Goldman tackles AI's missing link: The 'world model' that every AI godfather is racing to figure out
Goldman Sachs Global Institute report arguing "world models" are the next major leap in AI — moving beyond pattern-matching text prediction toward systems that internally understand physical and social reality.
Key People
- Yann LeCun — former Meta chief AI scientist; founded AMI Labs; developed joint-embedding predictive architecture (JEPA) for world models
- Fei-Fei Li — Stanford researcher; created ImageNet; founded World Labs focusing on "spatial intelligence"
- George Lee — co-head of Goldman Sachs Global Institute
- Dan Keyserling — managing director at Goldman Sachs Global Institute
Core Argument
Current LLMs "learned about the world by reading what humans wrote about it" without "ever encountering reality itself." They lack "the internal sense of the world those patterns describe" and "generate this understanding through second-order interpretation."
Two tracks of world models:
-
Physical world models — teach AI physics rules (gravity, friction, thermodynamics) through simulation. Robots practice millions of failures in digital environments before acting in the real world.
-
Virtual/social world models — digital environments populated by AI agents with goals and incentives, designed to approximate real-world behavioral profiles. "Patterns emerge. Markets behave. Organizations respond. Crises cascade."
Key distinction: "World models are not forecasts." They "don't predict the future in any narrow sense; they're meant to reveal plausible futures and expose hidden dynamics."
Investment Implications
Current AI infrastructure projections assume scaling LLMs. The report suggests world models create compute demands "not yet reflected in consensus supply-and-demand forecasts." Simulation needs "purpose-built data pipelines, synthetic data generators, and physics-based engines."
"Competitive advantage might depend as much on who trains the largest model as who builds the most faithful simulations of reality."
Closing Frame
"If large language models give AI fluency, world models give it situational awareness."