Neural representation is a key neuroscientific concept meant to bridge brain and mind, or brain and behavior. But what is meant exactly by a “neural representation”? Conventionally, a neural representation is a correspondence between something in the brain and something in the world, a “code”. The encoding view of representations faces two critical issues, empirical and theoretical. Empirically, I will show that neural codes do not have the properties required to naturalize mental representations. Theoretically, it raises the problem of “system-detectable error” (Bickhard): if the brain sits at the receiving end of the code, then how can it know if the representation is wrong? As John Eccles has concluded, the logical implication is dualism - there must be a “decoder” that translates brain properties to world properties. Consequently, a number of authors have argued that representations are not only homuncular but also unnecessary: adapted behavior results not from calculations on an internal copy of the world, but from coupling between body and world - “the world is its own best model” (Brooks). Anti-representationalism introduces crucial concepts missing from the conventional view (embodiment, autonomy, dynamicism) but it struggles to explain some aspects of anticipation and abstraction. I argue that the problem with representation is to think of it as a “thing” that can be manipulated and observed (like a painting), which collides with the dynamical nature of brain activity. I suggest to shift the focus from the encoding properties of brain states, a dualistic concept, to the representational properties of brain (and body) processes, such as anticipation and abstraction.