Beyond Language: What Yann LeCun’s World Models Mean for the Next AI Revolution (Augmented with GPT 6 Luna)
From sensory signals to national competitiveness, industrial transformation and the future of systems design.
By Leke | Founder & CEO, Wonda Designs
On 29 October 2026, at Dartmouth’s AI at 70 conference, Yann LeCun delivered a keynote titled World Models: Enabling the Next AI Revolution. The occasion marked seventy years since the 1956 Dartmouth Summer Research Project that helped establish artificial intelligence as a field of study. His keynote invited a timely question: what comes after our current generation of increasingly capable language models?
The question is larger than the next generation of AI products. It concerns how machines learn, how economies create value, how countries develop technological capabilities, and how we design the complex systems through which innovation becomes practical, reliable and beneficial.
What particularly interests me is the idea of signals beyond large language models. It points towards a distinction that deserves more attention in public conversations about artificial intelligence: the difference between generating convincing descriptions of the world and developing useful internal models of how the world behaves.
If this distinction becomes a defining feature of the next AI frontier, the implications will extend well beyond technology companies. They will reach into industrial policy, economic development, institutional design and the future of work.
1. The Next AI Frontier Is About Learning How the World Works
Large language models have transformed the ways in which we access information, work with knowledge, create content and interact with software. They can synthesise ideas, interpret instructions, generate code and assist with complex intellectual tasks. Their progress has been extraordinary.
Yet fluent language is not, by itself, evidence that a system possesses a sufficiently reliable model of physical reality or can anticipate the consequences of acting within it.
Consider a simple example. An AI system can explain why a glass falls when pushed off a table. Another system might learn from observations of objects, motion, gravity and collisions to predict what will happen when a glass moves towards the table’s edge. It could then use those predictions to assess whether a robot should move an object, adjust its grip or take another route.
The distinction is between describing an outcome and developing a model that supports the anticipation of outcomes.
This is central to LeCun’s long-standing research on autonomous intelligence. In his 2022 position paper, A Path Towards Autonomous Machine Intelligence, he proposed an architecture involving learned world models, hierarchical representations, prediction, planning and self-supervised learning. His subsequent work on Joint Embedding Predictive Architectures, including video-based approaches, has explored ways for AI systems to learn meaningful representations without reconstructing every detail of their observations.
The practical significance is straightforward: an intelligent system should not need to reproduce every pixel of a scene to understand the movements that matter. It should learn which information is relevant to predicting how a situation may evolve.
This is a different emphasis from treating next-token prediction as the central organising principle for intelligence.
It does not follow that language models are obsolete, that they cannot reason, or that multimodal language models cannot process images and audio. The distinction concerns the learning objective, the representations a system develops, and whether those representations reliably support prediction and action.
World models offer a complementary research direction: learning useful structure from observation so that a system can anticipate what may happen next.
2. What Are the Signals Beyond LLMs?
Signals are observations that provide information about the state of an environment and how it changes. They can include images, video, sound, spatial measurements, motion, touch, force, equipment telemetry and the outcomes of previous actions.
A factory robot, for example, may receive visual information about objects on a production line, measurements of its own position, force feedback from a gripper, and data about the performance of nearby equipment. A mobile robot may combine cameras, depth measurements and motion sensors to navigate an unfamiliar environment.
Individually, these signals are measurements. Their significance emerges when a system learns relationships among them.
It may learn that certain movements precede a collision, that a machine’s vibration changes before a fault, or that a particular action makes an object unstable. By learning such regularities, a system can begin to anticipate changes rather than simply react to what is immediately visible.
The central idea is not that every AI system needs every available sensor. Nor is it that collecting more data automatically produces greater intelligence. The challenge is to identify which observations matter, learn the relationships that govern the situation, and use those relationships to make better predictions.
This is where world models become particularly interesting. Rather than predicting every detail of a future image, a model can learn an abstract representation of the current situation and predict how that representation may change.
The objective is to preserve the information necessary for understanding and planning while avoiding the burden of modelling irrelevant detail.
In June 2025, Meta’s research team published V-JEPA 2, a video-based world model that demonstrated progress in visual understanding, prediction and robotic planning. The researchers reported that a model trained on large-scale video data, with a relatively small amount of robot interaction data, could support picking-and-placing tasks in unfamiliar laboratory environments without task-specific training in those environments.
This remains a research result rather than proof that general-purpose physical intelligence has been solved. Nevertheless, it illustrates the direction of travel: learning from observation, predicting aspects of the physical world and using those predictions to guide action.
The broader question is no longer only whether an AI system can produce a correct answer. It is whether that system can develop a useful model of its operating environment, anticipate the consequences of alternative actions and improve its decisions through feedback.
3. Why This Matters for Countries
The next phase of AI competition may depend on more than access to the most capable language model.
Countries will still need research talent, computing infrastructure, investment, advanced semiconductor supply chains and competitive AI enterprises. However, the ability to apply intelligence to local economic and social conditions may become equally consequential.
A country’s productive capabilities are embedded in its geography, institutions, industries, energy systems, transport networks, workforce, natural resources and public services. An AI system developed elsewhere may be powerful, but its value depends on how well it can be adapted to these specific conditions.
Consider agriculture. An intelligent system intended to support farmers must contend with local soils, weather patterns, crop varieties, machinery, logistics, water availability and market conditions. A general-purpose assistant may explain agricultural practices, but a system supporting operational decisions must work with the conditions that determine what actually happens in a field.
The same principle applies to manufacturing, healthcare, logistics, energy, construction and environmental management.
The strategic opportunity for countries is therefore not simply to consume AI capabilities produced elsewhere. It is to build the ability to contextualise, integrate, evaluate and govern those capabilities within their own economies.
The World Bank’s Digital Progress and Trends Report 2025: Strengthening AI Foundations identifies four foundational areas for inclusive AI development: connectivity, compute, context and competency. These encompass access to digital networks, computing capacity, locally relevant data and the skills necessary to use AI effectively.
These foundations matter even more when moving from general-purpose chat interfaces towards systems that must perceive conditions, anticipate outcomes and support real-world operations.
For countries with established industrial capabilities, the opportunity may lie in improving manufacturing, logistics, infrastructure and scientific research. For emerging economies, it may lie in applying AI to locally important challenges without needing to replicate the entire frontier research ecosystem.
Neither opportunity is automatic. The benefits will depend on investment, skills, institutional capacity, infrastructure and the ability to translate research into functioning systems.
National competitiveness will increasingly involve asking not only, “Which AI models can we access?” but also, “What can our people, institutions and industries reliably accomplish with them?”
4. From AI Adoption to Economic Transformation
The distinction between AI adoption and economic transformation deserves particular attention.
An organisation can deploy AI assistants across its workforce without fundamentally improving the way its core operations function. It can generate reports faster while retaining inefficient processes. It can automate individual tasks while leaving the larger system poorly coordinated.
World models make a different class of applications conceivable. In principle, AI systems that learn useful relationships between operational conditions, possible actions and resulting outcomes could support more adaptive production, predictive maintenance, robotic operations, logistics planning and resource allocation.
Imagine a manufacturing facility that combines equipment telemetry, machine-vision systems, production schedules, inventory data and the operational knowledge of its workers. An intelligent system could help identify relationships between changes in equipment behaviour and production delays, compare alternative responses and support decisions before a disruption escalates.
Such a system would need more than a capable model. It would require reliable data, integration with existing machinery and software, clear operational objectives, appropriate permissions, validation procedures and human accountability.
This is the essential economic point: the value is created not by the model in isolation, but by the quality of the system into which it is integrated.
Stanford University’s 2026 AI Index reports that organisational AI adoption continued to expand in 2025, while the use of AI agents across business functions remained at an early stage. This distinction is important. Widespread experimentation does not necessarily mean that organisations have redesigned their operating models or established reliable pathways from AI capability to measurable productivity.
The economic challenge is to connect technological potential to operational improvement.
That requires new forms of collaboration between AI researchers, engineers, industry specialists, designers, policymakers, workers and organisational leaders. It also requires a willingness to redesign processes rather than simply insert AI into existing arrangements.
The economies that benefit most may not always be those that adopt the largest number of tools. They may be those that become exceptionally good at translating technological capabilities into productive, resilient and adaptable systems.
5. Industry 5.0: Intelligence Must Serve the System, Not Just Optimise the Task
This brings the discussion into the territory of Industry 5.0.
The European Commission frames Industry 5.0 around three priorities: human-centricity, sustainability and resilience. The vision extends industrial progress beyond efficiency and productivity as isolated objectives, placing technology in the service of workers, society and environmental responsibility.
World models are relevant to this direction because they offer a potential basis for systems that can better anticipate changing conditions and evaluate alternative actions.
However, technical prediction alone is not enough.
A factory may optimise production speed while increasing worker fatigue. A transport system may reduce travel time while distributing benefits and burdens inequitably. A supply chain may minimise short-term costs while becoming more vulnerable to disruption. A highly automated organisation may improve measurable efficiency while weakening the human expertise needed when unusual situations arise.
These are not merely modelling problems. They are problems of system design, priorities, incentives and accountability.
The question is not only whether an AI system can predict an outcome. It is whether the organisation has defined a worthwhile outcome, understood the relevant trade-offs, established acceptable limits and retained the capacity to intervene when circumstances change.
For this reason, I see a natural connection between world-model research and a broader approach to designing intelligent, human-centred systems. The unit of design should not be the model alone. It should be the complete operating arrangement: people, technology, processes, physical environments, objectives, permissions, feedback and governance.
Such a perspective allows intelligence to be evaluated in context. It asks whether the system improves the performance of the whole entity, rather than whether one technical component performs impressively in isolation.
6. The Governance Challenge: Prediction Is Not the Same as Trustworthiness
The prospect of AI systems that can anticipate and act also raises important questions.
A model can be useful in one environment and unreliable in another. It can learn patterns from observations that fail to represent unusual but consequential situations. It can make predictions that appear plausible while overlooking a critical constraint. And a system that performs well in a research demonstration may not be ready for unsupervised deployment in a safety-critical setting.
As AI systems move closer to physical operations and consequential decisions, evaluation must extend beyond answer quality.
Organisations will need to assess how systems behave under changing conditions, how they respond to uncertainty, whether their predictions remain reliable outside familiar environments, and what happens when they fail.
They will also need clear rules governing authority. Which decisions may a system make independently? Which actions require approval? What conditions trigger a pause, rollback or shutdown? Who is responsible for reviewing failures and learning from them?
The more directly an AI system can affect people, equipment, infrastructure or public services, the more important these questions become.
A world model is not a guarantee of safety, common sense or sound judgment. Better prediction is a capability, not a substitute for accountability.
For countries and organisations, the objective must therefore be to develop systems whose capabilities are matched by appropriate validation, oversight and mechanisms for correction.
The aspiration should be intelligent systems that can anticipate and act, while remaining governable by the people and institutions responsible for their consequences.
7. What This Means for My Work at Wonda Designs
Through Wonda Designs, my interest is in systems design, Industry 5.0 innovation and the design of entities capable of achieving sustainable, resilient and repeatable results.
I use the term entity design to describe the intentional design of the whole operating arrangement: its purpose, people, structures, processes, technologies, constraints, incentives and relationships with its environment.
This perspective is particularly relevant as AI moves from generating outputs towards supporting increasingly complex operations.
When considering an AI-enabled entity, I believe we should begin with the system that needs to function, not with the model we want to deploy.
What is the entity trying to achieve? What conditions govern its performance? Which signals provide meaningful evidence of those conditions? What relationships must be understood? Which outcomes can be predicted? What decisions should be supported or automated? Where must human judgment remain decisive? How will the system learn from errors, adapt to change and demonstrate that it is delivering the intended results?
These questions help connect AI capability to an operational architecture.
They also reflect the interdisciplinary nature of the work. Technical models must be understood alongside organisational behaviour, physical constraints, economic incentives, institutional responsibilities and the lived experiences of people using the resulting systems.
This is consistent with the integrated role I have been articulating as an IODAI: an Integrated Orchestrator, Designer, Architect and Implementation specialist. The emphasis is on connecting disciplines, aligning the elements of a complex undertaking and helping move an idea from conceptual design towards practical implementation.
In this context, world models are an important development to watch, but they are not the whole solution. Realising their value will require collaboration among AI researchers, technical specialists, domain experts and implementation partners.
For Wonda Designs, the opportunity is to help frame the right problem, design the entity in which the technology must function, establish the requirements for dependable results and orchestrate the expertise needed to implement it. Where specialist technical capabilities are required, effective delivery will depend on the appropriate expertise and partnerships rather than the assumption that one organisation should possess every capability internally.
The larger ambition is to make innovation operational: to connect what technology makes possible with what an entity must reliably accomplish.
8. The Strategic Question for the Next Decade
LeCun’s world-model vision invites us to think beyond the immediate appeal of generative AI. It directs attention towards how machines might learn from observations, develop useful representations of their surroundings, anticipate change and support action.
If this direction continues to mature, it could reshape important areas of science, manufacturing, logistics, robotics and other sectors in which understanding the evolution of a real-world system is essential.
But technical progress alone will not determine the outcome.
Countries will need the foundations that enable them to build and apply these capabilities. Economies will need to integrate them into productive activities. Organisations will need to redesign their operating models. And society will need institutions capable of governing systems whose decisions increasingly affect the world beyond a computer screen.
For those of us working across innovation, systems design and Industry 5.0, this is an invitation to widen the frame of the conversation.
We should ask not only what a model can generate, but what a complete system can perceive, understand, anticipate and improve. We should ask how it behaves when conditions change, how its performance is measured, and whether its benefits are realised without compromising human responsibility or long-term resilience.
The next AI revolution may be shaped not simply by who develops the most capable model, but by who learns to design the most effective relationship between intelligence, people, institutions and the world in which they operate.
That is where the conversation becomes bigger than AI.
It becomes a conversation about the future of production, the design of organisations, the competitiveness of nations and our collective capacity to turn intelligence into meaningful progress.
The question I am left with is this: as AI learns to model the world, will we become equally capable of designing the systems in which that intelligence can create lasting value?
References and further reading
This article interprets the keynote’s theme alongside LeCun’s published research. It is an analytical reflection, not a verbatim transcript of the presentation.
Comments