I think a human-centric multi-modal dataset (i.e.: head mounted binocular cameras, steroes mics, text transciption, skeletal pose, skin area contact) would be a great starting point for building human-level artificial intelligence. Do you know if such a dataset exists or if there are open data efforts to build one. (It'd be terrible if Tesla or another private company were the first to become an AI superpower in this domain. It's unlikely they'd release any data)
https://en.wikipedia.org/wiki/List_of_datasets_for_machine-learning_research list many AI datasets. Note that private companies have released plenty of very useful AI datasets.