They are the people who made artificial intelligence what it is. They built the datasets, designed the architectures, and trained the systems that now write our emails, generate our code, and pass the bar exam. And increasingly, quietly, they are all working on the same problem—a problem that implies today’s most powerful AI, for all its staggering capability, is still missing something fundamental.
Goldman Sachs has a name for what’s missing. A new report from the Goldman Sachs Global Institute, authored by Co-Head George Lee and Managing Director Dan Keyserling, covers what is known in the industry as the “world model”—and argues that solving it represents the next decisive leap in artificial intelligence. Not a marginal improvement. A qualitative shift in what machines can do, and how consequentially they can do it.
The fact that the AI godfathers are already racing toward it suggests Goldman may be onto something.
The Gap Nobody Likes to Talk About
The large language model revolution produced something genuinely astonishing. Train a system on enough human text, optimize it to predict what word comes next, scale it up, and—almost inexplicably—it begins to reason, converse, write, and code at a level that routinely surprises its own creators. The commercial results have followed: trillion-dollar valuations, reshaped industries, a generation of white-collar workers rethinking their careers.
But beneath that capability sits a structural limitation the industry has been reluctant to confront head-on. “LLMs are powerful at completing patterns,” Lee and Keyserling write, “but they lack the internal sense of the world those patterns describe.” These systems, the Goldman authors note, “generate this understanding through second-order interpretation—they understand how our world works based on the data and text to which they have been exposed. They do not possess first-principles understanding of physics, motion, light, action/reaction, or other fundamental properties of our universe.”
Put plainly: today’s AI learned about the world by reading what humans wrote about it. It absorbed the description of reality without ever encountering reality itself. It can explain, in fluent prose, that a glass will shatter if dropped. It has no internal sense of the weight, the trajectory, or the consequence.
That distinction barely registers in the use cases dominating enterprise AI today—summarizing documents, drafting communications, generating code. It becomes a hard wall the moment AI is asked to navigate an unstructured physical environment, coordinate a complex organizational response in real time, or reason about how a strategic decision will cascade through a live market.
What the Godfathers Are Building
Here is where the Goldman report becomes more than a think piece. The researchers converging on world models aren’t a fringe movement. They are, in several cases, the same people whose earlier work produced the AI era now dominating headlines.
Yann LeCun, who spent years as Meta’s Chief AI Scientist before departing to launch his new venture AMI Labs, has made world models the explicit foundation of his vision for artificial general intelligence. His Joint-Embedding Predictive Architecture—JEPA—is designed to build machines that develop internal models of the world through observation, the way humans do, rather than through text prediction. LeCun has been publicly and persistently critical of the idea that scaling LLMs alone will reach general intelligence. World models are his alternative thesis.
Fei-Fei Li, the Stanford researcher whose ImageNet dataset helped ignite the deep learning revolution that produced today’s dominant AI systems, founded World Labs around a related idea: spatial intelligence. The premise is that genuine intelligence requires not just recognizing objects in images but understanding how those objects exist in space, interact with each other, and change over time. Li’s bet is that machines need to inhabit a model of three-dimensional reality, not merely classify it.
These are not peripheral figures staking out contrarian positions for attention. They are the architects of the current paradigm, arguing in their own research and ventures that the paradigm is incomplete.
Two Frontiers, One Idea
The Goldman report maps out what world models actually look like in practice—and identifies two distinct but related tracks.
Physical world models teach AI the governing logic of the material world: gravity, friction, thermodynamics, fluid dynamics. Rather than learning purely from real-world trial and error, these systems absorb the rules of physics through simulation, practicing in digital environments where failure is cheap and fast. A robot can fall thousands of times inside a simulator before ever touching a floor. When it finally acts in physical space, it does so having already internalized consequence.
The results are already visible in logistics, manufacturing, and autonomous systems—warehouse robots navigating crowded spaces with fewer collisions, autonomous vehicles rehearsing edge cases before encountering them on the road. The critical advance, as Goldman frames it, isn’t better hardware. It’s better internal models of reality.
Virtual, or social, world models pursue a parallel ambition in human systems. These are digital environments populated by AI agents with goals, memories, and incentives—each one designed to approximate a real-world behavioral profile. As those agents interact, patterns emerge. Markets behave. Organizations respond. Crises cascade. “Enterprises already spend enormous effort guessing how others will respond, how competitors will move, how markets will interpret signals, how boards will react under pressure,” Lee and Keyserling write. “Multi-agent simulations offer something closer to a living model of human systems.”
The Goldman authors draw a distinction here that matters enormously for how business leaders should think about these tools: world models are not forecasts. “These systems don’t predict the future in any narrow sense; they’re meant to reveal plausible futures and expose hidden dynamics,” they write. “Forecasting assumes a single correct outcome. World models reveal ranges, paths, and feedback loops.”
The Investment Question Wall Street Hasn’t Asked
Goldman being Goldman, the report ultimately lands on a financial argument—and it’s a pointed one.
The entire AI infrastructure buildout, the report notes, has been sized around a single assumption: that the future of AI is larger language models running on more compute. Current projections for chips, data centers, and energy capacity are built almost entirely on that foundation. Goldman’s question is whether those projections are measuring the right thing.
“The demands and opportunities surrounding world models are not yet reflected in consensus supply-and-demand forecasts for AI infrastructure,” Lee and Keyserling write. If world models develop as a complementary layer—built alongside LLMs rather than replacing them—the compute requirements could substantially exceed what current Wall Street forecasts anticipate. Simulation environments require purpose-built data pipelines, synthetic data generators, and physics-based engines that go well beyond text corpora. “The infrastructure story,” the authors write, “is one of partial overlap, not seamless reuse.”
The competitive framing is equally stark. “Competitive advantage might depend as much on who trains the largest model as who builds the most faithful simulations of reality, physical, social, and economic.”
The Missing Link
The Goldman report closes with a formulation that doubles as the clearest summary of what world models represent—and why the race to build them is drawing the field’s most credentialed minds.
“If large language models give AI fluency, world models give it situational awareness,” Lee and Keyserling write. “For much of its recent history, we’ve treated artificial intelligence as a system that produces answers. World models suggest something more ambitious.”
The AI that has reshaped the past decade learned to talk about the world with remarkable sophistication. The AI the godfathers are now building is trying to learn something harder, and more fundamental: what it actually feels like to be inside it.
For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing.
