How do they create first sample from which to "learn" without importing current bias?