There's some previous discussion of the "beats top professionals" part of this here: https://news.ycombinator.com/item?id=13534778 https://news.ycombinator.com/

There's some previous discussion of the "beats top professionals" part of this here: https://news.ycombinator.com/item?id=13534778 https://news.ycombinator.com/item?id=13534778 https://news.ycombinator.com/item?id=13358138 https://news.ycombinator.com/item?id=13358138 Note that this shouldn't really be a surprise - it's very likely that the best HUNL AIs were as good or better than the best humans maybe 5 years ago. The recent developments over the last couple years are more about widening the gap and increasing the confidence in the results. I haven't fully read this paper yet but my impression is that it's mostly about architectural and abstraction improvements, and possibly just chucking more hardware at it.