I agree - I find the particular tone and pitch he uses when saying that to be painful (it literally hurts my ears). The transition from learning to review is a good idea; the execution can be improved.
I don't think that the issue is computers only understanding numbers. Even numbers are not directly understood by computers and need to be represented via combinations of "1"s and "0"s. Perhaps the issue is more related to the fact that neural networks and similar models can't directly deal with non-numerical data, and hence the need for a numerical representation before any training can take place. You make a great point about the higher computational cost of using natural or default numerical representations for items like words and images, which explains the need for an 'embedded' representation.
Can you please make a video that showcases how we can generate custom word embedding on a custom dataset from scratch? Without using anything pre-built? Say IMDb dataset? and then later load them to train a classification model?
Thanks for video, you explain things in different difficulty level, that works. The quize and stuff is not working, for me breaks the flow of the content.