After ChatGPT, a new AI breakthrough claims it can truly “understand” our world – Futura-Sciences

Large language models like ChatGPT have gotten really good at summarizing documents, translating languages, and writing code. Yet a growing number of AI researchers argue that text-based fluency has a hard ceiling. Because these systems learn only from words on a page, they lack any real sense of how physical reality works. Now, a new wave of research is betting on “world models”, systems designed to understand space, movement,and cause and effect.

Two well funded bets on physical understanding

The push toward world models has drawn some of the biggest names in computer science, along with billions of dollars in new capital:

• World Labs: Stanford computer scientist Fei-Fei Li, known for creating the ImageNet dataset that helped launch modern deep learning, co-founded World Labs in 2024 alongside Justin Johnson, Christoph Lassner, and Ben Mildenhall.

  The company’s core platform, Marble, generates fully explorable 3D environments from text, image, or video prompts. Following a limited preview in late 2025, World Labs launched Marble for general availability in February 2026. Shortly after, the company raised $1 billion in new funding, bringing its total capital raised to $1.23 billion.

• AMI Labs: Yann LeCun, who shared the 2018 Turing Award for foundational work on deep neural networks, took a different path to the same problem. After 12 years leading Meta’s Fundamental AI Research lab, LeCun left the company in late 2025 to launch AMI Labs. Based in Paris, the venture raised $1.03 billion in seed funding in March 2026 at a $3.5 billion valuation.

  Led by CEO Alexandre LeBrun, who previously headed health-tech startup Nabla before stepping into a chairman role there, the company features LeCun as executive chairman while he continues teaching at New York University.

Why physical reasoning differs from language

A chatbot can easily describe how to pour a glass of water. A robot attempting the actual task, however, needs more than words, it must calculate finger placement, weight distribution, and grip strength, or risk shattering the glass.

This contrast illustrates the core premise behind world models. Understanding derived from reading text is fundamentally different from understanding grounded in physical interaction:

• New training inputs: while traditional models scrape billions of sentences from the web, world models train on video streams, 3D spatial data, and sensor feeds.
• Predictive architecture: AMI Labs builds on Joint-Embedding Predictive Architecture (JEPA), a technique LeCun championed at Meta. Instead of predicting the next word in a sentence, JEPA trains systems to predict missing spatial details within video and images.
• Practical applications: fields like robotics, autonomous driving, scientific simulation, and industrial design all require software that understands physical cause and effect, areas where current language models regularly fail.

LeCun has argued that humans and animals learn primarily by observing and interacting with their surroundings, not by reading. In his view, AI trained solely on internet text will inevitably hit a wall that simply adding more text cannot fix.

Questions around safety and deployment

As world models gain the ability to plan and act rather than just answer prompts, some researchers have raised concerns regarding oversight. A system that predicts how the physical environment will react, and then takes autonomous action based on those predictions, presents more complex monitoring challenges than a standard text bot.

Both leading startups are currently testing early commercial applications. AMI Labs has partnered with Nabla to develop healthcare software targeted for FDA certification, alongside planned tools for robotics and industrial automation. World Labs has focused its initial efforts on creative and spatial design tools, backed by investors including AMD’s venture arm and Salesforce CEO Marc Benioff.

Neither company has proven that its technology can operate flawlessly outside controlled testing environments. However, consensus within the field is shifting: the future of artificial intelligence will likely depend on combining language fluency with a grounded, practical understanding of the physical world.