AI Needs a New Language

GEUL — An artificial language for unambiguous communication between humans and AI

Why the Age of Prompt Engineering Is Over

As models get smarter, "what you show" matters more than "how you say it." The shift from prompts to context, and the seven pieces of infrastructure context engineering needs.

Why RAG Is Not Enough

RAG searches natural language as natural language. Three fundamental limitations of embedding similarity search, and why structured semantic search is the alternative.

Why Clarification Is Necessary

Natural language has to grow longer to resolve ambiguity. In a structure with identifiers and typed places, the trade-off between clarity and brevity breaks, and clarity becomes the precondition for indexing, validation, and consistency checking.

Why Mechanical Verification Is Necessary

Natural language has no concept of an "invalid sentence," so information enters the context unchecked. Why mechanical verification, filtering form against a specification, must come first in the pipeline.

Why Filters Are Necessary

Not all valid information is needed information. The filter that picks only what is needed now by relevance, reliability, and recency, and why those criteria become field operations in a structured representation.

Why Consistency Checks Are Necessary

Information that is correct on its own can contradict itself once gathered. The three kinds of contradiction (temporal, between sources, inferential), and why detection becomes field comparison and arithmetic in a structured representation.

Why Is an Artificial Language Needed?

Natural language evolved for human communication. Ambiguity, redundancy, and implication are strengths for humans but causes of hallucination for AI. Neither programming languages nor existing semantic frameworks are the answer. A new artificial language satisfying six conditions simultaneously is needed.

Why Exploration Is Necessary

When knowledge grows so large that even the index exceeds the context window, one-shot search hits its limit. Why exploration, choosing a direction at each level of a hierarchical map, is needed.

Why Does Natural Language Create Hallucinations?

Hallucination is not an LLM bug — it is a structural inevitability of natural language. Four flaws — ambiguity, absent sources, absent confidence, absent time — make it unfixable by scaling alone.

Why MD/JSON/XML Won't Work

Markdown, JSON, and XML were each made for a different era and purpose. They have structure but no unified meaning, and no place for sources. Why a separate format is needed to record and accumulate AI reasoning.

Why Programming Languages Are Not Enough

Programming languages describe procedures. They cannot describe the world. JSON provides structure but no meaning. Even LISP merely borrows syntax.

Why Embedding Vectors Are Not Enough

Rearranging embedding vectors breaks the model. Avoiding breakage means rebuilding the model from scratch. What we need is not transparency inside the black box, but a transparent layer outside it.

Why Esperanto Failed

Esperanto failed because it had to be learned, tried to replace existing languages, and existed for communication. An artificial language for AI takes a different path: humans need not learn it, it does not replace natural language, and its purpose is recording and verification.

Why Wikidata

GEUL does not reject Wikidata. It loads the dump without loss and turns each of about 120 million entities into one GEUL document holding every language. Designing a semantically aligned identifier (SIDX) from its classification and frequency statistics is ongoing research.

Why WordNet?

Build a verb system yourself and it has gaps, is arbitrary, and lacks grounding. GEUL learned the structure by building a provisional codebook from WordNet's 13,767 verbs, accumulated over 40 years. Because WordNet is not CC0 and is tied to English, this provisional edition will be replaced by a verb codebook without WordNet.

Why Claims, Not Facts?

Truth is inaccessible; all we are left with are someone's claims. Why a knowledge system should take claims, not facts, as its primary data, and give source, time, and confidence a place to be recorded.

Why a Semantically-Aligned Index?

If an ID were knowledge rather than an address, a single bitwise operation could narrow the candidates without any JOIN. This essay covers SIDX, the semantically-aligned identifier GEUL is researching, and why its width will be set from usage statistics.

Why 16-Bit?

A single 16-bit word is designed to be the unit of three worlds: the network stream, the file format, and the LLM vocabulary. Why 16 bits, and not 8, 32, or variable length.

Why Is Structured Memory Necessary?

AI writes files well but cannot find the files it wrote. The problem is retrieval, not storage. The path from a single map of what is where, like a table of contents or an index, to structured memory.

Why Cache Reasoning as Code?

Instead of having an LLM reason from scratch every time, crystallize each inference into code and cache it. A JIT-compiler-like structure that calls the LLM only on a cache miss, and what it achieves over time.

Why Annotations Should Be Indexes

Annotations are written for humans. But when you have 10,000 functions, machines need to read them too. Turn annotations from narrative into indexes, and full scans become instant lookups.

Why We Must Leave It Empty

GEUL leaves 75% of the 16-bit TAG WORD space that opens every record empty. The lessons of IPv4, Unicode, and ASCII — the cost of filling is irreversible, while the cost of leaving empty is three bits per record.