Prediction can happen in a representation
World models have renewed interest in an important AI question: can a system learn how its environment changes, then use that knowledge to choose an action? Joint Embedding Predictive Architectures, or JEPA, explore prediction in a learned representation space. For video, the aim is to capture useful structure without having to reproduce every pixel of a future frame.
Meta’s V-JEPA 2 research, published in June 2025, combines large-scale learning from video with action-conditioned post-training for robot planning. The distinction matters. Learning from observation and learning what changes after an action are related tasks, with different data requirements. The resulting robot experiments provide evidence for particular physical tasks; they do not establish a general model for running a company.
Keep the research result in context
The V-JEPA 2 paper describes limitations including camera sensitivity, error accumulation over longer planning horizons and goals specified through images. Those details matter as much as a demonstration. A representation may be effective for one task yet omit information another task requires. A useful deployment has to establish which states, actions and constraints the model can reliably handle.
For a business workflow, the transferable lesson is architectural: represent the current situation, distinguish possible actions and evaluate the consequences against an explicit objective. This is a design perspective, rather than evidence that a video model transfers to operational planning. Forecasting, optimisation and simulation already offer established tools for many of these problems. Choose the model from the decision requirements.
Give business events a structured meaning
A support operation might need to decide which cases require escalation. A planning team might compare staffing options under several demand scenarios. Both need reliable state: case status, available capacity, permitted actions and the cost of a poor decision. A language model can help extract or explain that state, while explicit rules and tested numerical models handle the parts with firm constraints.
My SENTiVENT research focuses on extracting company-specific events and the sentiment implied by financial news. It offers a concrete example of moving beyond surface wording: identify what happened, which participants were involved and what the text supports. Such structured evidence can be an input to a decision system. It does not, by itself, establish causality or predict a future business outcome.
Prototype one decision before a general planner
Pick a recurring decision with a clear owner. Define the available observations, permitted actions, constraints and outcome measure. Establish a rules-based or statistical baseline. If a learned model is proposed, evaluate it on held-out periods and changed conditions, then inspect how its errors affect the decision. Historical correlations alone are insufficient grounds for claims about an intervention.
Introduce uncertainty explicitly and provide a human review point when the system lacks relevant evidence. Keep explanation separate from the mechanism that scores actions, so a persuasive narrative cannot conceal a weak recommendation. The near-term opportunity is disciplined decision support: better structured information, measurable alternatives and bounded action. World-model research is worth following because it asks useful questions about prediction and planning; operational value still has to be demonstrated in the organisation’s own environment.
Sources & further reading
Written by Gilles Jacobs, AI & data engineering consultant at Exverbai.
