Yann LeCun launches AMI Labs to build $1B world models, moving past LLMs

Yann LeCun Challenges the Limitations of Generative AI

On 2 July 2025, Turing Award winner and AI pioneer Yann LeCun gave a pointed interview to the BBC in which he declared that large language models (LLMs) such as ChatGPT, Claude, and Gemini are structurally incapable of achieving human-level or animal-level intelligence. LeCun, who recently departed Meta to found Advanced Machine Intelligence Labs (AMI Labs), argues that the generative text-prediction paradigm underpinning today’s most prominent AI products has hit a hard ceiling. His position is not a minor technical quibble. It represents a fundamental disagreement among AI’s founding architects about what intelligence actually requires and whether the current trajectory of the industry can deliver it.

AMI Labs has backed this position with capital at a scale that demands attention. The company completed a seed funding round of USD 1.03 billion (approximately AUD 1.58 billion at current exchange rates), achieving a valuation of USD 3.5 billion (approximately AUD 5.4 billion). Backers include Nvidia and Jeff Bezos, both of whom are directing capital away from the prevailing text-scaling paradigm and toward physical and spatial AI architectures. This is not a fringe research project. It is a well-capitalised institutional bet that the dominant AI stack is wrong in a fundamental way.

For professionals operating in physical industries, including environmental consulting, civil engineering, logistics, and field operations, the implications are material. The current wave of AI tools sold to these sectors is built on the same LLM foundations LeCun is criticising. Understanding what world models are, why they are architecturally different, and where investment is flowing matters for anyone making medium-term decisions about technology adoption, workflow automation, or data infrastructure.

Key details of the world model architecture and AMI Labs technology

LeCun’s central technical argument is that autoregressive LLMs are next-token predictors. They process language sequentially, generating each word based on a statistical prediction of what word is most likely to follow the previous one. This linear process creates a compounding error problem: small miscalculations at each step amplify exponentially across long reasoning chains. For tasks that require multi-step physical reasoning, such as predicting how an object moves through space or how a robotic arm should adjust grip pressure in response to sensory feedback, this error amplification makes LLMs structurally unreliable regardless of parameter count or training data volume.

AMI Labs is building its systems on an architecture LeCun developed called the Joint Embedding Predictive Architecture, or JEPA. Rather than predicting every word, pixel, or discrete token in sequence, JEPA learns abstract, high-level representations of the world and uses those representations to predict the consequences of actions in a latent space. The distinction is important. Instead of reconstructing all the granular detail of a scene or a sequence of events, JEPA reasons about what will happen at a conceptual level, which is closer to how biological intelligence operates. A rat navigating a maze does not reconstruct every photon hitting its retina. It maintains an internal model of space and updates that model as it gathers new sensory information.

To address the limitation of static world models that fail when real-world conditions shift unexpectedly, AMI Labs recently introduced a refinement called AdaJEPA. This architecture embeds sensorimotor adaptation directly into a Model Predictive Control (MPC) loop. In practice, this means the AI system continuously updates its internal physical model based on live sensory feedback rather than relying on a fixed representation learned during training. If a robot encounters a surface with unexpected friction, or if a sensor returns anomalous data, AdaJEPA can revise its model of the physical environment in real time and adjust its predicted action sequences accordingly. This closed-loop adaptation is absent in standard LLM-based agents, which have no mechanism for updating their world understanding mid-task without an explicit external intervention.

Stanford professor Fei-Fei Li’s startup, World Labs, is pursuing a parallel direction, having secured approximately USD 1 billion (approximately AUD 1.54 billion) to develop 3D world models grounded in video and spatial data. The convergence of two separately capitalised research programmes, built by two of the most credentialled figures in AI research, on the same architectural thesis suggests this is not a contrarian position. It reflects a genuine and growing schism in the field between those who believe intelligence emerges from scaling text data and those who argue that physical embodiment and sensory grounding are prerequisites for reliable general intelligence.

Australian context: what this shift means for technology adoption in physical industries

Australia’s professional services economy, including environmental consulting, engineering, resources, and infrastructure, is in the middle of an AI adoption cycle that is largely built on LLM-based tools. Generative AI platforms marketed to environmental and engineering professionals are predominantly text-prediction systems. They can draft reports, summarise regulatory documents, and respond to natural language queries about legislative frameworks. What they cannot reliably do is reason about physical systems, interpret multi-sensor field data streams in real time, or make spatially coherent predictions about how a contaminated plume will migrate through a heterogeneous aquifer. LeCun’s critique names exactly this limitation.

For Australian environmental professionals, the practical consequence is that the AI tools available today are well-suited to documentation tasks and poorly suited to the physical modelling and spatial reasoning that underpins much of the sector’s core technical work. The architectural shift underway at AMI Labs and World Labs points toward a generation of AI systems that may eventually address this gap — though those systems remain in early development and are not yet available as commercial products.

References and related sources

• Primary source: letsdatascience.com
• defencepk.com
• sepe.gr
• aiweekly.co
• siliconcanals.com

How iEnvi can help

iEnvi integrates technology and data-driven approaches into environmental consulting. We monitor AI and technology developments that affect how environmental professionals deliver services to clients.

• Our capabilities
• About iEnvi
• Talk to the team

This is an iEnvi Machete news summary. Prepared by iEnvi to summarise the source article for environmental professionals tracking AI, data, and technology developments that affect consulting and project delivery.

Published: 04 Jul 2026

Need advice on this topic? Speak to an iEnvi expert at info@ienvi.com.au or 1300 043 684, or contact us online.