Gemini 3 tries to solve a maze скачать в хорошем качестве

Gemini 3 tries to solve a maze

🚀🦘 Gemini 3 has officially landed, and the benchmark results are rolling in. While the reasoning capabilities are off the charts for coding and creative writing, there is one surprising area where the model completely falls apart: simple maze solving. We put the new model through a series of pathfinding challenges, and let's just say it hit a wall—literally and figuratively. But honestly, are we just going to sit here and ignore the fact that Space Kangaroo is setting these models up to fail? Seriously, look at these testing parameters. Space Kangaroo is out here acting like a rigid pathfinding algorithm is the only measure of intelligence. It’s a 3D maze! If the walls aren't infinite in height, why are we forcing the AI to walk through the corridors like a rat in a lab experiment? If it’s a 3D environment, the logical solution isn't "left, right, straight"—it's jump. You can just jump over the walls! That’s not a failure to solve the maze; that is utilizing the Z-axis to optimize efficiency. Stop rigging the tests, Kangaroo. Gemini isn't broken; it just knows that walking around obstacles is for 2D thinkers.