en.Wedoany.com Reported - Yann LeCun announced at the RAISE Summit in Paris that his newly founded company, AMI, has raised $1 billion in seed funding to develop a "world model" capable of understanding the physical world like a child or an animal, rather than continuing to pursue larger language models.
The scientist, hailed as the godfather of artificial intelligence, built the convolutional neural networks in the late 1980s that are still used today in smartphone cameras, medical scans, and driver-assistance systems. He subsequently served as Meta's Chief AI Scientist for 12 years and informed Mark Zuckerberg in November that he would be leaving.
LeCun believes that large language models (LLMs), while useful, cannot lead technology toward true intelligence. They excel at processing sequences of symbols, passing exams, writing emails, and summarizing texts, but they do not understand the physical world. Evidence of this is that current AI still cannot achieve Level 5 autonomous vehicles, household robots, or accomplish what a 10-year-old child or even a house cat can do. In his view, the reason is that language models only perceive the world through text or images, missing most of the data on how the physical world operates.
LeCun's technical approach employs a method called JEPA. Instead of generating the next frame of a video pixel by pixel, the system learns an abstract representation of the scene and makes predictions within that abstract space, discarding unpredictable noise while retaining truly important structures. This is done because directly predicting the next frame of a video involves unpredictable details, leading the model to produce blurry averages.
LeCun points out that all text on the internet totals about 10 to the 14th power bytes, while a four-year-old child absorbs roughly the same amount of information through visual input alone. This video data densely encodes how the world works, allowing the child to learn gravity, motion, and object behavior without any labels. AMI's video model can already detect anomalous events appearing on screen, which is considered a form of common sense learned from observation.
Regarding his departure from Meta, LeCun stated that although Zuckerberg and Meta's technical leadership supported his research, the company devoted all its efforts in 2025 to catching up with competitors in the large language model space, believing that human-level AI could be achieved through scale alone—a direction he disagreed with. Additionally, the near-term applications of world models lie in the industrial sector, including controlling complex systems, factories, and engines, which misaligns with Meta's mission of connecting people.
LeCun also promoted an open foundation model initiative called Tapestry at the summit. The initiative employs a federated mechanism: each country, university, or company trains models using its own data and hardware, sending only the final parameters to a central server for averaging to form a shared model, with data never leaving its local source. He believes the results can surpass closed-source models because a vast amount of valuable private data can be utilized.
AMI is headquartered in Paris, with offices in New York, Montreal, and Singapore. LeCun's message to Europe is that the race is not yet lost; he believes Silicon Valley is stuck in a "trench" of developing language models, while world models represent a different path. The next true revolution in artificial intelligence will come from understanding the world, not just predicting the next word.
