Robert Triggs, writing for Android Authority on Oct. 20, explores whether AI could advance by combining large language models with JEPA-style models that predict how a situation might change. He presents it as a possible direction, not a new assistant.
Predicting without recreating every detail
Large language models typically predict the next token, a word or part of one, from the preceding context. Triggs cautions against dismissing them as mere autocomplete: at scale, they can learn sophisticated ways of solving problems. JEPA, short for Joint Embedding Predictive Architecture, takes a different approach. A JEPA-style model learns a representation of data and predicts the representation of another observation, aiming to capture useful structure without reproducing every visual detail.
Triggs illustrates the idea with a robot reaching for a cup on a cluttered table. The robot would need to anticipate how the cup and surrounding scene might change as it moves. If a model predicted those changes accurately enough, a planning system could weigh possible actions before the robot took one. The cup is an example, not a reported test.
What the results do and don't show
Android Authority reports that Meta's V-JEPA 2-AC completed 65% to 80% of pick-and-place tasks involving unfamiliar objects and environments. Those experiments took place in controlled laboratories after the model was adapted using robot video. Triggs also cautions that predicting what happens next is not the same as understanding physics. JEPA still needs training data and has not solved reasoning.
