> Isn’t new data required? No. For these systems to improve on their defects they should rely less on having seen something very specific in the training data,

> Isn’t new data required? No. For these systems to improve on their defects they should rely less on having seen something very specific in the training data, and more on reasoning and 'understanding' (forming conceptual models). The amount of training data available is already vastly more than strictly necessary given more data-efficient algorithms. I didn't have to read the whole internet to learn anything I know.