The concept of “Present is the key to the past” is widely applied in paleoclimate studies. By cross-referencing different proxy records, paleoclimatologists can piece together a picture of Earth's climate history, creating robust climate reconstructions spanning hundreds, thousands, or even millions of years, which are crucial for understanding long-term climate variability and trends. However, the potential of proxy records for interpreting the present and predicting future climate states remains underutilized. As artificial intelligence (AI) continues to rapidly evolve, data-driven models based on deep learning (DL) have been gradually applied to the domain of atmospheric science, and have made breakthroughs in weather forecasting and climate prediction, achieving results comparable to traditional numerical weather prediction (NWP) models or even better. ClimaX, a novel AI model built on Transformer architecture, is one of them. ClimaX is the first fundamental model for weather and climate, by pre-training heterogeneous datasets spanning different variables, spatio-temporal coverage, and physical groundings, it can be fine-tuned for various downstream tasks such as climate projection. Therefore, here, we explore whether machine learning (ML) can serve as a powerful tool to connect the "past-present-future" in climate science. We investigate the application of the ClimaX model to paleoclimate research, using climate fields reconstructed from proxy records by paleoclimate data assimilation (PDA) to pre-train the model and fine-tune it for future climate projections. As a case study, we focus on the tropical Pacific, conducting experiments on decadal climate variability projections. This approach demonstrates the potential of AI to integrate paleoclimate data for both the interpretation of present conditions and the prediction of future climate trends.