Google Gemini 3.5 Pro Delayed: When Scaling Laws Hit the Hard Wall of Code

Last week at our group meeting, my advisor asked me to try using the latest large models to write a custom reward function for the lab's reinforcement learning environment. I first tried GPT-4o, generated three versions, two of which wouldn't run, and one had logic bugs. Then I tried Gemini 1.5 Pro, but the results were worse—it mixed up gym interfaces with PyTorch tensor operations, causing immediate errors. My advisor sighed, "Current models still can't handle this kind of engineering code." I casually replied, "I heard Google's latest flagship model got stuck because its coding skills are too bad."