NLPR

“current action-conditioned world models do not actually learn physical dynamics; they memorize 2D visual patterns...current architectures behave more like 2D visual pattern matchers than systems that genuinely understand physical dynamics”