Automatic differentiation and Google

 

https://syncedreview.com/2021/06/07/deepmind-podracer-tpu-based-rl-frameworks-deliver-exceptional-performance-at-low-cost-35/

(The URL seems to have a confusing name: it is about automatic differentiation).

Not sure I am understanding the implications of this. I know many users of optimization libraries use finite differences (probably without even knowing it). This is almost always a bad idea: you lose half of the precision in the gradient (solvers typically assume exact gradients) and at the same time, this needs many extra function evaluations. 

Original paper:

https://arxiv.org/pdf/2105.15183.pdf