Imagine trying to predict a student's yearly test scores based on just a few weeks of performance. Or forecasting a patient's five-year health outlook after a brief initial treatment. This is the challenge of evaluating long-term strategies, or 'policies,' where the final outcomes take a long time to observe. Researchers are tackling this 'short-long' prediction problem, especially for AI policies that make sequential decisions. Traditionally, evaluating a new AI policy requires extensive testing over the entire period of interest. This can be incredibly time-consuming and expensive in real-world scenarios like personalized education or healthcare. A new research paper introduces innovative techniques to predict long-term AI policy outcomes using minimal initial data. The key is to combine this early data with historical data from previous policies. One method, called SLEV (Short-Long Evaluation of Value), treats the problem like predicting under a changing environment. It leverages the fact that the performance of a policy often depends on the situations, or states, it encounters. SLEV uses the short-term data to predict the long-term results, adjusting for the differences between the new policy's behavior and the historical data. Another approach, SLED (Short-Long Estimation of Dynamics), goes a step further by modeling how the AI's actions change the situation over time. It learns a general model from historical data and then adapts it to the new policy using a small set of initial observations. This adaptation helps SLED simulate the future states the policy is likely to encounter and predict the ultimate long-term outcome. Experiments on simulated healthcare scenarios (HIV treatment and kidney dialysis) show these methods provide more accurate predictions than traditional approaches. They also demonstrate the potential to quickly identify risky policies that might lead to poor long-term results. This research opens exciting possibilities. Imagine quickly identifying effective teaching methods or safe and efficient battery charging protocols. While more research is needed, these early results offer a promising path to faster innovation and improved decision-making in a variety of fields.