It's also well established that an algorithm or architecture alone are not enough to produce a useful model. The same architecture can produce vastly different results depending on the training data, post-training, harness, etc.