I think there's been a big impact on having commonly used tools too. Tensorflow and the like have allowed people to share and easily build on each others work,

I think there's been a big impact on having commonly used tools too. Tensorflow and the like have allowed people to share and easily build on each others work, including on shared models. Not unique to 2016, but I remember the struggles of trying to get hacked together matlab code working that it turned out was actually slightly different than must have been run for the paper. Now I'd expect a lot of work to come with a setup that anyone can easily get and try themselves.