It was introduced and popularized by the Stanford Institute for Human-Centered Artificial Intelligence when first large-scale multitask models started to appear in 2019-ish.