First, demonstrate how classic statistics centered at type-I error control fail to consider the totality of scientific evidences, fail to address the bias issues in data analysis of adaptive trials, fail to address the controversial findings in multiregional trials, and fail to deal with low power issue in rare disease trail and precision medicine. Then, propose similarity-based machine learning (SBML) and demonstrate its superiority to the classic modeling methods in terms of predictions in clinical trials. Further, illustrate the utilization of SBML in a drop-loser adaptive trial design, monitoring and adaptations. Machine learning methods have been used widely in drug discovery for nearly two decades, but there is little research on such a technology in clinical trials, in which sample size is often limited. SBML and other artificial intelligences emphasize on learning and prediction of drug benefits to future patients, and on real world experiences; our vision for the future of drug development is: smaller trials, stage-wise marketing authorization, and enhanced post-marketing monitoring, with the support of artificial intelligence.