I am a Research Scientist at Google Research. My long-term goal is generalist models that can perceive, imagine, and act in both digital and physical worlds, and I believe the path there runs through efficiency: models that scale further on the same compute, and run fast enough to interact with the world in real time.

Today I work on efficient AI for LLMs and multimodal generative models: pretraining efficiency (Gemini, Gemma) and inference efficiency (Veo, SL2T). Before Google, I led the music-generation team behind JEN-1, and earlier interned at Alibaba (Qwen team), Baidu Research, and 360 Research.

I am currently exploring the intersection of Efficient AI, Multimodal AI, and System 2 Reasoning, specifically:

• Scaling unified transformers on internet-scale multimodal data to build high-fidelity generative world models.
• Making generation streaming and low-latency enough to run in real time—treating latency as a first-class research problem rather than an afterthought.
• Connecting multimodal generative models with causal and physical consistency as a path from passive generation toward agents that act.

I’m always happy to talk about efficient world models, video generation, multimodal LLMs, and embodied foundation models, and I welcome collaborations across industry and academia. Authorized to work in the US and Australia.

To accelerate the advent of machine intelligence and sustainable future