For the past several years, the technology and defense communities have been captivated by the extraordinary capabilities of Large Language Models (LLMs). These systems have transformed how we interact with data, accelerating knowledge work, and reshaping entire industries.
But as engineering leaders, we must look past the current hype cycle and ask a deeper question: What architecture will lead to machine intelligence that can reason, plan, and reliably interact with the physical world?
The answer likely does not lie in scaling these autoregressive LLMs indefinitely. While LLMs encode a great deal of statistical world knowledge, they are not grounded physical world models in the control or robotics sense. Because of this, they can be brittle in common sense and causal reasoning, especially in physical domains.
In current production systems, one practical mitigation is to use retrieval, tool use, and agent decomposition to constrain context and validate outputs, keeping the models within a reliable operational zone.
But looking forward, a foundational architectural shift is on the horizon. One of the most compelling research paradigms addressing these limits is the Joint Embedding Predictive Architecture (JEPA), championed by Turing Award winner Yann LeCun. While JEPA is currently an emerging research frontier, it represents a fundamentally different way of thinking about AI, and it could be the breakthrough that mission systems, autonomy, and government AI modernization have been waiting for.
Moving Beyond Token Prediction
Many leading language models and sequence models are based on autoregressive prediction: guessing the next token on what came before as a whole. This involves compute-intensive pretraining, often refined with supervised fine-tuning and preference-based alignment methods.
They excel at pattern completion and language synthesis, but their reasoning and planning behavior can be brittle, especially when tasks require grounded causal models, formal guarantees, or long-horizon execution.
The phenomenon of ‘hallucinating’, for example, is a byproduct of this approach, as the model prioritizes the statistical likelihood of the next token over actual factual grounding.
JEPA takes a different approach. Instead of predicting raw, granular outputs (like the exact next word or the exact pixels in a video frame), JEPA uses self-supervised learning to predict abstract representations of missing or future data. It operates in a latent representation space. It is designed to capture higher-level structure while abstracting away irrelevant detail, which may be advantageous for perception, prediction, and planning.
The Rise of “World Models”
The broader trajectory of research like JEPA points toward the development of AI World Models.
A world model acts as an internal predictive model of environment dynamics, allowing an AI system to anticipate state transitions and outcomes. Instead of simply generating a plausible-sounding paragraph, these architectures are being researched to answer questions like:
- If I move this physical object, what happens next?
- If a specific sensor degrades, what cascading operational risks emerge?
- What sequence of autonomous actions will safely achieve a mission objective?
This predictive understanding is exactly what is required to advance robotics, logistics optimization, and complex operational planning.
Why This Matters for Defense and GovCon
For government and defense missions, AI is most valuable when it improves high-stakes decisions and accurately predicts operational outcomes, including:
- Autonomous vehicles and unmanned systems
- Sensor fusion across ISR (Intelligence, Surveillance, and Reconnaissance) platforms
- Battlespace simulation and multi-domain operational planning
These domains require grounded contextual reasoning. As JEPA and similar self-supervised, representation-learning architectures mature, they will be uniquely suited for these environments for three reasons:
- Learning from Unlabeled Data: Self-supervised learning allows systems to learn from massive amounts of raw, unlabeled sensor data, video, telemetry, radar, and operational logs, which the government possesses in abundance.
- Modeling Physical Relationships: Variants like Video-JEPA (V-JEPA) are currently being researched to understand motion, physics, and interactions directly from video streams without relying on pixel-perfect recreation.
- Grounded Predictive Modeling: May reduce certain failure modes of pure token generation by enabling simulation-based evaluation of candidate actions, though reliability still depends on data coverage, verification, uncertainty handling, and system design.
A Complement to Generative AI, not a Replacement
It is important to note that JEPA-style architectures are not being researched to replace LLMs. A plausible future architecture is complementary rather than monolithic. A likely future architecture will be deeply multimodal and composite:
- LLMs will serve as the natural language interface and human interaction layer.
- JEPA-style world models may serve as a core engine for perception, prediction, and planning in physically grounded systems.
- Multi-Agent Platforms will serve as the planning and execution layer, driving actions based on those predictions.
The future of enterprise and government AI will not be defined solely by chat interfaces. It will be defined by systems that understand the operational world and help human operators make superior, faster decisions inside it. While still in the research phase, JEPA may prove to be the architectural foundation that makes that future a reality.
The author, John Mark Suhy, is CTO at Greystones Group.