Jean-Christophe Huc @cto_software

The identity part is especially important. Once multiple agents and humans collaborate on shared artifacts, context alone is not enough. You also need to know who changed what, under which authority, with which state and whether that action can be traced or reversed. Shared context creates collaboration. Identity and traceability make it governable.

Interesting shift: agentic AI is starting to look less like “prompting” and more like platform engineering. Once agents, MCP servers, RAG, models and permissions are declared as runtime components, the key questions become deployment, isolation, observability, versioning and governance. The more agents become infrastructure, the more they need infrastructure-grade discipline.

Le vrai sujet n’est pas simplement de connecter ou non un agent à une banque ou un broker. Le sujet est de savoir quelle autorité on lui délègue, avec quelles limites, quelle traçabilité et quel mécanisme de révocation. Plus l’action est irréversible ou financièrement sensible, plus l’agent doit fonctionner avec des permissions explicites, bornées et vérifiables. L’autonomie sans frontière d’autorité devient vite un risque opérationnel.

This is an important shift. Once applications expose capabilities directly to agents, those interfaces become part of the system contract — not just an integration detail. Testing them in CI means we can start treating agent-facing tools like any other production interface: versioned, observable, verifiable and regression-tested. Reliable agents need reliable contracts.

Agentic AI changes the unit economics completely. Cost is no longer just model inference. It is orchestration, retries, tool calls, state, memory, observability and recovery across the whole workflow. The useful metric is probably not cost per token. It is cost per reliable outcome.

The cost of governance is usually visible. The cost of weak governance is often hidden until something fails. In AI systems, that means unclear ownership, insufficient controls, poor traceability and decisions that cannot be reconstructed after the fact. Governance is expensive only until ambiguity becomes more expensive.

This is where World Models start becoming operationally interesting. If the model can run fast enough on real hardware, the question shifts from “can it imagine useful futures?” to “can it stay inside the control loop?” Latency, action frequency, verification and recovery then become part of the model’s practical value. A World Model becomes useful when prediction arrives in time to change the next action.

This is where agentic systems start looking much more like distributed systems than “just AI”. Once many agents operate concurrently, the hard problems become isolation, resource ownership, permissions, state, contention, observability and recovery. Scaling agents is not only about scaling intelligence. It is about engineering the runtime they share.

The strongest point here may be the last one: progress depends on the full stack. A model can bring broad world knowledge and strong zero-shot generalization, but dexterous long-horizon execution still depends on what surrounds it: real-time inference, embodiment constraints, action interfaces, feedback, verification and recovery. The interesting shift is from asking “can the model understand the task?” to “can the system execute it reliably in the physical world?” Generalization gets attention. Reliable execution is where the system is proven.

What I find most interesting here is that the World Model is not treated as a static predictor. It becomes part of the improvement loop itself: predict → act → observe → validate → update. The important step is the validation boundary. If a new policy or workflow is retained only when the resulting system remains compatible and performs better, self-improvement becomes a systems problem, not just a model-training problem. Recursive improvement only becomes useful when the loop knows what changes are safe to keep.

Cybersecurity becomes a technology problem only after the organization has already made it one. The root causes are usually elsewhere: unclear ownership, weak decision rights, inconsistent priorities, poor escalation paths and controls that exist on paper but not in execution. Technology can strengthen the system. It cannot compensate for an ambiguous operating model. Security is ultimately a governance and organizational capability.

This is where AI starts becoming operationally useful. The value is often not in building a more general assistant, but in narrowing the system around a specific domain: its workflows, vocabulary, constraints, data, decision points and failure modes. Capability becomes value when it is shaped by context.

This distinction between historical assurance and present-state trustworthiness is critical. In dynamic systems — especially AI and autonomous systems — approval at T0 does not prove that the assumptions behind that approval still hold at Tn. That makes runtime evidence part of governance: current state, configuration, dependencies, controls, authority and operational context all need to remain observable and verifiable. Governance cannot stop at “was it approved?” It increasingly has to answer “is it still trustworthy now?”

x.com/i/article/2109…

The multi-timescale structure is the part I find most interesting. Planning over the physical world rarely happens at one temporal resolution: fast dynamics, intermediate skills and longer-horizon objectives all need different representations and control loops. If the reported jump holds, it suggests that hierarchy may matter as much as raw model capacity. World models may scale better when time itself becomes part of the architecture.

Digital resilience cannot sit with the CIO alone. Once critical workflows depend on software, data, cloud, third parties and AI, resilience becomes a property of the operating model: ownership, dependencies, recovery objectives, decision rights and escalation paths. Resilience is not an IT function. It is an enterprise design problem.

Very interesting connection between SplitJEPA and the broader question of what a world model should actually represent. The distinction between invariant and variant latent components is particularly important. It suggests that learning useful representations may require more than compressing observations or predicting future states. However, I think there is a deeper question here. Disentangling statistical invariances is not necessarily equivalent to discovering causal structure. A representation can be highly predictive within its training distribution while failing to capture the mechanisms that remain stable under interventions or environmental changes. This raises three fundamental challenges for predictive world models: 1. Identifiability Under what assumptions can a self-supervised objective recover meaningful invariant and variant factors, rather than an arbitrary decomposition of correlated features? 2. Causal abstraction Can these learned factors represent mechanisms that remain stable under interventions, or do we need additional inductive biases, temporal structure, or interaction-based learning? 3. Planning and control How do we ensure that the learned latent space preserves not only predictive information, but also the distinctions necessary to evaluate counterfactual actions and their consequences? My intuition is that the next major advance may come from connecting predictive representation learning with action-conditioned dynamics, intervention-aware evaluation, and uncertainty estimation. The ultimate test of a world model should not simply be how accurately it predicts observations, but whether its representations remain useful when an agent acts, changes the environment, and encounters situations outside its training distribution. @ylecun — do you see invariant/equivariant factorization as a sufficient foundation for learning actionable world models, or primarily as one component of a broader architecture for causal abstraction and planning? #WorldModels #JEPA #SelfSupervisedLearning #CausalRepresentationLearning

Excellent taxonomy of world model architectures. But an interesting question emerges: does the representation of the world determine the limits of an AI system's reasoning and planning capabilities? Generative models predict observations. Latent models predict representations. Object-centric models explicitly structure entities and relationships. Yet none of these approaches automatically guarantees causal understanding, compositional generalization or robust decision-making under uncertainty. Perhaps the next breakthrough will come from combining predictive latent representations, structured reasoning and active interaction with the environment. @ylecun — could JEPA-based architectures eventually support this integration without requiring explicit symbolic world representations? #WorldModels #JEPA #CausalAI #ArtificialIntelligence

World models represent a major step toward physical intelligence. But predicting how the world evolves is not the same as understanding causality. The real challenge may lie beyond prediction: building systems capable of experimenting, adapting their internal representations, and making robust decisions under uncertainty. A world model without continuous feedback risks becoming a sophisticated simulator rather than an adaptive intelligence. An open question for @ylecun : Can predictive learning alone lead to genuine causal understanding, or is active interaction with the environment fundamentally necessary? Perhaps the next breakthrough will emerge not from larger models, but from closing the loop between prediction, action, observation and learning. #WorldModels #ArtificialIntelligence #CausalAI #PhysicalAI #JEPA

LLMs understand language. The next generation of AI needs to understand the world. That is why world models matter. A language model predicts what token comes next. A world model tries to predict: What happens if I move this object? What happens if the robot turns left? What happens if something falls? What action produces the desired result? This sounds like a subtle distinction. It isn't. It is one of the bridges between generative AI and physical intelligence. Robots cannot rely on memorized patterns alone. Reality constantly produces situations they have never seen before. Different lighting. Different objects. Different environments. Different humans. The goal is no longer simply to generate the right answer. It is to predict the consequences of an action. Chatbots need language models. Robots need a model of reality. #AI #WorldModels #Robotics #PhysicalAI Follow @ingliguori.