This is one of the reasons why I think AI algorithm development is going to appear to go nowhere until it is nearly done, and then change more in a matter of months than it has in decades before. Nobody really has a clue how to teach the English language, or how to walk, or how to create goals. Once we can get the problems humans are naturally good at solving with very little help out of the way, programming an AI to understand the scientific method, existential risks, and how to overcome its own biases is most likely going to seem easy in comparison.