“The AI research community is in a "bad equilibrium" where labs publish benchmark results as a simple grid, despite knowing it's a flawed evaluation method, due to industry inertia and expectations.”