Write Like It's 1866: LLMs Relearn Telegraphese

What it claims

A one-sentence instruction to write in telegraph-style "cablese" roughly halves output tokens for text that another model will read, with no loss in downstream question-answering. Two details decide whether it pays off: where in the pipeline you compress, and whether the writer model's reasoning can be switched off. An older-style codebook (dictionary substitution) saved only about 10%. The author's reason: "substitution can't remove words the author already wrote."

How it works

The exact instruction tested:

"Write a complete record in telegraphese; drop articles and filler; abbreviate; keep every fact, number, and proper noun verbatim — in lowercase, not all caps."

The word "lowercase" matters. Without it, models write in ALL CAPS, and savings fall by 14–19 points: gemma 25.0%, qwen 29.8%, GLM 33.9%. With it, GLM-5.3-Flash saves 48.4%. Later the page gives gemma 40% and Qwen 49% "on their own meters." It does not reconcile these with the "25% to 49%" range it quotes elsewhere.

Placement:

• Compressing while the model is still composing its answer (answers written in cablese) scores 0.81 relative to plaintext, partly recoverable to 0.86.
• Compressing after the content is settled, into a record a machine will read, scores 1.08–1.09.

The recommended uses are scratchpads, memory stores and agent-to-agent handoffs. The advice is to expand back to plaintext only when a human reads the record. Expansion gives back about a third of the savings.

Evidence

• Benchmark size: 50 passages, about 24 questions each, roughly 1,300 total. 694 anchored questions per condition in the cross-family matrix.
• Grading: answers are short (a date, name or count). They are matched after normalisation at a fixed 0.8 overlap threshold, with no LLM judge.
• Baseline: answering from the full plaintext passage scores about 91%. The shortfall comes from strict grading, ambiguous questions and real misses. Every condition is reported as a ratio to its own plaintext control.
• Statistics: comparisons are paired and tested with McNemar's test.
• Cross-family matrix:
• Readers are other model families answering from GLM-5.3-Flash's records.
• Writers are other families, whose records GLM-5.3-Flash reads.
• All ratios fall between 0.99 and 1.10.
• In-family recovery is 1.09 (read-back p<0.0001). Decoded records give p=0.03.
• Readers were never shown the instruction. Four reader families got the compressed records cold.
• Why cablese can score above plaintext: the author's guess is that compression suppresses copy-the-record phrasing, which the strict grader penalises.
• Model versions: every GLM call is glm-5.3-flash on the z.ai coding plan. gpt-5-mini is also tested.

How to use it

The harness, frozen ledgers and a one-click notebook are at github.com/Travis42/telegraph-test. The author estimates "an afternoon of compute per new model" to measure:

• compression ratio
• cross-family legibility
• read-back cost and human-transcription cost, which are measured separately

The page gives no install or run commands.

Limits

• Mandatory-reasoning models lose money. gpt-5-mini cannot disable reasoning. Cablese makes it think about 3× harder, so its writes bill about double plain output despite the shorter text. Use writer models whose thinking you can control.
• No terseness control. The author has not run the same instruction without the historical framing (e.g. "write as tersely as possible"). The "latent register" explanation is called "an educated guess."
• No breakdown of the plaintext shortfall. The roughly 9% gap at baseline cannot be split into grading, ambiguity and real misses.
• Narrow coverage. Only four model families were tested, mostly small or fast variants, and only on short-answer factual QA. Tasks needing reasoning over the compressed record are not measured.
• Not the densest option. Emergent protocols compress further: BabelTele reports 27.9% of original length at 99.5% semantic fidelity, and other work cites 3–6×. The page argues cablese trades that density for stability and human auditability.