“ You need a place where robots can learn to be robots,” he emphasises.“ If you buy a window-cleaning robot, you don’ t want it to learn the laws of gravity in your kitchen.”
Each working robot then becomes a source of new data, feeding a process known as federated learning, in which many devices contribute to improving a shared AI model.
“ Each of us gains experience doing our jobs, and we meet once or twice a year at a conference to share it,” Michael points out.“ But robots can share experience by updating their AI model in the big data centre every day – or even every hour.”
The AI safety debate While a raft of AI leaders have recently called for the pace of model development to slow, Michael firmly disagrees.
“ I think slowing down is exactly the wrong thing to do,” he asserts,“ because the only way to make AI safe is to use AI. You need to accelerate AI development to make it safe.”
Michael accepts that, like any tool, AI can fall into the wrong hands – and the industry must build safeguards. One practical example is NVIDIA’ s OpenShell, an open-source runtime that places AI agents in isolated sandboxes and enforces limits on
30 October 2026