COMPUTATIONAL REALISATION AND THEORETICAL ELIMINATION The Ontology of I-language, Causal Abstraction, and the Limits of Linguistic Explanation

Eliminative proposals hold that, because large language models gen- erate language through continuous tensor operations, the attribution of a discrete grammatical architecture is theoretically redundant. This essay argues that the step from numerical implementation to theoret- ical elimination requires an additional argument which has not been supplied. Continuous computation is compatible with discrete func- tional organisation; transformer attention is not inherently symmetric; and a lower-level description does not exhaust the levels at which a system may be explained. The essay proposes causal abstraction, tested by interchange intervention, as the appropriate criterion for distinguishing an organisation that a researcher can describe from one that the system actually implements. It then applies the same standard to continuous accounts of negation, to contextual domain restriction, to the internalist treatment of reference, and to inferences from linguistic fluency to mental states