Pioneers like Yann LeCun and Oxford's Ingmar Posner are building next-gen AI models to tackle the physical world and power advanced robotics, moving past the limits of current LLMs.
The artificial intelligence landscape is witnessing a profound re-evaluation, as pioneering figures like Yann LeCun, formerly Meta's chief AI scientist, declare current large language models "not smart" and fundamentally limited for real-world applications. This bold assertion is not just academic; it is backed by significant capital, with LeCun's new venture, Advanced Machine Intelligence Labs (AMI Labs), recently securing over $1 billion in seed funding, signaling a crucial pivot for investors and the broader tech ecosystem towards next-generation AI architectures.
AMI Labs' monumental seed round, one of Europe's largest for an early-stage startup, attracted heavyweights including US chip giant Nvidia and the private wealth fund of Amazon founder Jeff Bezos. This capital infusion underscores a growing consensus among deep-tech investors that the path to truly intelligent, adaptable AI systems lies beyond the scaling of existing generative models like ChatGPT or Claude. The funding validates a strategic shift towards models capable of nuanced interaction with the physical world, a domain where current LLMs demonstrably falter.
LeCun, who departed Meta in 2025 to found AMI Labs, is developing a novel system dubbed Joint Embedding Predictive Architecture (JEPA). This architecture is designed to create abstract representations of reality, enabling AI to reason about outcomes and filter out extraneous information, a critical capability missing from statistical pattern-matching LLMs. The move reflects a belief that future AI advancements will be driven by systems with an inherent understanding of physics and causality, rather than solely linguistic fluency.
The Stakes
The core of LeCun's argument, and the driving force behind this investment surge, is the inherent limitation of LLMs when confronted with the unpredictable complexities of the physical world. While LLMs excel at well-defined tasks like coding, mathematical problem-solving, and text generation, they essentially "regurgitate" patterns from vast datasets without true underlying understanding. This architectural constraint renders them largely unsuitable for dynamic environments, such as those encountered by humanoid robots.
My read is that this isn't merely an academic debate; it's a strategic inflection point for the entire AI investment thesis. For years, the narrative has centered on scaling up transformer models and expanding training data, leading to a substantial market capitalization built on generative AI's prowess in digital domains. However, LeCun's move, backed by over a billion dollars from astute investors, signals a bifurcation: while LLMs continue to refine their digital capabilities, a new frontier is opening for foundational research into embodied AI and world models. This directly impacts where venture capital flows, favoring startups addressing physical intelligence over purely digital applications.
The robotics industry, having poured billions into advanced hardware, stands to benefit immensely from this paradigm shift. Humanoid robots are becoming increasingly sophisticated in their mechanical feats, yet their ability to perform common household chores or navigate unstructured industrial settings remains constrained by the limitations of their AI brains. LeCun explicitly states, "LLMs are largely hopeless for robotics," a sentiment echoed by many researchers who envision a future where robots operate autonomously and intelligently.
AMI Labs raised over $1 billion in seed funding, one of the largest seed rounds in European history, underscoring investor confidence in new AI architectures beyond current large language models.
Background on the Shift
The concept of "World Models" is not entirely new, having been explored conceptually for decades. However, recent advances in machine learning and computational power have reignited interest, transforming theoretical constructs into viable research avenues. Ingmar Posner, Professor of Applied Artificial Intelligence at Oxford University and an Amazon Scholar, leads a team developing a "mechanistic world model" that structures knowledge for efficient recall and modification, aligning with LeCun's vision.
This renewed focus on world models gained significant momentum following an influential 2018 paper by David Ha and Jurgen Schmidhuber, which demonstrated how AI could learn through a "mental simulation" of the world. Since then, major players have entered the fray, including Google's Dreamer World Model, which successfully navigated complex tasks in Minecraft by imagining future scenarios, and DeepMind's Genie model. London-based Wayve is also developing its Gaia system, and AI pioneer Fei-Fei Li founded World Labs in San Francisco in 2023 with similar objectives. This concerted effort across leading research institutions and well-funded startups underscores a critical industry-wide recognition that the current architectural approach to AI has limitations for certain, highly valuable applications.
What strikes me here is the sheer velocity of this pivot. Just a few years ago, the idea of generative AI reaching its current capabilities seemed decades away to many. Now, we are seeing a similar acceleration in the development of these next-generation models. This rapid evolution is not just about incremental improvements; it represents a foundational re-thinking of intelligence itself, moving from pattern recognition to predictive modeling and causal inference. For startups, this creates immense opportunities for those building foundational layers or applications leveraging these new model types, but also poses a challenge for those heavily invested in the pure LLM stack.
What It Means for the Ecosystem
This shift to world models and embodied AI has significant ramifications for the broader tech ecosystem, extending beyond venture capital. The demand for specialized talent capable of working with these complex architectures will intensify, potentially drawing researchers away from traditional LLM development. Furthermore, these new models, with their emphasis on physical understanding and simulation, could drive innovation in AI hardware, potentially moving beyond general-purpose GPUs to more specialized processors optimized for predictive encoding and spatial reasoning.
LeCun anticipates AMI Labs will spend the remainder of this year refining their JEPA model, with initial industrial deployments targeted for next year. If successful, the ambition is to develop "general generic intelligence systems that can be applied to just about anything in the world with minimal training or fine tuning." This long-term vision suggests a future where AI systems, even if more intelligent than humans in specific domains, will serve as sophisticated assistants. "Our interaction with future AI systems — even if they are smarter than us — is going to be like the interaction between a captain of industry or a political leader with their staff of assistants — many of whom are smarter than they are," LeCun observed, framing AI as an augmentation, not a replacement, for human ingenuity in defining goals and creativity.
The coming years will be crucial for validating these new AI paradigms. Key triggers to watch include the successful deployment of AMI Labs' JEPA in industrial settings in 2025, further funding rounds for World Model-focused startups, and the performance of Google's, DeepMind's, and Wayve's alternative architectures in real-world benchmarks. How established tech giants, currently heavily invested in LLM infrastructure, adapt their strategies to incorporate or compete with these emerging architectures will define the next phase of the global AI race.
Frequently asked questions
Why does Yann LeCun say current AI is 'not smart'?
LeCun believes current large language models (LLMs) like ChatGPT are limited because they cannot understand or interact with real-world data and physical environments effectively, unlike even animals or toddlers. They excel at predictable problems but lack true reasoning about reality.
What is AMI Labs developing to advance AI?
AMI Labs, founded by Yann LeCun, is developing a new AI system called Joint Embedding Predictive Architecture (JEPA). This system aims to create abstractions of the real world to better assess action outcomes and overcome the limitations of LLMs in physical environments.
What are 'World Models' in the context of new AI development?
World Models are a category of AI systems designed to learn and simulate how the world works, allowing the AI to understand causes, effects, and predict outcomes based on a 'mental' simulation. This enables more flexible decision-making, especially for robotics.
How will new AI models like JEPA and World Models benefit robotics?
These new AI models are crucial for robotics because they enable machines to understand and navigate the unpredictable real world. This will allow humanoid robots to safely perform complex tasks like household chores, which current LLMs are largely incapable of.
Who else is working on next-generation AI beyond LLMs?
Besides Yann LeCun, other key figures include Ingmar Posner at Oxford University, who is developing a 'mechanistic world model,' David Ha and Jurgen Schmidhuber (influential paper authors), and Fei-Fei Li, founder of World Labs.
What is the future vision for these advanced AI systems?
Developers hope to refine these new AI models for industrial use next year, with the eventual goal of creating general, generic intelligence systems applicable to almost anything with minimal training. These AIs are envisioned as powerful tools working for humans.
