It is the idea that humans think any task easy for humans is hard for robots and things that appear hard for humans are easy for robots. I can see the point she is trying to make as we have seen a bunch of benefits of using data in developing large language models but I am also generally a bit skeptical of the idea that we use data for everything when sensors are inherently noisy- if all we care about is outputs then this is fine but separating process (or signal) from noise is a deeper thing that I think we will lose.