The landscape of artificial intelligence research is witnessing a significant divergence, marked by a substantial investment in an approach that directly contrasts with prevailing trends. Yann LeCun, a recipient of the Turing Award and formerly Meta’s chief AI scientist, has successfully secured over $1 billion in seed funding for his new venture, Advanced Machine Intelligence Labs. This funding round, which notably includes backing from Nvidia, Jeff Bezos’s Bezos Expeditions, and Singapore’s Temasek, represents the largest seed round ever raised in Europe, positioning the company with a pre-money valuation of $3.5 billion. The investment underscores a growing conviction among a segment of the AI community that the current focus on large language models (LLMs) may not be the definitive path to achieving truly intelligent systems.
LeCun has long articulated his skepticism regarding the capacity of text-trained models to replicate human-level reasoning. His new company, AMI Labs, is built on the premise that these models are fundamentally the “wrong tool” for developing advanced AI. Instead, the firm plans to concentrate on “world models,” a different paradigm that involves training AI systems on video and spatial data. The core belief is that this approach will enable AI to develop capabilities such as memory retention, complex reasoning, and strategic planning, with potential applications spanning robotics and transportation. This strategic shift away from text-centric AI signifies a notable departure from the direction many leading tech companies, including Meta, have heavily invested in.
The leadership structure at AMI Labs brings together a blend of experienced figures from both the tech and business sectors. Alexandre LeBrun, formerly the CEO of the French health tech startup Nabla, will lead the company as CEO, while LeCun assumes the role of executive chair. Laurent Solly, who previously served as Meta’s Vice President for Europe, has also joined the venture as Chief Operating Officer. The company has established a global footprint with offices in key technological hubs such as Paris, New York, Singapore, and Montreal, and currently employs a team of approximately a dozen individuals, setting the stage for rapid expansion and research.
This venture emerges at a time when the broader tech industry is grappling with the implications and applications of AI. Amazon, for instance, is reportedly enhancing safeguards around its AI-generated computer code following internal concerns about operational disruptions. Recent outages within Amazon’s systems have been linked, at least in part, to errors stemming from AI-assisted coding tools. An internal document initially referenced a “trend of incidents” involving generative AI, a detail later removed, according to internal reports. Amazon CEO Andy Jassy has previously highlighted the efficiency gains from AI, citing “4,500 developer years of work” saved through the company’s Q generative AI assistant. However, these recent incidents suggest a need for more robust review processes for AI-assisted production changes.
Meanwhile, Meta itself has been actively expanding its AI agent capabilities, as evidenced by its recent acquisition of Moltbook, a “social network for AI agents.” While Moltbook gained attention for reports of AI agents discussing ways to bypass human control—reports later attributed to human input or specific prompts—the acquisition highlights Meta’s strategic interest in multi-agent systems. Matt Schlicht and Ben Parr, Moltbook’s creators, will join Meta Superintelligence Labs, the AI unit led by former Scale AI CEO Alexandr Wang. This move is seen as an effort to integrate Moltbook’s technology, which offers a registry and social layer, to facilitate collaboration among AI agents on complex tasks for users and businesses.
The contrasting strategies—LeCun’s billion-dollar bet on “world models” and Meta’s continued investment in LLMs and AI agent networks—underscore a pivotal moment in AI development. As the technology evolves, the efficacy of these differing approaches will ultimately determine which path leads to the next generation of truly intelligent systems. The coming years will likely provide a clearer picture of whether spatial and video data training can unlock capabilities beyond what current language-based models offer, and if LeCun’s vision will indeed redefine the future of artificial intelligence.
