Abstract

This paper provides a comprehensive analytic critique of epistemological claims made on behalf of self-supervised video learning systems, with particular focus on Meta's V-JEPA 2 architecture and its claimed achievement of "world understanding." Drawing exclusively on the resources of analytic philosophy and cognitive science , Frege, Quine, Putnam, Carnap, Tarski, Fodor, Goodman, Kripke, Sellars, Dretske, and Brandom , we demonstrate that such systems cannot achieve understanding in any philosophically robust sense. The critique proceeds through four interconnected problems. The problem of reference: neural networks operate entirely on representations without making contact with referents, achieving Sinn without Bedeutung. The problem of models: latent spaces are formal constructions that repeat the failures of Carnap's Aufbau, achieving structure without content. The problem of concepts: learned detectors have no determinate content (Fodor's disjunction problem), learned patterns may not project (Goodman's grue), and concept boundaries are no more determinate than training data allows. The problem of normativity: no finite training determines a rule (Kripke), perception provides no foundation (Sellars), and information-processing is not representing (Dretske). We establish that the success of large language models does not transfer to vision systems because language is already a compressed world-model saturated with human meaning, while visual data carries no such semantic inheritance. We provide a positive account of understanding as operation within the space of reasons (Sellars, Brandom), demonstrate that neural networks cannot enter this space, and prove three impossibility theorems establishing that the failure is architectural, not empirical: no representation-manipulating system can ground reference, no statistical learner can acquire normativity, and no syntactic engine can enter the space of reasons. The paper distinguishes legitimate applications (warehouse robotics, autonomous vehicles in controlled environments, video compression) from fraudulent claims (world understanding, semantic world models, physical reasoning). We argue that the systematic misuse of cognitive vocabulary , "understanding," "world model," "learning" , constitutes epistemic laundering with practical consequences for policy, investment, and deployment. The conclusion is unambiguous: V-JEPA 2 does not understand the physical world, and no computational system operating on the same principles ever will.