LLMs are not the end of AI! Yang Likun founded AMI, securing $1 billion in funding and betting on the World Model approach.

Winner of the Turing Award and former Meta Chief AI Scientist Yann LeCun (Yong Li Kun) founded the artificial intelligence research company Advanced Machine Intelligence (AMI), which recently completed a large seed round of approximately $1.03 billion (about €890 million). The company’s valuation is around $3.5 billion, reaching unicorn status in less than three months. This indicates that the capital market is betting on a different path from the large language models (LLMs) dominated by companies like OpenAI and Anthropic.

In recent years, two heavyweight figures in AI academia, Yann LeCun and Fei-Fei Li, known as the “AI Godmother,” have both chosen to focus on a new generation AI architecture called World Model. Currently, mainstream LLMs rely on vast corpora and statistical relationships to understand context and predict the next word. They can recognize that “apple falls” often appear together but do not truly understand gravity or causal relationships in the physical world.

This pattern performs well in text generation, coding assistance, or question-answering tasks, but has fundamental limitations in scenarios requiring understanding of real-world structures, reasoning about causality, and long-term planning. A bigger issue is that data sources are gradually depleting. Training LLMs heavily depends on high-quality human data, but industry warnings suggest that available human text data may be exhausted within a few years. When that happens, it could be like inbreeding, leading to genetic defects, ultimately causing models to drift away from reality and degrade in performance.

Yann LeCun’s AMI completes $1.03 billion seed funding, reaching unicorn valuation in three months

Founded by Turing Award winner and former Meta AI chief Yann LeCun, Advanced Machine Intelligence (AMI) recently completed a large seed round of about $1.03 billion (approximately €890 million). After this round, the company’s valuation is about $3.5 billion, and it has achieved unicorn status in less than three months, making it one of the largest seed funding rounds in European history, according to multiple media outlets.

The goal is to develop the next-generation AI systems, called World Models, enabling AI to understand and reason about the real world, rather than merely predicting the next token.

This funding round was led by multiple venture capital firms, including Cathay Innovation, Greycroft, Hiro Capital, and HV Capital, with participation from major tech and industry giants. Strategic investors include Nvidia, Toyota, Samsung, and others. Additionally, Jeff Bezos participated through his family investment firm, Bezos Expeditions.

It is noteworthy that such a large seed round in AI startup history is rare, indicating that the capital market is willing to pay a premium for AI research that diverges from the LLM route.

AMI’s focus on World Models reveals flaws in LLMs

AMI’s core research direction is world models, a type of architecture that enables AI to learn about physical world structures and causal relationships. The difference between World Models and existing LLMs is quite clear. Large Language Models rely on massive datasets to understand context and generate output based on statistical relationships. They know that “apple falls” often appears together but do not understand gravity.

More concerning is that LLMs depend on feeding data to train models. High-quality human data is running out. The industry is now relying on synthetic data, but if AI continues to learn from AI-generated content, it could lead to the “Habsburg effect.” Repeatedly training generative models on their own outputs is akin to inbreeding, which can cause genetic defects. Even under ideal conditions, models will gradually forget reality, leading to degradation.

(Cambridge research: By 2026, human data will be exhausted for LLM training, and AI may collapse due to the “Habsburg phenomenon.”)

AMI is developing systems that create abstract representations of the world, ignoring unpredictable details, and performing prediction and reasoning within this representation space. They aim to develop new AI systems capable of understanding the world, possessing persistent memory, reasoning, and planning, with controllability and safety. Their research continues the JEPA (Joint Embedding Predictive Architecture) approach proposed by LeCun during his time at Meta FAIR.

How AMI differs from World Labs, founded by another academic entrepreneur

Yann LeCun’s AMI Labs shares similarities with another AI startup founded by Fei-Fei Li, known as the “AI Godmother.” In 2024, Li founded AI startup World Labs and serves as CEO. The company positions itself as a frontier foundational model development and product company, with the core goal of creating large-scale World Models (LWM). These models enable AI not just to speak or interpret images but to understand, generate, and reason within 3D worlds and interact accordingly.

(Who is Fei-Fei Li? New unicorn World Labs receives support from Nvidia, AMD)

The goal is to enable models to understand 3D spatial structures and object relationships from images or videos, generate freely movable viewpoints of 3D scenes from text or images, and perform reasoning and simulation in virtual spaces, laying the foundation for robotic and physical AI behaviors.

Yann LeCun (YAM LeCun) and Fei-Fei Li (Li Fei-Fei) focus on JEPA, general world models, spatial intelligence, and 3D world models. Currently, their research is at the stage of product development, with no mature products yet, but they have released a 3D generation tool called Marble.

In summary, the World Model approach advocates that robots will be a key carrier for AI development.

This article states that LLMs are not the end of AI! Yann LeCun’s AMI secures $1 billion in funding, betting on the World Model route. Originally published by ABMedia on Chain News.