We need benchmarks that can distinguish between continuous learning and long-context extrapolation.