When AI Learns How the World WorksAfter a decade defined by systems that recognize patterns and predict text, the frontier of AI is shifting toward models that understand how the world works.
Overview
What We Do Insights Our Firm Careers Investor Relations Investor Relations Investor Relations Investor Relations Pressroom Pressroom Pressroom Pressroom Worldwide Worldwide Worldwide Worldwide Alumni Alumni Alumni Alumni Client Login Client Login What We Do Insights Our Firm Careers Investor Relations Investor Relations Investor Relations Investor Relations Pressroom Pressroom Pressroom Pressroom Worldwide Worldwide Worldwide Worldwide Alumni Alumni Alumni Alumni Client Login Client Login What We Do Insights Our Firm Careers Investor Relations Investor Relations Investor Relations Investor Relations Pressroom Pressroom Pressroom Pressroom Worldwide Worldwide Worldwide Worldwide Alumni Alumni Alumni Alumni Client Login Client Login What We Do Insights Our Firm Careers Investor Relations Investor Relations Investor Relations Investor Relations Pressroom Pressroom Pressroom Pressroom Worldwide Worldwide Worldwide Worldwide Alumni Alumni Alumni Alumni Client Login Client Login
When AI Learns How the World Works
When AI Learns How the World WorksAfter a decade defined by systems that recognize patterns and predict text, the frontier of AI is shifting toward models that understand how the world works. The next advances in AI may come from world models, which represent a quiet but decisive change in how machines become intelligent. How AI World Models Create an Operating System for Decision-Making After a decade defined by systems that recognize patterns and predict text, the frontier of AI is shifting toward models that understand how the world works. The next advances in AI may come less from bigger models and more from systems that can simulate reality, test actions before taking them, and reason about consequences. This new category of models, known as world models, represents a quiet but decisive change in how machines become intelligent. For the past few years, artificial intelligence has been defined by large language models (LLM). Trained on vast swathes of text, they learned to predict the next word with uncanny accuracy. From that simple objective emerged systems that write, translate, code, and converse with startling fluency. That achievement is real and transformative, but it also reveals a limitation to the current generation of AI models. LLMs are powerful at completing patterns, but they lack the internal sense of the world those patterns describe. They respond well to prompts but struggle to reason through consequences or act reliably in environments where mistakes carry costs. This limitation has become clearer as these systems have been pushed beyond text. When they’re asked to control robots, manage entire supply chains, or coordinate complex enterprise workflows, prediction alone proves insufficient. Intelligence, in these settings, requires more than correlation. It requires an internal model of how the world works. Consider the layers in practice: An LLM can extract covenants from a stack of loan documents or draft an investment committee memo. A physical world model can simulate how a hurricane season reshapes insured-loss distributions across a reinsurance portfolio. A social world model can forecast how a policy shock cascades through markets and behavior. The most consequential decisions may eventually draw on all three capabilities—yet plenty of high-value financial tasks remain squarely in LLM territory today. What’s changing is that building these natural evolutions of LLMs is no longer a fringe ambition. It has become a strategic priority for some of AI’s most influential researchers. Yann LeCun, who recently left his position as Chief AI Scientist at Meta, has made world models the centerpiece of his vision for artificial general intelligence and his new venture, AMI Labs. His Joint-Embedding Predictive Architecture (JEPA) framework explicitly aims to build machines that learn world models from observation, much as humans do, focusing on predicting abstract representations or concepts about what comes next without reconstructing exact details. Meanwhile, Fei-Fei Li, the Stanford professor whose ImageNet dataset helped catalyze the deep learning revolution, has founded a new venture focused on spatial intelligence. Her work at World Labs emphasizes that true intelligence requires not just recognizing objects in images but understanding how those objects exist in space, how they interact, and how they change over time. In other words, instead of asking models to respond to inputs, researchers create internal representations of the world so they can run simulations inside them. These so-called world models allow systems to imagine outcomes before taking an action. They run mental experiments. They test possibilities. You could call it a primitive form of machine foresight. But the term world model hides an important distinction. There are two kinds of worlds AI is learning to model. One is the physical world of gravity, friction, heat, and force. The other is a virtual or social world, populated by many inte
Subscribe to Briefings
Our signature newsletter with insights and analysis from across the firm By submitting this information, you agree that the information you are providing is subject to Goldman Sachs’ privacy policy and Terms of Use. You consent to receive our newsletter via email.
