معرفی
Tri Dao serves as Assistant Professor of Computer Science at Princeton University and Chief Scientist at Together AI, where he bridges machine learning theory with systems optimization. His work focuses on developing efficient algorithms for deep learning training and inference, particularly in hardware-aware computation and sequence modeling.
Education
- Ph.D. in Computer Science, Stanford University (2023)
Research Focus: Dao pioneers techniques for long-range memory in sequence models and structured matrices for compact deep learning, evidenced by breakthroughs like FlashAttention and Mamba. His research integrates theoretical insights with practical systems constraints, targeting memory efficiency, computational speed, and scalability in large language models. Key contributions span attention mechanisms, state space duality, and hardware-aware algorithm design.
Publication Impact: Recent work (2022-2025) reveals a trajectory from foundational theory (e.g., state space duality in Mamba-2) to deployable systems (FlashAttention-3), addressing critical bottlenecks in LLM training and inference. Publications consistently emphasize hardware-aware optimizations, memory efficiency, and linear-complexity alternatives to attention mechanisms.
Scientific Recognition
- Outstanding Paper Honorable Mention (MLSys 2025)
- Outstanding Paper Honorable Mention (NeurIPS 2024)
- Outstanding Paper (COLM 2023)
- Best Paper award (ICML Workshop 2022)
- Stanford Open Source Software Prize (2024)
- Outstanding Paper runner-up (ICML 2022)
Mentorship and Funding: Dao actively advises PhD students Ted Zadouri, Berlin Chen, and Wentao Guo on sequence modeling and systems research. While specific grants aren't detailed, his dual Princeton/Together AI roles indicate sustained funding for open-source LLM development and hardware-aware ML research.
Research Ecosystem: Leading Princeton's machine learning systems efforts while driving Together AI's open-source initiatives, Dao operates at the academic-industry nexus. His work directly informs real-world LLM deployment through tools like FlashAttention (adopted by major frameworks) and Mamba (enabling efficient long-context modeling).
