World Models and Vertical Models: A New Frontier of AI Innovation for Europe
- Marc Griffith
- Mar 1
- 3 min read
Summary Analysis of the emergence of world models as a new axis of AI innovation in Europe: LeCun, Ami Labs, World Labs, and European startups are driving a shift from imitation to abstraction. Data, investments, and real-world case studies highlight where to invest, how to compete, and the scenarios shaping the industry's future. Key takeaways
|
The WAICF in Cannes sparked a lively debate on the evolution of AI, highlighting world models as a potential structural alternative to LLMs. World models and vertical models represent a tangible breakthrough for AI, with potential that goes beyond LLMs. This vision is fueled by European investments and talent aiming to transform the ability to learn from the real world into reliable, controllable systems, a central theme for startups and policy.
There is a large gap between the learning capabilities we observe in humans and animals and the kind of efficiency we are able to reproduce in machines.
Context and Key Players
Yann LeCun, the former AI chief at Meta, outlines a vision in which persistent memory, the ability to plan complex sequences of actions, and robust, safe control are essential elements for world models. The key is to build systems that master memory, planning, and safety. In recent months, interest has shifted from guessing dates to pursuing more solid structures that allow robots to operate coherently in the real world.
LeCun also notes that, although LLMs are incredibly efficient in some tasks, they still lack essential elements for true physical and perceptual understanding of the world. There is a large gap between the learning abilities we observe in humans and animals and the kind of efficiency we reproduce in machines.
Fei-Fei Li and her World Labs have advanced a similar path, focusing on 3D worlds and abstract representations to simulate real contexts and scenarios. World models as the next frontier for AI systems capable of understanding and simulating the world. Ami Labs, LeCun's creation, is developing architectures that learn abstract representations and forecast the future in a space of representations, handling data from real and dynamic sensors.
Generative architectures trained through self-supervised learning have succeeded in understanding and generating language, but most real-world data remain unpredictable and hard to model with generative approaches.
Outlook and Concrete Examples
In parallel, Fei-Fei Li and Fei Li have inspired investments and European startups in the sector, including initiatives aimed at creating world models capable of learning robust and abstract representations. The pursuit of abstract representations and predictions in the representation space is the direction indicated by the experts. The focus is not only on language models, but on how to integrate perception, action and reasoning in real-world contexts, such as robotics and industrial automation.
Several players are exploring hybrid formats and hybrid approaches: specialized language models, with vertical data and self-supervised training, paired with real sensors, to create more reliable systems. Advanced generative architectures are not enough if they are not integrated with robust sensory data and clear use contexts. In Europe, companies like DeepL illustrate how models can evolve into contextualized services, able to maintain security and privacy while offering competitive performance.
European competitiveness depends on the combination of research, business applications and infrastructure, not on passive adoption of large models.
Practical Implications for Startups and Policymakers
The shift from large generalist models to vertical solutions can translate into reduced costs, greater data control, and less reliance on massive infrastructure. Ami Labs is exploring how world models learn abstract representations of real-world sensory data, ignoring unpredictable details to formulate useful predictions in the representation space. Opportunities extend beyond industries traditionally associated with AI, including sectors requiring high reliability and compliance with sensitive data.
From an investor perspective, the landscape suggests a growth trajectory based on specialized models, with potential use cases in robotics, industrial automation, and automated business processes. Moreover, European leadership could stem from a mix of advanced research, real B2B applications, and policies that incentivize responsible AI adoption. European competitiveness depends on the integration of research, business and governance.
Conclusions: A Shared Trajectory
The path to world models requires cooperation among universities, startups, established companies, and public policy to build an ecosystem capable of turning theoretical discoveries into practical applications. The synergy of creativity, data, and adequate infrastructure will be decisive in building reliable and globally competitive AI. In this scenario, Europe has the opportunity to lead a transformation that goes beyond simply imitating existing models, steering innovation toward highly specialized and robust AI systems.
