20.7 Path to AGI & World Models

There is no established engineering recipe for AGI. The term bundles different capabilities—reasoning, learning, planning, acting, and oversight—and different researchers assign different importance to each. A world model is one useful system concept in this debate: a model that predicts how an environment may change, especially when an agent takes an action.

That definition is narrower and more testable than saying a system “understands the world.” It leads to engineering questions about state, actions, transition error, planning horizon, uncertainty, and control.

1. From Sequence Prediction to Action-Conditioned Prediction

Language modeling estimates a distribution such as $p(x_{t+1}\mid x_{\le t})$ . A world-model objective may instead include observations $o$ , actions $a$ , rewards or success signals $r$ , and a goal $g$ :

p(o_{t+1:t+k}, r_{t+1:t+k} \mid o_{\le t}, a_{t:t+k}, g)

The action variable is the key addition. The system must estimate not only what is likely to happen, but how outcomes differ when an agent chooses another action. A model can generate plausible video while being a poor action-conditioned simulator, so visual quality alone is not an adequate world-model metric.

World models have a long lineage in model-based reinforcement learning. World Models and Dreamer are influential examples of learning latent dynamics and using imagined trajectories to improve control [1], [2]. Modern multimodal work extends the setting to richer observations and larger, more open-ended environments.

2. JEPA and Representation-Space Prediction

JEPA is an influential proposal for predicting useful representations instead of reconstructing every surface detail [3]. The motivation is not that pixels or tokens never matter. It is that an agent may benefit from representing stable, decision-relevant structure while avoiding the cost of predicting unpredictable detail.

This remains a design family, not a proof that representation prediction is sufficient for general intelligence. An implementation still needs safeguards against representation collapse, a definition of what context and target contain, temporal and action conditioning, and evaluations showing that the learned state supports downstream decisions.

3. Capability Axes Beyond One Architecture

AGI discussions become more concrete when separated into capability axes.

Axis	Engineering question	Example measurement
Learning and retention	Can the system acquire new skills without erasing old ones?	Forward transfer, retention, and forgetting curves
Reasoning and verification	Can it solve and check multi-step work?	Held-out tasks with executable or human-verified outcomes
State and world modeling	Can it predict action-conditioned changes?	Multi-step rollout error and counterfactual accuracy
Planning and tools	Can it choose and revise actions under constraints?	Task success, tool errors, recovery rate, and cost
Grounding and embodiment	Does behavior remain valid under real observations and interventions?	Simulator-to-real transfer and intervention tests
Alignment and oversight	Can people constrain, inspect, and stop it?	Policy compliance, escalation, auditability, and rollback drills

Scale and test-time compute can improve several axes, but neither guarantees the others. Tool use can compensate for missing factual or computational capability, while also introducing permissions, prompt injection, side effects, and recovery problems. Capability must therefore be evaluated as a system, not inferred from parameter count or one benchmark.

4. World Models as an Engineering View

A world model is best treated as a component in a controlled system.

Component	Contract	Failure to test
State representation	Which facts persist, expire, or remain uncertain?	Stale or contradictory state
Action schema	Which actions and parameters are legal?	Invalid or over-privileged action
Dynamics model	How does state change after an action?	Compounding rollout error
Planner or policy	How are candidate actions selected?	Reward hacking or brittle search
Observation update	How is real feedback reconciled with prediction?	Ignoring tool errors or new evidence
Verifier	Which constraints can reject a plan or result?	Confident but invalid transition
Recovery control	How are pause, retry, rollback, and escalation handled?	Irreversible side effects

This view applies beyond robotics. A support agent has a state consisting of ticket status, policy, permissions, and prior actions. Its “world model” includes predictions about what a refund, database write, or email will change. High-risk actions should use previews, dry runs, idempotency keys, approval boundaries, and an external source of truth rather than relying on the model’s imagined state.

5. Evidence Boundaries: Genie and Other Frontier Claims

Google DeepMind describes Genie 3 as a general-purpose world model that generates interactive environments from text and runs them in real time [4]. The same official page reports important limits: restricted action spaces, imperfect multi-agent interaction, consistency failures, and interaction horizons measured in minutes. DeepMind’s statement that this could be a stepping stone toward AGI is a provider research claim, not evidence that AGI has been achieved or that this is the only path.

OpenAI’s earlier Sora research used the “world simulators” framing for video generation [5]. That work is evidence that scaled video models can show useful simulated behaviors; it does not establish a reliable causal model, persistent state, or action-conditioned planning in every generated scene.

For frontier systems, record what is disclosed and what is undisclosed: training data, action interface, parameter count, intervention protocol, evaluation set, failure distribution, and access conditions. Do not infer internal architecture from a demo.

Established

Model-based control and latent dynamics are established research areas.
Exact surface reconstruction is not the only useful prediction objective.
Error compounds across long imagined rollouts and must be measured.

Emerging

Large multimodal models that combine generation, interaction, and agent training.
Learned simulators used to create training curricula and counterfactual data.
Hybrid systems in which an LLM plans, a simulator predicts, and external tools verify.

Speculative

One world-model architecture is the decisive missing ingredient for AGI.
Success in a generated environment transfers safely to the physical world.
Scaling alone resolves grounding, long-horizon planning, and oversight.

6. Practical Engineering Roadmap

An engineering team can use the world-model view without claiming to build AGI.

Define the state and source of truth. Specify fields, freshness, uncertainty, ownership, and which external system is authoritative.
Define action contracts. Give each tool a typed schema, permissions, cost, idempotency behavior, timeout, and reversible or irreversible classification.
Predict before execution. Ask for the expected state delta and risk, then compare it with the real observation after execution.
Use safe simulators. Run SQL dry runs, payment previews, email drafts, and code sandboxes before side effects.
Measure multi-step error. Evaluate one-step accuracy, rollout degradation, calibration, recovery after an unexpected observation, and goal completion.
Build counterfactual tests. Vary tool failures, malicious observations, missing data, delayed results, and user goal changes.
Gate side effects. Set human approval and automated abort thresholds for money movement, deletion, external communication, and privilege changes.
Rehearse recovery. Pin the artifact bundle, log every state transition, maintain a last-known-good policy, and test rollback.

A release gate should include task success, critical-slice minima, unsafe-action rate, tool-call validity, state-consistency errors, recovery rate, latency, and cost. Compare candidates with paired trials and uncertainty rather than a few curated trajectories.

7. Practical Takeaway

World models sharpen an engineering problem: predict how actions change state, expose uncertainty, and correct predictions from real observations. They are a valuable research and system-design direction, but they do not settle the AGI debate. The practical standard is observable capability, documented limitations, controlled actions, and recoverable operation.

Quizzes

Quiz 1: What distinguishes an action-conditioned world-model objective from ordinary next-token prediction?

It explicitly conditions future observations or outcomes on candidate actions, allowing counterfactual questions about how a chosen action changes the state.

Quiz 2: Why can a visually convincing video model still be a poor world model for planning?

Plausible pixels do not guarantee correct causal transitions, persistent state, calibrated uncertainty, or accurate responses to interventions over multiple steps.

Quiz 3: What is the practical motivation for JEPA-style representation prediction?

It may focus capacity on stable, decision-relevant structure instead of reconstructing every unpredictable surface detail, but downstream usefulness still has to be demonstrated.

Quiz 4: Why should Genie 3’s AGI framing be labeled a provider claim?

The official page presents both a research direction and significant limitations. It does not prove that the system is AGI, that the approach is sufficient, or that alternatives cannot work.

Quiz 5: What makes a business agent a small world-model system rather than only a text generator?

It tracks operational state, predicts the effects of tool actions, reconciles predictions with real observations, and uses permissions, verification, and rollback to move toward a goal safely.

References

Ha, D., & Schmidhuber, J. (2018). World Models. arXiv:1803.10122.
Hafner, D., et al. (2020). Dream to Control: Learning Behaviors by Latent Imagination. arXiv:1912.01603.
LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. OpenReview.
Google DeepMind. (2026). Genie. Official model page.
OpenAI. (2024). Video generation models as world simulators. OpenAI.