Delimiters in Languages
I propose that languages employ a universal duality in which the same set of symbols can serve two distinct roles. First-order expressions, such as directives or computations intended for immediate execution, contrast with second-order expressions, such as quoted text or stored information that is meant to be represented without being interpreted. This distinction enables complex cognitive processing through the use of delimiters: explicit markers that signal a shift in interpretation between these layers.
| Article | Description | Revisions |
|---|---|---|
| What Quotation Marks, String Literals and DNA Have in Common | I show the duality at work in three symbolic systems: human speech (quotation marks), programming languages (string literals) and genetic code (repeated DNA sequences that, like quotation marks, set apart the viral fragments a bacterium stores as a memory instead of evaluating them). | July 6, 2024: initial publication / November 2026: retitled (formerly “First-Order, Second-Order Expressions, and Delimiters in Languages”) |
| Special Tokens: The Quotation Marks of LLMs | I apply the duality to LLMs, whose special tokens mark roles, turns and reasoning, and look at what happens when such a delimiter appears in ordinary text. | February 22, 2025: initial publication / November 2026: retitled (formerly “LLM Delimiters and Higher-Order Expressions”); description of the Llama 3.2 experiment corrected |
| Formulaic Delimiters in the Iliad and the Odyssey | I examine how Homer, without quotation marks, marked the shifts between narration and speech in the Iliad and the Odyssey: formulaic phrases that open a speech, close it, or pass the word from one speaker to the next, cues that a listening audience could follow. | December 26, 2025: initial publication / November 2026: references to Milman Parry and Mark W. Edwards |
| Why XML Tags Are So Fundamental to Claude | I argue that the XML tags Anthropic recommends are more than a formatting tip: by teaching Claude to read tags as delimiters, its creators made it “aware” of the boundary between first-order and second-order expressions, which may be what makes it so effective at interpreting layered meaning. | March 1, 2026: initial publication / March 2, 2026: updated to address the inference framework vs. training ambiguity / November 2026: Anthropic’s own statements on XML tags added |
| Transitive Policies for Agents | Agent permissions do not automatically compose. I connect the first-/second-order distinction to authorization, information-flow restrictions, and action history that must survive delegation. | November 14, 2026: initial publication |