What if the next AI breakthrough isn't a better chatbot?

Love it or hate it, AI is increasingly good at helping write your emails, summarise a 100-page report, explain quantum physics, and generate a convincing image of just about anything. Give a modern AI model the right tools, and it can tackle complex, multi-step tasks, write code, and even build a small app or game with surprisingly little input.

ChatGPT, Claude, Gemini, and their competitors showcase increasingly impressive reasoning and agentic capabilities. But you’ve probably noticed some limitations, too. Sycophantic responses, overly verbose or formulaic prose, and difficulty grappling with the subtleties of complex problems remain familiar frustrations. It’s far from perfect. And while AI can be remarkably capable inside a computer, getting it to reliably do things in the physical world is a different challenge altogether.

Consider asking a robot to pick up a cup from a cluttered table without knocking over anything around it. The robot needs to do more than recognize what a cup looks like. It needs to estimate where the cup is, anticipate how it might move, predict what will happen when its gripper makes contact, and account for the objects in its path. It needs to anticipate the consequences of its actions before committing to them.

Today’s AI models can be adapted to tackle these problems, but generating a plausible answer is very different from reliably predicting the consequences of a physical action. It’s one reason some AI researchers believe the next major advance in machine intelligence may not come from making large language models (LLMs) even larger. Instead, progress could come from teaching AI systems to build useful predictive models of the world around them.

Meta is one of the companies pursuing this idea through the Joint Embedding Predictive Architecture (JEPA). Its V-JEPA 2 research offers an intriguing glimpse of what this alternative approach to learning could achieve.

AI is good at talking, but can it really understand?

Today’s most prominent generative AI systems are built around a deceptively simple idea: learn patterns in enormous amounts of data, then use those patterns to produce an appropriate output. For an LLM, that typically means predicting the next token — a word or part of a word — based on the preceding context. Train a model at sufficient scale, and it can learn grammar, factual associations, coding patterns, and surprisingly sophisticated ways of solving problems, including problems it hasn’t encountered in exactly the same form before.

This approach has proved extraordinarily powerful. It would be a mistake to dismiss LLMs as mere autocomplete. Modern models can reason through complex problems, interpret images, use external tools, and work through multi-step tasks. Multimodal systems can also process video and other information about the physical world.

But a model that can describe what should happen next is not necessarily a model that can reliably predict what will happen if it takes a particular action. That distinction matters as AI moves beyond chat windows. An assistant that gives you a slightly wrong answer is frustrating. An agent that sends the wrong contract email, a robot that drops an expensive device, or an autonomous system that misjudges a moving obstacle can cause real problems.

JEPA offers an alternative approach. Instead of trying to predict the next token or reconstruct every detail of an input, a JEPA-style model learns representations of data and predicts the representation of another observation from the information available. The goal is to capture meaningful structure without reproducing every pixel, texture, or other detail that may be irrelevant to the task.

Consider our robot reaching for a cup. It doesn’t necessarily need to predict every reflection on the cup’s surface. It needs a useful representation of the cup’s position, its relationship to the gripper, and how the surrounding scene might change as the robot moves. If the model can predict these changes accurately enough, a planning system can use those predictions to evaluate possible actions before carrying them out.

That is the promise of a world model: not simply describing the world, but learning to anticipate how it might change.

The idea has some intuitive appeal. Humans don’t always reason in words, after all. We can visualize an object falling, imagine how a movement will affect something we’re holding, or intuitively anticipate the consequences of an action without first translating the problem into a verbal explanation. A JEPA-style model doesn’t need to express its predictions in language either.

There’s a catch, though. Predicting what happens next isn’t the same as understanding physics, and a model’s predictions are only useful if they’re sufficiently accurate to guide it to choose the right action. JEPA isn’t a shortcut around the need for training data, either. Its promise lies in learning to predict something different, not learning everything with less data.

What does Meta say about JEPA?

Even so, Meta’s results offer a glimpse of what’s possible. Its V-JEPA 2-AC model successfully completed 65–80% of pick-and-place tasks involving unfamiliar objects and environments. The experiments took place in controlled laboratory settings, and the model had already been adapted using robot video data, so it wasn’t learning to control a robot from scratch. However, it showed that knowledge learned from video could help a robot handle tasks it hadn’t been specifically trained to perform.

Other researchers are exploring how world models might plan further ahead. A very recent October 2026 H-JEPA study, for example, reported a jump from 18% to 73% success on a simulated maze-navigation benchmark by using a hierarchical planning approach. That’s an encouraging result, but a simulated maze is a long way from a robot navigating your home or a wearable assistant helping navigate your bike ride home.

World models clearly have a lot left to prove. But they don’t need to replicate human understanding to be useful. If they can help AI predict the consequences of an action more reliably before taking it, that could be a meaningful step towards machines that do more than simply tell us what to do.

What could this mean for your next AI agent or phone?

World models like JEPA may seem confined to robotics labs, but their longer-term implications could extend far beyond.

For instance, today’s phone assistants are increasingly expected to do more than answer questions. They need to understand what’s on your screen, interpret photos, navigate apps, remember relevant context, and complete multi-step tasks without losing the plot. Imagine asking an assistant to organize a weekend trip. It needs to compare travel times, check hotel options, fit activities around opening hours, and revise the itinerary when a booking falls through.

An LLM can already help with much of this. It can interpret your preferences, generate an itinerary, use search tools, and reason through the constraints you provide. But an agent also needs to track the task’s changing state. If a train is delayed or a museum is closed, it must understand which other parts of the itinerary are affected and decide how best to adapt.

More generally, predictive models could help an agent evaluate the likely consequences of different choices, particularly when those choices affect a complex environment. For a travel assistant, that might mean estimating how a change in departure time affects later activities. For a robot, it might mean predicting whether a particular movement will knock over an object. Wearable assistants could also benefit from real-time information about obstacles and hazards.

Of course, not every AI application needs JEPA, and a conventional algorithm may be more suitable for many tasks. But future AI systems may benefit from combining language models with other specialized models that predict how the relevant environment or task state evolves. An LLM could interpret an instruction, clarify the user’s goal, and break it into steps. A world model could predict the consequences of candidate actions. A planning system could compare those outcomes, while conventional software carries out the chosen action.

That kind of architecture would combine the strengths of language-based reasoning with a more explicit mechanism for anticipating the consequences of action.

Beyond bigger models

The LLM boom has encouraged a compelling narrative: bigger models, more data, and better reasoning will deliver increasingly capable AI. There are good reasons to expect continued progress along that path. Language models are becoming more capable, and multimodal systems are expanding what they can perceive and do. But scaling isn’t the only avenue for progress. Another possibility is that more capable AI will emerge from combining different learning approaches, each designed to capture different aspects of the world.

Language understanding isn’t going away. JEPA isn’t about to replace ChatGPT for writing emails or Claude for generating code. JEPA also hasn’t solved the problem of reasoning. What ideas like JEPA do offer is a different way to think about machine intelligence: learning useful predictive representations directly from observations, then using those predictions to help decide what to do next.

The next leap in AI may not come from choosing between a model that talks and a model that predicts the world. It may come from combining the two — giving AI systems the ability not only to reason about what they should do but also to anticipate what is likely to happen when they do.