NVIDIA World Model Researcher Ruilong Li Departs
NVIDIA world model researcher Ruilong Li announced today is his last day at NVIDIA.
He graduated with a bachelor's degree in Mathematics and Physics from Tsinghua University and later obtained a PhD in Computer Science from UC Berkeley, under the supervision of Angjoo Kanazawa at BAIR. He has previously conducted research at Google Research, Meta Reality Labs, and NVIDIA.
His research has consistently focused on machine reconstruction, generation, and simulation of interactive 3D worlds: early contributions include nerfacc and Nerfstudio, and as a core author of gsplat, he promoted the Gaussian Splatting ecosystem; at NVIDIA, he led NuRec for real scene reconstruction, OmniDreams for generative driving simulation, and FlashDreams for high-performance interactive world model inference systems.
He believes that "the future of video generation is simulation," emphasizing that models need to understand actions, predict world changes, and continuously generate interactive data to support training for robotics and autonomous driving.
Top 3D and world model talents are leaving large companies, with funding and attention shifting towards independent or new platforms to advance interactive simulation. Beneficiaries will be researchers and related applications that continue to delve into world models, while laboratories relying on internal talent will face pressure. This event drives an increase in research priorities from reconstruction to real-time generative simulation.
Source: Public Information
ABAB AI Insight
Ruilong Li's academic path from Tsinghua to Berkeley has long focused on neural rendering and 3D reconstruction, establishing influence through open-source tools (nerfacc, gsplat) before entering NVIDIA, driving a complete chain from NuRec to FlashDreams.
Resources are shifting from real data reconstruction to generative and interactive world models, motivated by the desire to break through the limitations of "only reconstructing the captured world" and achieve action-driven real-time simulation to generate training data.
Similar to the early NeRF researchers moving towards Gaussian splatting and generative models, the current 3D/world model field is transitioning from static reconstruction to dynamic interactive simulation, with the industry as a whole shifting from rendering tools to robotics and autonomous driving data engines.
This is essentially a technological replacement, where video generation shifts from prompt-based output to action-responsive simulation, with the mechanism being autoregressive world models combined with low-latency inference, allowing generated content to directly become a closed-loop training source.
ABAB News · Cognitive Laws
The endpoint of video generation is interactive simulation.
Reconstruction is just the starting point of generation.
Understanding actions determines the value of world models.