Just landed some really, really dumb changes to T3 Code that reduce GPU utilization by ~85%. LLMs failed to find any of them. Both Fable and gpt-5.6 kept getting caught on things that didn't matter. Most weren't even in the render path! They had all the tools they needed, but they got too distracted by the code to solve the real issues. I did get them to be helpful. I had gpt-5.6 write me a super useful script I could paste into the console to enable/disable the CSS features I assumed were…