The shift in three lines
A decade of pattern recognition — and what comes next
For a decade, the frontier of artificial intelligence was defined by a single organising principle: train a system on enough human-generated data, optimise it to predict what comes next, scale it up — and watch capabilities emerge that routinely surprise even the researchers who built the system.
The results have been genuinely astonishing. Large language models can reason, converse, write, and generate code at levels that have reshaped entire industries and triggered a multi-trillion-dollar infrastructure investment wave. The commercial momentum is real, the capability advances are substantial, and the enterprise adoption curve is still early.
And yet, beneath the surface of this achievement lies a structural limitation that the field has been slow to confront openly. These systems are powerful at completing patterns, but they lack an internal sense of the world those patterns describe. They understand how our world works through a form of second-order interpretation — absorbing descriptions of reality without ever encountering reality itself. They have no first-principles grasp of physics, motion, cause and effect, or consequence.
A language model can explain, in fluent prose, that a glass will shatter if dropped from a table. It has no internal representation of the weight, the trajectory, the contact, or the outcome. That distinction barely registers when AI is asked to summarise a document or draft a communication. It becomes a hard constraint the moment AI is asked to navigate an unstructured physical environment, coordinate a complex organisation in real time, or reason about how a strategic decision will cascade through a live system.
The transition now under way is from systems that recognise and reproduce patterns to systems that build internal models of the world — and use those models to plan, simulate, and act with consequence-awareness.
Systems trained on human text learn the statistical structure of language. They predict what word, sentence, or argument typically follows another. They are extraordinarily capable within this frame — and bounded by it. They have no causal model of the world; they have a very sophisticated map of how humans have described it.
Systems that build internal simulators can ask, repeatedly: if I take this action, what happens next? They can rehearse in digital environments where failure is cheap. They can reason about physical constraints, social dynamics, and causal chains before committing to a course of action in the real world.
This analysis explores what world models are, why the field's most credentialed researchers are converging on them, what they imply for enterprise AI deployment, and whether current AI infrastructure projections are adequately sized for a world in which they succeed.
What world models are — and why they matter
A world model, at its core, is an internal simulator. It allows a system to answer a simple question, repeatedly and quickly: if I do this, what happens next? Humans rely on this faculty constantly — we picture a glass tipping before it falls; we rehearse a difficult conversation before having it. Until recently, machines could not do this well.
Teaching a robot to recognise a cup is straightforward. Teaching it to pick one up without shattering it is another matter entirely. The physical world is unforgiving: objects have weight and inertia; surfaces have friction; forces compound in ways that text descriptions cannot fully capture. World models offer a path to machines that have internalised these constraints rather than merely read about them.
There are two distinct categories of world model, each pointing toward a different frontier — and each with significant implications for how organisations compete and make decisions.
These systems learn the governing logic of the material world: gravity, friction, thermodynamics, fluid dynamics, structural mechanics. Rather than learning purely through real-world trial and error — slow, expensive, and potentially dangerous — they absorb the rules of physics through simulation, practising in digital environments where failure is fast and cheap.
- Warehouse robots navigating dynamic, unstructured spaces with fewer collisions
- Autonomous vehicles rehearsing low-probability edge cases before encountering them on real roads
- Industrial process optimisation where physical constraints bound every decision
The most consequential decisions will ultimately draw on both capabilities simultaneously. Physical world models guide the robots moving goods through a port; social world models simulate the demand shock, labour response, and geopolitical disruption that changed the routing. Decisions in one domain inform actions in the other — and planning becomes continuous rather than episodic.
What makes social world models especially powerful is that they offer something beyond faster analysis: a higher-fidelity operating environment for decision-making. Enterprises already spend enormous effort attempting to anticipate how others will respond — how competitors will move, how markets will interpret signals, how boards will react under pressure. Today, those judgements rely on experience, static analysis, and intuition. Multi-agent simulation offers something closer to a living model of human systems.
Critically, world models are not forecasting tools in the conventional sense. They do not predict the future. They reveal plausible futures — ranges, paths, and feedback loops that conventional single-point forecasts cannot surface. Forecasting assumes a single correct outcome. World models reveal the shape of possibility.
The researchers building the next paradigm
What distinguishes the current world model moment from prior cycles of AI research enthusiasm is the calibre and credibility of the researchers converging on it. These are not peripheral voices staking out contrarian positions. They are, in several cases, the architects of the current AI paradigm — and they are arguing, in their own research and ventures, that the paradigm is incomplete.
Yann LeCun
LeCun spent years as one of the most influential figures inside a major AI lab before departing to found AMI Labs with world models as its explicit foundation. His Joint-Embedding Predictive Architecture — JEPA — is designed to build machines that develop internal models of the world through observation, in the way humans do, rather than through text prediction. LeCun has been publicly and persistently critical of the thesis that scaling language models alone will reach general intelligence. World models are his alternative.
Fei-Fei Li
Li's ImageNet dataset helped ignite the deep learning revolution that produced today's dominant AI systems. She founded World Labs around a related but distinct idea: spatial intelligence. The premise is that genuine intelligence requires not just recognising objects in images but understanding how those objects exist in space, interact with each other, and change over time. Li's bet is that machines need to inhabit a model of three-dimensional reality — not merely classify it at a frame level.
The significance of this convergence is hard to overstate. These are the researchers whose earlier work produced the current AI era — and they are now directing their energy toward what they believe the current era is missing. Their departure from leading research positions to build dedicated ventures around world models signals something more than academic interest: it signals strategic conviction.
The people who built today's AI are quietly telling us it isn't enough. That signal deserves more attention than it has received from investors and enterprise strategists.
Not a competition — an integration
Much of the current discourse frames LLMs and world models as competing paradigms, asking which one will prevail. This framing fundamentally misunderstands how complex intelligence emerges — and how the most capable AI systems of the near future are likely to be structured.
Complex intelligence does not arise from a single dominant approach. It arises from the orchestration of many. Just as the human brain integrates specialised modules into coherent thought — perception, memory, language, motor control, social cognition — advanced AI architectures are likely to combine large language models, physical simulators, and social reasoning engines into unified systems.
The most capable AI system of the foreseeable future will likely use language models as the interface for instruction, explanation, and communication — while relying on world models for planning, consequence modelling, and action. Physical laws constrain motion; social rules constrain behaviour. In both cases, intelligence emerges from understanding how local actions ripple outward through a system.
The analogy that holds here is instructive: language gives AI fluency; world models give it situational awareness. For much of the recent history of enterprise AI, we have treated these systems primarily as tools that produce answers. World models suggest something more ambitious — systems that understand the context in which those answers will be acted upon.
High-value financial, legal, and knowledge tasks that are fundamentally about language — document analysis, contract review, communications, code generation, structured reasoning over text — remain squarely in LLM territory. The transition to world models does not displace this; it extends beyond it.
Tasks requiring consequence-aware planning, physical coordination, or the simulation of complex adaptive systems — logistics, manufacturing, organisational strategy, market dynamics, risk modelling — are precisely where world models extend the frontier of what AI can usefully do.
Enterprise and financial applications
World models are not a distant research project. The early applications are already visible across logistics, manufacturing, financial services, and strategic planning — with the most consequential use cases only beginning to be defined.
| Domain | Model type | Application | What changes |
|---|---|---|---|
| Logistics & supply chain | Combined | Physical robots navigate warehouses and ports; virtual models simulate demand shocks, labour responses, and routing disruptions in parallel. | Planning becomes continuous rather than episodic — physical and social intelligence inform each other in real time. |
| Insurance & reinsurance | Physical | High-fidelity simulation of hurricane seasons, flood events, or wildfire dynamics to stress-test insured-loss distributions across complex portfolios. | Risk modelling moves from actuarial tables to dynamic physical simulation — probability distributions become richer and more realistic. |
| Financial markets | Social | Multi-agent environments where each agent reflects a different market participant with distinct incentives, information sets, and behavioural constraints. | Strategy rehearsal replaces static scenario analysis — decisions are tested against adaptive adversaries, not fixed assumptions. |
| Policy & macroeconomics | Social | Simulation of how a policy shock — a rate change, a tariff, a regulatory shift — cascades through heterogeneous populations with different response functions. | Policy evaluation gains a dynamic component: second- and third-order effects become visible before implementation. |
| Manufacturing | Physical | Digital twins of production facilities where physical constraints, equipment states, and process parameters are modelled at high fidelity. | Process optimisation and fault prediction become physics-grounded rather than purely statistical — the model understands why, not just when. |
| Corporate strategy | Social | Board-level governance simulations where strategic options are tested against simulated competitor responses, regulator behaviour, and market interpretations. | High-stakes decisions are rehearsed against adaptive, realistic counterparties rather than static scenario trees. |
The common thread across these applications is leverage. In domains where mistakes are expensive and foresight creates disproportionate advantage, the value of reliable simulation compounds faster than its cost. The barrier has historically been computational — high-fidelity simulation of physical dynamics or complex social systems is expensive. That cost calculation is now shifting, and the shift is consequential.
Infrastructure implications
The entire AI infrastructure build-out — chip projections, data centre construction timelines, energy capacity planning — has been sized around a single assumption: that the future of AI is larger language models running on more compute. The question world models raise is whether those projections are measuring the right thing.
Readers of our companion note on the AI build-out scenario framework will recognise the implication: the four supply-side variables that determine total AI capital — silicon service life, data-centre cost, chip architecture mix, build-out elongation — were calibrated against today's LLM-dominated workload mix. A world-model-augmented future raises every one of them, and adds a fifth: the cost of constructing high-fidelity simulation environments at scale.
The competitive dimension
If the transition to world models develops as described here, it will not merely be a technical refinement. It will reshape the competitive landscape of the AI industry itself — and the enterprises that depend on AI as a source of advantage.
In the current paradigm, competitive advantage in AI accrues primarily to those with the largest models, the most data, and the most compute. These advantages are real but also scale with capital — and capital, at sufficient scale, is accessible to a broad range of well-resourced actors.
Competitive advantage may soon depend as much on who builds the most faithful simulation of reality — physical, social, and economic — as on who trains the largest model.
World models introduce a different dimension of differentiation: the fidelity and accuracy of the simulated reality. A superior world model is not simply a function of compute spend — it reflects the quality of the underlying training environment, the richness of the synthetic data, the sophistication of the physics engine, and the depth of the behavioural modelling. These advantages are harder to replicate through capital alone.
For enterprise decision-makers, the implication is that the relevant question is no longer only "how much AI can we afford?" but "how accurately can our AI represent the systems we operate within?" Organisations that develop or access world models of their own competitive environments — their markets, their supply chains, their regulatory contexts — will gain a form of strategic foresight that has no precedent in conventional analysis.
Organisations that invest in world-model capabilities now — building simulation environments, synthetic data pipelines, and multi-agent modelling infrastructure — gain the advantage of early calibration. World models improve with use: the organisation that has been running simulations for two years will have materially better models than one starting from scratch.
The risk of waiting for world models to mature before investing is asymmetric. If the transition develops as the field's leading researchers expect, early movers will have accumulated both proprietary model quality and organisational capability that is difficult and slow to replicate. Competitive position in AI is increasingly path-dependent.
From producing answers to understanding the world
The large language model era gave AI the ability to communicate about the world with remarkable sophistication. It produced systems that can summarise, argue, translate, code, and converse at levels that have genuinely surprised their creators and reshaped industry after industry.
What it did not produce is systems with an internal sense of what the world actually is — how its physical constraints work, how its social dynamics unfold, how local actions propagate through complex systems over time. That gap has been the quiet limitation of every deployed AI system of the past decade.
World models are the field's most serious attempt to close it. The convergence of the AI era's most credentialed researchers on this problem — departing leading institutions to build dedicated ventures around it — is a signal that deserves to be taken seriously by investors, operators, and strategists.
This transition is not imminent in the sense of being months away. The commercial deployment of sophisticated world models is a multi-year journey, and large language models will remain the dominant paradigm for enterprise AI for the foreseeable future. But the trajectory is now clearly established, the researchers are now clearly committed, and the infrastructure investment required — which current forecasts have not yet adequately priced — is now clearly materialising.
Organisations that recognise this shift early will be better positioned not just to adopt AI, but to deploy it where it matters most: in decisions that shape real systems, in real time, with real consequences.
If large language models gave AI fluency, world models will give it situational awareness. For much of its recent history, we have treated artificial intelligence as a system that produces answers. World models suggest something more fundamental: machines that understand what it means to be inside the world they are describing.
This report has been prepared by Lualdi Advisors for informational and educational purposes only. It draws on publicly available research, academic literature, industry commentary, and Lualdi Advisors' proprietary analytical framework. References to named researchers and organisations are purely illustrative and informational; they do not imply endorsement, affiliation, or any commercial relationship with Lualdi Advisors. This material does not constitute investment, legal, tax, or financial advice and should not be used as the basis for any investment decision. All forward-looking statements and scenario analyses are inherently uncertain; actual outcomes may differ materially. Lualdi Advisors makes no representations or warranties regarding the accuracy or completeness of the information herein. Past performance is not indicative of future results.
