Opinion: exploring AI‑IR, an intermediate language for ChatGPT and Grok

An experiment was run using a Japanese article about gold‑buying conditions. ChatGPT transformed the prose into a compact tag‑based format (e.g., @DEFINE AU=金, C1|GEO↑ = AUDEMAND↑) that encodes facts, inferences, triggers and causal links. Grok then restored the tags to readable Japanese, preserving the logical structure while changing the surface wording. The author argues that the benefit is not merely token compression but a clear separation of meaning components, which could aid auditing and reduce the risk of mis‑classifying inferences as facts.

Building on this, the piece proposes an “AI‑IR” (AI Intermediate Representation) – a lightweight schema for AI‑to‑AI communication. It references older agent‑communication standards such as FIPA‑ACL, KQML and the A2A protocol, noting that they handle message routing rather than meaning decomposition. The author plans to draft a minimal AI‑IR spec with ChatGPT, have Grok audit it, and iterate, while keeping human‑readable Japanese for parts that involve people.