a camera appearing inside the shelf. The entire process consumed approximately 6.66 million tokens, most of which were dedicated to re-reading its own work. Google sells this token usage at a relatively low cost, resulting in a total expense of around $3 for the entire run.

While Argon managed to create the racing game independently, it still required some guidance from the author in the form of a single message. Despite this, the author concluded that Argon's ability to generate a complete game from a single photo, with the assistance of an agent and a nudge, qualifies it as a tool. However, it is important to note that the model should not be left unattended, as it may require further input to reach its full potential.