Future Development

12.1.1 The Genesis of the World Model: From Mental Maps to AI Simulations

The Conceptual Foundation. The genesis of this field lies in "Mental Model" theory, posited by Kenneth Craik in 1943. Craik suggested that human cognition relies on small-scale internal simulators that allow us to anticipate consequences before acting. We do not need to drop every glass to know it will shatter; our internal physics engine rehearses the event in latent space. This mental rehearsal is what makes planning possible. The brain evaluates outcomes without paying the price of physical failure.

Formal Definition. The modern technical architecture was formalized in the 2018 framework by Ha and Schmidhuber, which breaks the World Model into three interlocking modules. The Observation module (V) acts as the system's vision, compressing high-dimensional sensory input, millions of raw pixels, into a manageable latent representation. The Prediction module (M) functions as the memory of the system: a differentiable physics engine that anticipates future world states based on current observations and prospective actions. The Controller (C) is the action module, optimizing policy by "dreaming" inside the simulator created by the Memory module and executing the best solution in the real world only after virtual validation. The conceptual pillars have stood for decades. What changed recently is the arrival of massive video datasets and H100-scale compute, which together push simulation from theory into working practice.

12.1.2 Lexical vs. Physical Intelligence: Why World Models Surpass LLMs

The "Moravec Paradox" remains the primary barrier to entry for trillion-dollar robotics and autonomous markets: high-level reasoning is computationally cheap, but low-level sensorimotor skills are immensely difficult to automate. LLMs, despite their linguistic prowess, are "word smiths in the dark." They possess statistical knowledge of gravity but lack the grounded, physical understanding required to navigate it. World Models change the question itself: not which word is probable, but what motion is caused.

Comparative Analysis: LLMs vs. World Models

The "Black Box" Critique. LLMs function as containers of human culture, yet they remain untethered from reality. They can describe a trajectory but cannot simulate it. This lack of spatial intelligence leads to a performance plateau in embodied AI. To reach AGI, systems must move from observing "what follows a word" to predicting "what happens to a voxel" when a force is applied. Without that grounding, an AI stays a stochastic parrot, however fluent it sounds.

12.1.3 The Tripartite Technical Landscape: Three Approaches to World Modeling

The current "technological arms race" is a trade-off between visual fidelity (skin), structural understanding (skeleton), and computational efficiency (nervous system).

Approach A: Video Generation as Simulation (The "Skin" Layer). Models like Google DeepMind's Genie 3 treat pixel prediction as a proxy for world logic, generating interactive 720p environments at 24 FPS with a visual memory that can extend up to 60 seconds. The limitation is that this approach masters the appearance of reality, predicting the next pixel from statistical probability, without necessarily understanding the 3D geometric skeleton underneath. The result looks consistent but carries no explicit spatial coordinates.

Approach B: 3D Native and Spatial Intelligence (The "Skeleton" Layer). Championed by Fei-Fei Li's World Labs (Marble), this approach prioritizes explicit 3D structures, outputting meshes and spatial coordinates using technologies like 3D Gaussian Splatting (3DGS) instead of 2D frames. The advantage: 3DGS is faster and easier to edit than traditional voxels, and when you know exactly where an object sits in a coordinate system, you can plug it straight into physics engines like Unreal or Unity. That gives robotics a much sturdier "skeleton" to work with.

Approach C: Abstract Latent Reasoning (The "Nervous System" Layer). Yann LeCun's JEPA (Joint-Embedding Predictive Architecture) argues that predicting every pixel is computationally wasteful and instead works in latent representations, ignoring environmental noise (the specific glint of light on a leaf) in favor of causal structure (the direction the leaf is falling). For high-level decision-making and planning this is far more efficient, because it strips away redundant data to find what actually drives a scene.

12.1.4 Technical Deep-Dive: NVIDIA Cosmos and the 3D Paradigm Shift

NVIDIA's Cosmos platform marks the move from passive generative AI to a "Digital Twin 2.0" infrastructure. It connects pixel generation to production-ready interactive assets.

The architectural pieces of these World Foundation Models build on one another to turn raw video into a usable representation of reality. Scene perception uses advanced tokenization to convert high-dimensional visual data into compact semantic tokens, allowing the model to "understand" 360-degree panoramas and complex scene layouts with discrete precision. On top of this, trajectory planning applies causal predictive intelligence to navigation: Cosmos imagines multiple future trajectories and picks the path that respects learned physical constraints.

Two further capabilities anchor these dynamic predictions in a stable world. Temporal coherence uses physics-aware generation to maintain view consistency, so that when a camera pans away and returns, the 3D environment remains unchanged, a persistent world state rather than a fleeting video. Feed-forward reconstruction then enables the rapid generation of 3D-native assets such as depth maps, normals, and 3DGS, which can be immediately exported to simulation platforms and rendered at 24 FPS on consumer-grade hardware.

The result is a move away from "pixels as video" and toward "pixels as world states": a foundational layer where developers can post-train generalist models on task-specific robotics data, cutting development cycles from years to days.

12.1.5 The Enterprise World Model: Applications in Business and Operations

In the corporate sphere, World Models introduce what might be called "Starcraft for CEOs." With a live operational model of the business, leaders stop relying on retrospective reporting and start running predictive simulations instead.

Robotics and Embodied AI. World Models attack the "Sim-to-Real" gap by giving agents like SIMA 2 an "unlimited curriculum." A robot can master a difficult task in a safe, high-fidelity virtual environment where the physics are indistinguishable from reality, then carry that skill into varied industrial settings without manual reprogramming.

Autopilot and Logistics. For autonomous systems, World Models enable "counterfactual reasoning." A system can ask, "What if that cyclist swerves?" and simulate 10,000 variations of that edge case in milliseconds. This move from "local capability" to "verifiably safe" navigation is the key to scaling AV fleets in complex urban terrains.

Supply Chain and Operations (Digital Twin 2.0). Following Rohit Krishnan's vision, the "Enterprise World Model" treats the business as a dynamic environment that can be rehearsed before it is run. In a real estate context, this plays out across several intertwined decisions. For capital allocation, a model can simulate the ROI of a $60k roof repair against the risk of a $500k total capex failure four months later, surfacing trade-offs that a static spreadsheet would obscure. For operational efficiency, managers can simulate the revenue lift of enforcing a 15-minute response-time SLA versus the increased staffing cost it implies, finding the Goldilocks zone for conversion. And for market strategy, companies can model price-matching responses to competitors and watch the impact ripple through occupancy and margin before the first dollar is spent.

12.1.6 Limitations, Responsibility, and the Horizon of AGI

The transition to a simulation-governed world brings ethical and technical risks at the system level. The stakes change too. We are no longer managing data errors; we are managing physical safety.

Technical Constraints and Constraints of the "Dream." Even advanced models like Genie 3 face hurdles that bound what the simulation can faithfully represent. The action space available to an agent inside these worlds is still narrow, restricting the range of physical behaviors that can be rehearsed. Temporal consistency is another frontier: maintaining high-fidelity coherence over hours or days, rather than minutes, remains unsolved, and a world that drifts over time cannot be trusted for long-horizon planning. Finally, text and geography lag behind: legible signage and accurate geographic detail in generated environments fall well short of the visual quality of the rest of the scene.

The "Simulation Gap" and Hallucinations. "World Model hallucinations" are a far more dangerous liability than LLM fact errors. If an LLM hallucinates, a sentence is wrong. If a World Model hallucinates gravity or object mass, a robot damages property or injures someone. This "Sim-to-Real" gap makes alignment a matter of physical liability, which is why rigorous "reward modules" are needed to track and penalize physically impossible outcomes.

Future Outlook: The Sensory Cortex of AGI. The AI "grandmasters" remain divided. Yann LeCun views current LLMs as a potential "dead end" for true autonomy, a conviction he backed with his career: in late 2025 he left Meta, where he had been chief AI scientist since 2013, to found a startup dedicated to world models, the clearest possible signal of where he believes the next paradigm lies. Others see LLMs as the linguistic interface for a more powerful sensory engine. The consensus, however, is that World Models will serve as the "sensory cortex" for a future AGI, giving it the ability to not just talk, but to navigate, anticipate, and exist.

Digital systems are beginning to navigate our reality, not merely process our information. The companies that build and control these internal simulators will hold a rare advantage: they can rehearse the future before it happens.

12.2.1 Introduction: The Shift to Delegated Shopping

Agentic commerce hands the act of shopping to AI agents that work on our behalf, and research by McKinsey & Company suggests the shift is moving from concept to strategy. McKinsey projects that by 2030, the US B2C retail market could orchestrate up to $1 trillion in revenue from agentic commerce, with global projections reaching $3 trillion to $5 trillion. The change unbundles the traditional shopping trip. Instead of visiting vertical destinations such as Amazon or Expedia, consumers state an intent, and a personal agent acts as a concierge across a horizontal, integrated ecosystem to fulfill it.

12.2.2 The Mechanics of Agentic Commerce

According to McKinsey's analysis, agentic commerce currently takes shape across three primary interaction models.

Agent to site: Personal AI agents interact directly with merchant platforms, such as a travel agent scanning hotel websites to find and book options that fit user preferences.

Agent to agent: AI agents autonomously transact and negotiate with vendor AI agents, such as a personal shopping agent securing a bundle discount directly from a retailer's in-house AI.

Brokered agent to site: Intermediary systems facilitate multi-agent and multi-platform interactions, like a personal agent connecting with an OpenTable broker agent to secure a reservation and apply loyalty discounts.

McKinsey highlights that this ecosystem depends on infrastructure that is still being built. Three enablers stand out: the Model Context Protocol (MCP), an interoperability standard allowing agents to share context and intent across tools; the Agent-to-Agent (A2A) Protocol, which lets autonomous agents securely negotiate and coordinate across platforms; and the Agent Payments Protocol (AP2), an open standard allowing agents to make verifiable, cryptographically signed purchases. Competing with the Google-led AP2 is the Agentic Commerce Protocol (ACP), developed by OpenAI and Stripe and launched with Instant Checkout inside ChatGPT in late 2025, while Visa and Mastercard have both introduced agentic payment credentials. A protocol war is underway, and merchants are not neutral bystanders: Amazon has blocked third-party shopping agents from its storefront, and Perplexity's agent was the subject of high-profile litigation over exactly this question. Who owns the customer relationship when an agent does the shopping is the commercial fight of the coming years, and the protocols are its weapons.

12.2.3 The Automation Curve

McKinsey emphasizes that agentic commerce does not arrive in a single leap to full autonomy. It unfolds along a six-level agentic commerce automation curve, based on how much of the journey consumers are willing to delegate. McKinsey defines the levels as follows:

Level 0: Programmed convenience. The pre-agentic baseline, consisting of rules-based, "set it and forget it" subscriptions, such as recurring coffee bean deliveries.

Level 1: Assist (the cognitive sidekick). Agents gather and present information by scanning catalogs and summarizing trade-offs, but the human shopper retains all assembly and execution duties.

Level 2: Assemble (the personal shopper). Agents orchestrate purchase-ready baskets by resolving constraints (balancing price, compatibility, and delivery speed), but stop short of buying until they receive human approval.

Level 3: Authorize (the supervised executor). Consumers delegate rules and set clear guardrails (for example, "buy my preferred sneakers if they drop below $80"), and the agent executes the workflow end-to-end within those boundaries.

Level 4: Autonomize (the intent steward). Agents operate against standing, long-term goals such as maintaining airline loyalty status at the lowest cost or keeping household essentials under a monthly budget, acting proactively with only episodic human intervention.

Level 5: Networked autonomy (multiagent commerce). A multi-agent world where personal agents autonomously negotiate directly with specialized vendor agents across pricing, logistics, and payments with minimal human input.

McKinsey research notes that this curve also applies to B2B commerce, with the caveat that B2B delegation is institutional rather than personal; autonomy advances more slowly due to strict procurement policies and compliance rules, but scales far more powerfully once it takes hold.

12.2.4 Implications for Merchants and Value Pools

McKinsey's analysis shows agentic commerce collapsing the traditional sales funnel: search, comparison, and consideration merge into a single agent-mediated moment. For retailers, the competition moves away from front-end brand storytelling and toward operational trust and delivered value. Here is the hard part. If a retailer's catalogs, pricing, shipping promises, and substitution policies are not exposed cleanly through machine-readable APIs, McKinsey warns that AI agents will simply bypass them.

Because agents bypass traditional ad channels, McKinsey notes that retail media networks reliant on ad-based models face a decline in revenue. To survive, businesses will need to adopt new monetization models identified in McKinsey's research, such as charging real-time negotiation fees, facilitating multi-brand bundle revenue sharing, offering premium subscriptions for specialized vertical agents, or monetizing anonymized agent-filtered consumer analytics.

12.2.5 Navigating Trust and Risk

McKinsey stresses that in this model, trust stops being a marketing asset and becomes infrastructure. Building it comes down to two things: explainability, so consumers understand why an agent made a specific choice, and reversibility, so cancellations and human overrides are easy.

Agentic commerce also introduces new systemic risks. Payments, compliance, and fraud systems built to stop bots must learn to verify legitimate agents instead, through a "Know Your Agent" (KYA) standard. McKinsey's studies emphasize that the industry needs new accountability protocols to prevent failures where a single faulty prompt sets off a chain of unintended autonomous errors across interconnected systems.