Fei-Fei Li Positions Spatial Intelligence as Independent AI Frontier

Current large language models remain fundamentally incomplete without architectures that perceive and reason within three-dimensional space, according to computer scientist Fei-Fei Li.

Speaking in an interview hosted by Stanford University's Hoover Institution, the World Labs co-founder framed spatial intelligence as the decisive frontier for artificial intelligence research. The argument positions spatial reasoning not as a downstream feature of text prediction, but as a separate foundational axis of machine intelligence.

Li drew on evolutionary biology to highlight the divergence between current frontier model development and biological cognition. Vision and spatial comprehension evolved over 500 million years ago during the Cambrian explosion, long preceding the emergence of human language. Biological agents survived by parsing depth, calculating trajectories, and manipulating physical matter. Language developed relatively late in evolutionary history as an abstraction built atop these grounded sensory systems.

Dominant AI research paradigms remain centred on sequential token prediction. While autoregressive transformers exhibit fluent reasoning across textual benchmarks, they lack an intrinsic grasp of geometry, physical affordances, and persistent spatial environments. Flat pixels and two-dimensional video frames capture only projections of reality, omitting the underlying three-dimensional coordinate frames and physical constraints governing the real world.

Spatial intelligence requires systems to infer complete 3D structures from partial sensory inputs. Such models must resolve occluded geometries, anticipate physical interactions, and predict how scenes alter when an agent shifts perspective. This capability forms the technical prerequisite for autonomous robotics, real-time simulation, and spatial computing.

World Labs focuses on developing large world models designed to perceive, generate, and interact with persistent three-dimensional environments. Rather than relying on flat probabilistic image outputs, the approach prioritises models that preserve geometric consistency and causal mechanics across space and time.

Frontier labs continue to concentrate compute on scaling language and multimodal corpora. Yet physical tasks demand models that operate directly on spatial geometry and material physics. Whether spatial world models can scale with the data efficiency of text transformers, or require entirely distinct inductive biases, remains an open empirical question.