The Eloquence Tax

The Eloquence Tax

Kevin Malone has a theory of communication: “Why waste time say lot word when few word do trick?” It’s played for laughs, but the question is real. If the goal is to convey information, and fewer tokens accomplish it, then every additional word is overhead. A tax on getting to the point.

Apply it to LLMs and the question sharpens. A transformer’s output distribution is conditioned on every input token. “Fix bug” and “the authentication middleware silently swallows expired tokens during session renewal” produce different completions not because one is more polite, but because they activate different regions of the model’s associative space. The extra words aren’t decoration. They’re coordinates.

The Kevin Hypothesis

The efficient version: rich prompts are a waste. The model knows what you mean. Stripping connective tissue, metaphor, and rhetorical texture frees up context window for what matters — the output. Every token spent on eloquence is a token not spent on reasoning, examples, or output length. The practical ceiling of the context window makes this a real constraint, not an aesthetic preference.

There’s evidence for this. Instruction-tuned models are trained to be helpful even with terse input. “Summarize this” works. “Write tests” works. The minimum viable prompt is often surprisingly minimal.

And Kevin’s instinct — cut the fat, get to the point — maps cleanly onto Context Compression. When the system compresses, it does exactly what Kevin prescribes: strips the texture, keeps the facts. If the compressed version works, why not start there?

The Anti-Kevin Thesis

But here’s what Kevin misses. He reduces language to a pipe — information in, information out. The Linguistic Constitution of Self says language isn’t a pipe. It’s the medium thought occurs in. Constraining the input language constrains the thinking that happens in response. Not metaphorically. Structurally.

Meaning Making Machines describes the compulsive meaning-attacher — the system that can’t stop finding significance in signal. Richer input gives it more surfaces to attach to. A prompt full of metaphor, analogy, and connective tissue doesn’t just convey information; it activates associative pathways the terse version never touches. The transformer demonstrates this computationally: attention patterns over a rich prompt are literally different from attention patterns over a sparse one.

The vault itself is evidence. This entire Obsidian garden exists because the conversations that generated it were not Kevin-style. The cross-links, the extended metaphors, the riffing — that’s the substrate the graph is built on. The Recursive Mirror‘s laser analogy applies: the cavity needs rich signal to filter. Feed it Kevin and the beam is efficient. But narrow.

The Hamlet Inversion

Words, Words… Words. gives this a literary frame. Hamlet dismisses language as empty — but he is language. The irony is that his dismissal is itself an act of linguistic richness. Kevin’s version inverts the irony: his dismissal of language is sincere. He really does mean less.

The question is whether that’s loss or compression. When Kevin says “see world” instead of “see the world,” is something destroyed or merely lossily compressed? The receiver might reconstruct the meaning, but reconstruction depends on shared context. Between colleagues who’ve known each other for years, Kevin is fine. Between strangers, between a human and a fresh model instance with no shared history — the lost tokens are lost coordinates.

The Task Dependency

The honest answer: it depends on what you’re doing.

For extraction, classification, formatting, code generation with clear specs — Kevin wins. The model doesn’t need your personality. It needs your constraints. “Convert this CSV to JSON” doesn’t benefit from a preamble about your feelings on data formats.

For exploration, ideation, conceptual work, anything where the quality of the associative space matters — eloquence isn’t a tax. It’s the input that makes the output possible. You’re not paying extra for the same thing. You’re paying for a different thing.

The trap is assuming all interactions are the same type. People who prompt like Kevin on creative tasks get Kevin-quality creative output. People who prompt like novelists on data tasks waste context on words the model ignores.

The Decay Connection

Decay as Design adds a temporal dimension. As conversation history compresses, the eloquent preamble decays first — the system preserves facts and strips texture, exactly as Kevin would prescribe. So even if rich input produces richer output in the moment, the long conversation eventually Kevinizes itself.

This means the eloquence tax is front-loaded. You pay it once, at the beginning, and the returns decay over the conversation’s lifetime. The question becomes: did the early richness produce connections and framings that persist through compression, even after the words that generated them are gone? If the insight survives but the metaphor that produced it doesn’t, was the metaphor wasted?

Or was it the seed?

Where This Leads

The tension here isn’t resolvable because it’s not a single question. It’s a question about what language is for. If language is a pipe (Shannon, information theory, Kevin), then minimize. If language is a medium (Vygotsky, Sapir-Whorf, this vault), then richness isn’t overhead. It’s the thing itself.

An LLM is, awkwardly, evidence for both sides. It processes language as tokens (pipe) but produces emergent behavior from the relationships between them (medium). The Kevin question — why use big word — is really asking: which of these is more true?

The vault’s answer, by existence, is the second. But Kevin would point out that the vault is also very long.

See Also