V-JEPA 2: the new visual brain that learns the world in order to anticipate it

By Kike Gandia · AI & Security · 4 min read

What if an AI could imagine what is going to happen without ever having seen it before? That's the promise of V-JEPA 2, Meta AI's new model, which doesn't just "see" but understands the physics of the world. At QuantumSec we have spent years anticipating the convergence of AI, cyber defense and autonomy. V-JEPA 2 marks a turning point: a model that doesn't depend on labels but on physical reasoning learned from video. And that changes everything.

What is V-JEPA 2 and why does it matter for security and applied AI?

V-JEPA 2 (Video Joint Embedding Predictive Architecture, version 2) is a self-supervised video model that learns the physical laws of its environment through predictive reasoning in latent space. But unlike other models, it doesn't predict pixels. It predicts abstract representations (tubelets) that encode motion, gravity, occlusion, object permanence and spatial structure.

That makes it a "world model", a category that is still rare but extremely promising. Why? Because it lets you anticipate what is going to happen, not just describe what is happening. A qualitative leap that transforms robotics, autonomous surveillance, predictive control and decision-making in physical systems.

How V-JEPA 2 works

The model is trained in two phases:

Phase 1 — Action-free pretraining
More than 1 million hours of video and 1 million images. The goal is to learn the dynamics of the environment by predicting masked representations of the visual future, not full images.
Phase 2 — Action-conditioned fine-tuning (V-JEPA 2-AC)
Trained on 62 hours of robot videos from the DROID dataset. It learns to link actions to future consequences, which makes it possible to use MPC-style controllers (model predictive control) to imagine multiple trajectories and pick the most effective one.

It doesn't predict pixels: it predicts tubelets, abstract representations that encode motion, gravity, occlusion, object permanence and spatial structure.

Direct application: robot control without retraining

With V-JEPA 2, a robot can carry out physical manipulation tasks in unfamiliar environments without task-specific training or reinforcement learning. Examples tested by Meta:

80% success on pick-and-place tasks in new scenes.
Robots that act on vision and latent simulation alone, with no explicit rewards.
Planning with fast inference for real-time execution.

80%: success rate on pick-and-place tasks (in new scenes, with no task-specific training)
1M h: of video in pretraining
62 h: of robot videos in fine-tuning (DROID dataset)

This is crucial in adversarial environments, like the ones we deal with at QuantumSec:

Intervention in physical infrastructure with no prior connection.
Security operations at remote or unmapped facilities.
Defensive robots or autonomous agents with independent reasoning.

Benchmarks: it beats previous models on almost everything

+8% Something-Something V2: Action and motion: eight points above the previous state of the art.
27% → 39.7% EPIC Kitchens: Video anticipation, the biggest improvement of the set.
~84 and 76.9 Perception Test and TempCompass: Accuracy achieved on each.

V-JEPA 2 isn't just good at predicting motion: it also anticipates intent and the structure of the physical world.

This shows that V-JEPA 2 isn't just good at predicting motion; it also anticipates intent and the structure of the physical world.

Implications for security, autonomy and defense

From QuantumSec's point of view, V-JEPA 2 represents a key step toward:

Anticipation Predictive cyber-physical agents: Able to anticipate physical threats before they happen.
Intervention Autonomous inspection robots: In conflict, sabotage or containment scenarios: nuclear, industrial, SCADA.
Simulation Physical threats in virtual environments: As part of a proactive defense system.
Deception Dynamic physical honeypots: Systems that adapt their behavior in real time based on predictions about their environment.

V-JEPA 2 opens the door to models that don't just detect anomalies but predict them before they surface.

V-JEPA 2 opens the door to models that don't just detect anomalies but predict them before they surface.

Compared with LLMs: visual intelligence vs. textual intelligence

While LLMs like GPT work on language and symbolic patterns, V-JEPA 2 operates on the physical world. There's no text, only spatial, dynamic and visual reasoning. That makes it the perfect complement for integrated AI systems. At QuantumSec we foresee dual architectures:

Language model (LLM) textual intelligence

Works on language and symbolic patterns.
Reasoning, strategy, instructions.

Visual and spatial model physical intelligence

Operates on the physical world: spatial, dynamic and visual reasoning.
Physical anticipation, perception, control.

The dual architecture the article foresees: the future of autonomous agents isn't textual, it's multimodal and predictive.

The future of autonomous agents isn't textual: it's multimodal and predictive. And this is the first serious step in that direction.

Access and experimentation

Meta has released the code and weights for V-JEPA 2:

🧠 GitHub repository
🧪 Checkpoints on Hugging Face.
🧰 PyTorch examples for inference, classification and latent prediction.

At QuantumSec we are already evaluating integration with active defense agents in physics simulators, and MPC connectors for industrial tasks.

V-JEPA 2 doesn't just observe. It intuits. It learns without labels, predicts without pixels and acts without instructions. We are entering a new era of visual intelligence capable of physical reasoning. And in security, that changes everything.

At QuantumSec we are exploring its integration in scenarios such as:

Autonomous visual reconnaissance of critical threats.
Defensive robotics with predictive control.
Simulation of hostile environments to train agents.

War, defense and automation won't be dominated by whoever sees the most data, but by whoever anticipates it first. And V-JEPA 2 is the first step toward that predictive dominance.

In one sentence. V-JEPA 2 doesn't just observe: it intuits. It learns without labels, predicts without pixels and acts without instructions.

Related service

AI and LLM penetration testing

Topics

#AI #Maritime-OT-ICS-cybersecurity #Computer-Vision #Artificial-Intelligence

Keep reading

Audit our AI systems