How to build fearless, human-centred AI: Fei-Fei Li and Reid Hoffman

Executive overview

Language models are one keyhole into intelligence — not the whole door. The next phase of AI is world modeling: machines that can represent, reason about, and participate in the physical and virtual world beyond text.

Spatial intelligence is the cognitive foundation of embodied action, from early evolution to robotics. Without it, AI cannot build, discover, or interact — only describe.

Trust in the AI age cannot be outsourced to machines; it must be built into products, governance, and culture from the start.

What world modeling is and why it matters

• Language captures symbolic description; the world beyond language includes geometry, physics, dynamics, and 3D space
• World modeling — representing and generating interactive, spatial environments — is the next frontier after LLMs
• Creators, designers, and industrial users all benefit: immersive experiences, simulation, and embodied AI all depend on it
• Simulation is critical for robot learning, just as it has been for self-driving cars
• Data is harder to obtain than for language: video helps, but 3D geometry and physics data are not easily scraped from the internet

Why spatial intelligence is the brain of embodied AI

• Evolution explains the link: the Cambrian Explosion (~530 million years ago) marks the beginning of perception, nervous systems, and movement
• Perception evolved not as passive sensing but as the foundation for action and interaction
• Complex movement — from tool use to surgery — requires nuanced spatial world understanding
• Robots operate in 3D and must touch things correctly; they are fundamentally harder than cars or Roombas
• The robotics journey will take time; trust infrastructure, supply chains, and manufacturing must all mature

Spatial reasoning lifts all of intelligence, not just robotics

• Building the pyramids required abstract geometry — not transactional perception
• The discovery of DNA's double helix required deep spatial reasoning; language alone could not have produced it
• Spatial AI capability will augment human discovery, not just automate physical tasks
• AI as civilizational technology means it will be embedded wherever there is a chip — and chips are already everywhere

On hype and timing

• AI is not overhyped as an intellectual future; it is the new computing
• Self-driving cars went from Sebastian Thrun's Nevada demo to Waymo over 20+ years, even with deep learning acceleration
• The car industry had 100+ years of infrastructure; robotics does not — expect a longer runway
• Long-term trajectory is clear; near-term timelines for specific applications are routinely underestimated

Building trust and fearlessness

• Trust is fundamentally human — individual, community, and societal — and cannot be delegated to AI systems
• Founders should embed trust and human agency into products from day one, regardless of sector
• Governance must evolve beyond company norms into societal-level frameworks
• Fearlessness means removing the shackles on creativity, courage, and execution — running toward uncertainty and contrarian hypotheses
• Tasks with uncertain outcomes demand more creativity than certain ones; that is where breakthroughs happen