Machine Learning based weather prediction (MLWP) is becoming increasingly common. Trained on reanalyses, these models incorporate no underlying physical equations, but rather learn common patterns in the data. The largest advantage of MLWP is the reduction in computing resources for each forecast compared to traditional numerical weather prediction (NWP). Here we consider 3 years of the Global Forecast System v16 (GFSv16) operational forecasts and two versions of GraphCast. The GraphCast versions are identical in their training on ERA5 reanalysis and differ in their initial conditions, GFS vs Integrated Forecast System (IFS). We evaluate these models for their representation of tropical dynamics and coupling to large-scale convection. A well-known characteristic of MLWP is that forecasts tend to be more "smooth" than NWP forecasts and analyses. We quantify this behavior using space-time power and coherence spectra and show that MLWP forecasts loose power at higher wavenumbers and frequencies very quickly with lead time. This leads to less small-scale variability and higher coherence for larger-scale phenomena in MLWP, for example for convectively coupled Kelvin waves. Further we show preliminary results for precipitation-convection coevolution and differences in GraphCast forecasts related to the initial conditions.