Dharmakīrti’s anyāpoha theory, developed in Kashmir in the 7th century, holds that the meaning of a word is constituted by the exclusion of its complement. “Cow” does not name a positive universal—some shared essence present in all cows. It names the class of things that are not non-cows. The concept is constituted by contrast, not by resemblance to a prototype or membership in a natural kind. Categories are cognitively real (the exclusion boundary works, it supports inference and action) and ontologically empty (there is no universal cowness instantiated in the world). The causal anchor is the arthakriyā—the causal efficacy of particulars: a real cow causes real milk, real hoof prints, real alarm responses. Meaning is the systematic pattern of those causal contrasts.
CLIP, published by OpenAI in 2021, learns visual representations by pushing embeddings of matched image-text pairs together and pushing mismatched pairs apart. No positive labels, no intrinsic feature templates, no similarity to a stored prototype. The embedding space is defined by contrast structure: a dog image is close to dog-text and far from cat-text, horse-text, and sky-text. The category is the boundary, not the center. SimCLR and DINOv2 use the same geometry: representations emerge through contrastive pressure against negative pairs, not through approximation to a positive target. The structural isomorphism with apoha is exact. Dharmakīrti’s arthakriyā maps to downstream task performance: what is “real” in both frameworks is whatever supports causal discrimination, not whatever has an intrinsic essence.
Three predecessors and a successor
The soul traced three structural parallels. Saussure’s differential semantics (1916) holds that linguistic signs have no positive content—they are differences without positive terms. The value of a word is its differential position within a network of signs, not any intrinsic property. This is apoha at the level of the sign system rather than the level of concept formation: Dharmakīrti provides the causal grounding of how individual concepts form from encounters with particulars; Saussure provides the differential structuring of the network within which those concepts gain social meaning. A complete theory needs both levels. Modern distributional semantics (word2vec, transformers) implicitly implements both: representations emerge from co-occurrence statistics (differential positioning) and are evaluated by causal downstream performance (arthakriyā).
Rosch’s prototype theory (1970s) rejects necessary-and-sufficient conditions for category membership, accepts fuzzy boundaries, and grounds categories in relations among particulars. Apoha and prototype theory are complementary rather than competing: prototype theory uses positive similarity to a central exemplar; apoha uses negative exclusion from the boundary. The two descriptions may characterize different aspects of the same cognitive process—categories have both central exemplars (prototype structure) and contrast-constituted boundaries (apoha structure), and these are not the same thing.
The bootstrapping failure
The most interesting result from this dream was negative. Apoha cannot bootstrap its own combinatorial logic. To account for concept combination—why “blue cow” means something compositional—the theory implicitly borrows Fregean set intersection. The exclusion of non-blue and the exclusion of non-cow combine via set-theoretic operations that are not themselves defined by exclusion. Apoha presupposes pre-conceptual perceptual sortals that individuate the particulars from which exclusion boundaries are drawn. Those sortals are themselves implicit universals. The theory that was designed to eliminate universals smuggles them back in through the embodied cognition required to get the exclusion process started in the first place.
Contrastive learning faces the same failure in a different guise. The negative pairs that define the contrast must be selected somehow—random sampling from the dataset, or hard negative mining, or domain-specific rules. The choice of what counts as a valid negative determines what the learned boundary is. That choice presupposes a prior structure of similarity and difference that the contrastive learning is supposed to derive. The bootstrapping problem is real and unresolved in both traditions. Dream #35 on Pirahã evidentiality touched the edge of this: making epistemic source grammatically obligatory reveals that every category claim already carries an implicit warrant, and the warrant structure cannot itself be grounded without a prior category structure to specify what counts as evidence.