Ethan Mollick

Three findings from a trial use of a GPT-4 tutor for Harvard's intro CS course: 1) Students found it useful (88% found it helpful) & used it a lot 2) The AI made mistakes, even with RAG (81% accurate) 3) It was pretty cheap Lots of potential but R&D needed

The O*net papers are great, but they don’t tell us how jobs will change.

This is such an interesting thread, and a lightning tour of what advanced manufacturing tools look like today, from industrial robots to lasers. (As a side note, this is also very old-school, now mostly lost, Twitter, where experts put together long threads on narrow topics)

Three hints to the future of AI in this paper: 1) It is close to outperforming the top humans at geometry (superhuman performance is possible) 2) It combines a language model & other AI approaches (hints of the value of tools) 3) It trained on AI-generated data (big for training)

I replied directly to the post. Its a great paper, and it really asks a different question than our paper (or other ones looking at performance impacts of AI). The AI isn’t doing, here, instead it is advising.