LeCun's Paris AI startup

Avatar
Lisa Ernst · 06.12.2025 · Technology · 5 min

The current development in the field of Artificial Intelligence (AI) shows a shift away from the pure dominance of Large Language Models (LLMs). Two recent news items illustrate that the focus is increasingly on systems that can develop a deeper representation of the world and plan actions within that representation. This points to an upcoming wave of products that go beyond text-based approaches.

AI development

On December 4, 2025, Yann LeCun announced the founding of a globally oriented AI startup at the AI Pulse conference in Paris. LeCun, who is leaving Meta after around twelve years, having shaped it since 2013 FAIR and built it up as Chief AI Scientist, , will focus his new company on 'World Models'. These systems will learn an internal representation of the world and be able to plan actions within that representation.

Yann LeCun, a key figure in AI research, plans to found a new startup in Paris.

Source: winbuzzer.com

Yann LeCun, a key figure in AI research, plans to found a new startup in Paris.

The choice of Paris as a location and the 'World Models' approach are significant. LeCun stated that Meta would be a partner but not an investor, indicating a strategic decoupling. He is pursuing a research track that, in his assessment, goes beyond what Meta currently wants to cover. For years, LeCun has argued that simply scaling LLMs does not automatically lead to human-like intelligence, as fundamental abilities such as a stable understanding of the world, robust planning, and lasting memory structures are missing. His group has therefore JEPA-Ansätzen worked, on learning predictions in an abstract representation space, rather than reconstructing pixels or tokens. With V-JEPA this direction was concretely described for video in 2024, , i.e., for learning from dynamic scenes rather than text.

From a user perspective, a 'World Model' could enable a robot to internally simulate the risk of an action before executing it, for example, when learning a prone-to-error work step in a factory. The practical difference from many current agents is that the planning will then not just seem 'plausible' in a text-based way, but will be better grounded in physics and causality.

LeCun's new startup could redefine the boundaries of AI models and significantly shape the future of technology.

Source: mischadohler.com

LeCun's new startup could redefine the boundaries of AI models and significantly shape the future of technology.

The anchoring of this project in Paris also carries an industrial policy message. President Macron had publicly signaled his desire to attract LeCun to a location in France. For European startups and research institutions, this could be an opportunity to participate in a leading agenda that not only targets chatbots but also perception, action, and robotics.

AI applications

While LeCun is conceptually driving forward the 'next generation of artificial intelligence research,' the weather side of AI is already showing concrete effects. In the 2025 hurricane season, AI models were included and evaluated in real forecast processes. According to ABC News, NOAA reports that one DeepMind-Modell bei Track- und Intensitätsprognosen das stärkste Nicht-Official-Modell war was only surpassed by the official NHC forecasts overall.

These results are the outcome of a formal collaboration between NOAA and Google, announced in July 2025, to rapidly and scientifically evaluate AI models at the National Hurricane Center. In June 2025, DeepMind and Google Research also launched 'Weather Lab,' a platform that makes experimental AI models for tropical cyclones openly accessible and displays ensembles with up to 50 scenarios. The company explicitly emphasizes that these forecasts are for research purposes and do not replace official warnings.

Technically, the comparison with classical models is interesting. DeepMind reports that its experimental cyclone model surpassed the average intensity errors of NOAA's HAFS, a high-resolution physics-based regional model, in internal evaluations. Nature highlighted in the fall of 2025 that the training base consists of large weather observation data and specialized cyclone datasets – an example of data-driven learning that directly impacts critical infrastructure.

For disaster management, the benefit is easy to translate. If a model indicates the probability of rapid intensification earlier and more reliably, authorities can plan evacuations more precisely, staff resources, and prepare protective infrastructure. This is not an abstract 'AI for Good' narrative, but a tool that must fit into the decision-making rhythm of operations centers and is apparently proving its worth there.

The fact that weather AI is leading the way here is also an indication of a broader direction. Atmospheric dynamics, sensor technology, satellite data, and ensemble thinking offer a natural bridge between pure pattern recognition and the 'World Models instead of LLM Trend' research approach, as the models must not only describe but also probabilistically anticipate future states of the world.

Future of AI

The announcement of Yann LeCun's new AI startup in Paris and the results of the 2025 AI hurricane forecasts may seem separate at first glance. However, viewed together, they tell a coherent development story: AI is shifting from pure language competence towards systems that can model world states, assess risks, and better justify actions.

LeCun provides the theoretical and entrepreneurial banner for this, while DeepMind and NOAA offer a tangible example of how such approaches are already providing benefits in safety-relevant domains today. When these lines converge, the real debate of the coming years will likely be less about which LLM is the most eloquent, but which system understands the world most reliably and is used responsibly enough to improve real decisions.

Further reports

For more information on the topics mentioned, the following sources can be consulted:

Share our post!
You might also like