IMO the rate of change or volatility of model changes, RLHF, user-behavior, the definition of 'better' is much higher than the rate of change of model degradati

IMO the rate of change or volatility of model changes, RLHF, user-behavior, the definition of 'better' is much higher than the rate of change of model degradation (if any) so it's hopeless to try and measure this