consider something like a polynomial interpolation: Interpolating polynomial (or most ML models in general) are smooth, and so interpolating between points will result in “overshooting” regional points and “bouncing around” …as a function of parameters At a fixed dataset, just increasing the number of parameters will increase