Statistical Pragmatic Competence

In the AI Story

Winograd's framework distinguished sharply between processing (formal manipulation of representations) and understanding (embodied, situated engagement with the world). The distinction was both philosophical and practical: genuine understanding required being-in-the-world, and without it, open-domain competence would be impossible. The large language models falsified the practical prediction while leaving the philosophical distinction intact. They demonstrate that the territory accessible to processing-without-understanding is vastly larger than anticipated—the models handle genuinely ambiguous instructions, maintain conversational context across thousands of tokens, and produce artifacts satisfying human intentions with remarkable reliability, all without possessing experiential grounding in the domains they operate across.

The mechanism is utterly unlike classical AI. SHRDLU parsed sentences according to explicit grammars, mapped them onto semantic representations, and evaluated representations against world models—each step inspectable and rule-governed. Large language models process tokens through layers of attention mechanisms weighting relationships based on statistical co-occurrence patterns learned from trillions of words. There is no grammar in the classical sense, no semantic representation mapping onto a world model, no world model. There is a vast, distributed, implicit encoding of how language is used across billions of contexts—so complex that no engineer can explain precisely how any particular output was generated. The outputs work not because the system understands but because statistical patterns in training data encode, implicitly, the contextual knowledge that makes linguistic phenomena interpretable.

The implications force a revision of Winograd's framework. Understanding remains necessary for caring—for judging not just correctness but value, not just whether an output matches specifications but whether the specifications themselves are worth matching. But understanding is not necessary for vast territories of pragmatic competence previously assumed to require it. The question shifts from 'What can machines do without understanding?' (far more than predicted) to 'What is understanding for?' (knowing why outputs matter, caring about who they serve, asking what is worth building). The expansion of competence without understanding makes the human capacity to care—to have stakes, to judge value—the only remaining check on capability deployed without purpose.

Origin

The concept emerged as a theoretical necessity when large language models began passing tests—the Winograd Schema Challenge, open-domain question answering, contextual instruction following—that Winograd's framework had classified as requiring 'full-bodied thinking.' The models achieved the results through statistical patterns rather than embodied reasoning, creating a phenomenon neither classical AI (which claimed processing could produce understanding) nor Winograd's critique (which claimed open-domain competence required understanding) had predicted. The term 'statistical pragmatic competence' is deliberately awkward, reflecting the awkwardness of the phenomenon: a capability that should not exist according to frameworks dominating thought about language and computation for half a century.

Key Ideas

Processing vs. understanding distinction survives. The gap Winograd identified between formal manipulation and genuine comprehension remains real—language models demonstrate exactly what processing without understanding looks like at scale.

Practical implications narrower than predicted. Open-domain competence does not require being-in-the-world; it can be approximated through statistical sampling of how beings-in-the-world use language—a correction to Winograd's framework.

Fossils of understanding. Training data encodes traces of situated human engagement—not understanding itself, but its textual residue, which models reconstruct paleontologically to predict how understanding would behave.

The caring boundary. What statistical pragmatic competence cannot provide is stakes—judgment informed by vulnerability, caring about whether outputs serve or harm, the question of what is worth building at all.