Focusing on $pp\to ZZ\to 4\ell$ which is phenomenologicaly interesting at the LHC, I will discuss the remarkable progress from proof-of-concept to implementable approximate two-loop amplitudes. The idea is to replace the exact amplitudes in monte carlo event generators with very precise approximations. This can be naturally achieved with machine learning algorithms. Most of the effort to achieve this goal, however, centered around deconstructing the amplitudes, leveraging their symmetries, and conditioning them in order to build an optimal set of functions to approximate. The end result is that the two-loop virtual amplitude can be evaluated with percent or sub-percent precision (which propagates to less than 1 part in $10^4$ on the differential cross-section) and in milliseconds. This translates to more than a thousandfold speedup over the exact amplitudes. Finally, I will discuss how this work generalizes to other amplitudes starting with the straightforward extension the full di-boson set of processes and the gluon-initiated two-loop matrix element to more generic cases.