Recent advances in AI not only created promises for what AI can do, but also introduced questions about how to bring this promise to reality in real-world applications in a responsible way. In this talk, I will describe my journey at Microsoft Research from being amazed by the sparks of GPT-4 to understanding limitations of current family of models and driving research on what comes next. I will discuss research directions we are pursuing to make future AI systems more efficient, sustainable, controllable and valuable through innovations in model training, agent technologies and engineering practices. I will conclude with reflections on our unified responsibly in balancing the promise of AI with rising risks and concerns.