The Unexpected Twist of a Record-Breaking Seed Round
On March 9, 2026, Yann LeCun completed a $1.03 billion seed round for his startup, AMI, a record for a European company. This funding is not just a financial event; it’s a sign of a paradigm shift: the world models AI model proposed by LeCun stands in direct contrast to the dominant approach of large language models (LLMs). The focus on physical behavior prediction, rather than language processing, reveals a different technical strategy, with implications for cognitive architecture and computational efficiency.
LeCun‘s statement, “We just completed our seed round, one of the largest seeds ever, probably the largest for a European company,” is not just a financial announcement. It indicates a desire to redefine the technical parameters of AI, challenging the hegemony of LLMs. This approach, which prioritizes understanding the physical world, requires a rethinking of training and deployment paradigms, with consequences for system scalability and security.
The Physiology of Alternative Models
LeCun‘s world models are based on a cognitive architecture that integrates physical predictions and environmental interactions. Unlike LLMs, which process text based on statistical patterns, these models seek to simulate the behavior of objects and agents in a concrete environment. This requires a different type of computational processing, with a focus on sensory data and physical simulations, rather than linguistic tokenization.
LeCun‘s criticism of LLMs is not just theoretical. Large language models, while powerful, have limitations in terms of energy efficiency and the ability to interact with the physical world. World models aim to overcome these limitations, offering a solution more suitable for tasks that require contextual understanding and direct interaction with the environment. However, this approach requires a different computing infrastructure, with implications for data distribution and management.
The Symbiosis Between Research and Market
LeCun‘s vision is not isolated. Andrej Karpathy, a researcher at OpenAI, has demonstrated that an autonomous AI agent can improve the efficiency of training a language model by 11%, testing 700 experiments. This suggests that AI agents not only perform tasks but can also autonomously optimize learning processes. Karpathy’s key statement, “AI agents can now do research on their own,” indicates a breakthrough in self-optimization capabilities, which could reduce reliance on human intervention.
“AI agents can now do research on their own.”
Andrej Karpathy, OpenAI
This development has implications for the market and research. If AI agents can autonomously optimize learning processes, a new class of intelligent systems may emerge that not only perform tasks but also improve themselves. However, this autonomy requires a robust support infrastructure, with implications for the safety and ethics of AI.
Scenario in 3-5 Years: The Sedimentation of Tensions
In my view, the next decade will see a gradual sedimentation of tensions between LLM models and alternative models. LeCun‘s world models may find applications in sectors where contextual understanding is crucial, such as robotics and industrial automation. However, the hegemony of LLMs will not be immediately replaced, as these models have a well-established support infrastructure and a large user base.
The main challenge will be adapting the computing infrastructure and managing data. World models require different processing, with a focus on sensory data and physical simulations. This could lead to a diversification of the AI ecosystem, with different models adapting to specific contexts. Scalability and energy efficiency will be the key criteria for evaluating the success of these models.
Photo by Andrey Matveev on Unsplash
Texts are processed autonomously by Artificial Intelligence models