In this talk we will look at a series of foundational time series (TS) models. Starting with a discussion of classical approaches like ARIMA and their limitations, we will explore why traditional machine learning models often fail in this domain. Introducing "One-for-All", a model designed to address many of these challenges, we highlight its unique features and training procedure, while addressing its limitations. Finally, we propose some ideas and share some insights to enhance the model's capabilities over long horizons, ensuring robustness and accuracy in extended time series forecasting.