Did NVIDIA predict OpenClaw? | Jensen Huang and Lex Fridman
Jensen Huang explains how NVIDIA anticipated agentic AI systems like OpenClaw by reasoning through what digital workers would need to do - access files, use tools, and do research. He describes how this led to designing the Vera Rubin rack system before OpenClaw's release and NVIDIA's subsequent security contributions.
Summary
In this conversation, Jensen Huang discusses how NVIDIA predicted and prepared for agentic AI systems like OpenClaw through logical reasoning rather than insider knowledge. He explains that by thinking through what a digital worker would need to accomplish - accessing ground truth data from file systems, conducting research, and using existing tools - it became obvious that AI agents would need these capabilities. Huang uses the analogy of a humanoid robot, arguing it would more likely use existing tools like microwaves rather than transform its hands into different implements. This reasoning led to NVIDIA's design of the Vera Rubin rack system, which includes storage accelerators, new CPUs, and additional components specifically for running agents, contrasting with the previous Grace Blackwell racks designed for LLM inference. Huang notes that NVIDIA was discussing these agentic system concepts at GTC two years prior to OpenClaw's release. He credits OpenClaw with doing for agentic systems what ChatGPT did for generative systems, making the technology accessible to consumers. The discussion also covers security concerns, with NVIDIA contributing Open Shell and Nemo Claw to address these issues through a 'two out of three' security model that limits agents' capabilities around accessing sensitive information, executing code, and external communication.
Key Insights
- Huang argues that predicting agentic AI needs required simple reasoning rather than insider knowledge - digital workers must access ground truth data, do research, and use existing tools
- Huang claims that AI agents will use existing tools rather than replace them entirely, comparing it to a humanoid robot using a microwave instead of beaming microwaves from its fingers
- Huang states that NVIDIA was discussing agentic systems matching OpenClaw's architecture at GTC two years before OpenClaw's release
- Huang argues that OpenClaw did for agentic systems what ChatGPT did for generative systems in terms of making the technology accessible
- Huang describes NVIDIA's security approach as giving agentic systems only two out of three capabilities - accessing sensitive information, executing code, or communicating externally - but never all three simultaneously
Topics
Full transcript available for MurmurCast members
Sign Up to Access