I see these things and while interesting they are just a lab curiosity.
From the paper:
By tuning the phase delays of each waveguide at each layer of the directional coupler mesh and the coupling region’s effective optical length using heaters it should be possible to achieve an arbitrary unitary transfer matrix
As these phase sections are thermally tuned, they are slow and will have thermal crosstalk to their neighbors. While this is manageable it’s not exactly fast to create a new matrix transfer. Additionally, these devices drift over time and operating temperature. Changes over time are due to changes in local stress that cause refractive index to change. So this device as published has no ability to put the MZM bias into a known state and keep it there. Drift is a part of working with photonic devices and even when temperature controlled (very inefficient) they still need compensation. This compensation is the difference between a lab device and something that could be commercialized.
Furthermore such devices have what we call parasitic reflections. When the phase is adjusted, the parasitic reflected phase is also adjusted creating an error in the output. For an analog computer that can only look at amplitude any numerical resolution would have to be coarser than the size of these parasitic reflections.
This is just the tip of the iceberg. Yet none of these papers ever address these problems.