AI Trained on Old Scientific Papers Makes Discoveries Humans Missed

I’m skeptical. Sounds VERY much like overfitting. Remember the Shakespeare and CS paper generators? Word2Vec isn’t real AGI, it’s a “magic trick”

EDIT: oh, it’s a Vice article

Our track record is already embarrassing :) https://www.pnas.org/content/112/24/7426

Now if only we were less unsure of the accuracy of published results

This feels like a good use case for HN's mysterious title revision policy. Something that sounds less like we blew past the singularity overnight?