# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/write-like-its-1866-llms-relearn-telegraphese-49991580 # Write Like It's 1866: LLMs Relearn Telegraphese 54 points · 37 comments [Full discussion]() [Read original]() Category: [Research & Evaluation]() ## Skept-o-meter & Hotness Skept\-o\-meter: High\. Estimated from 5 comments\. 2 comments for the summary\. Peak rank: \#10 Time in Top 10: 2\.0 hours Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\. 57 recorded rank observations from 2026\-10\-07T15:01:57\.541362\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\. Hotness — latest 57 recorded Hacksnap ranks: 2026\-10\-07T15:01:57\.541362\+00:00: rank \#14 2026\-10\-07T16:01:28\.474022\+00:00: rank \#10 2026\-10\-07T17:01:15\.036908\+00:00: rank \#11 2026\-10\-07T18:00:53\.300079\+00:00: rank \#10 2026\-10\-07T19:01:49\.456657\+00:00: rank \#12 2026\-10\-07T20:01:59\.032487\+00:00: rank \#12 2026\-10\-07T21:05:36\.655392\+00:00: rank \#14 2026\-10\-07T22:02:27\.105241\+00:00: rank \#15 2026\-10\-07T23:01:17\.452894\+00:00: rank \#14 2026\-10\-08T08:03:55\.083808\+00:00: rank \#13 2026\-10\-08T09:03:58\.68898\+00:00: rank \#11 2026\-10\-08T10:04:05\.364235\+00:00: rank \#11 2026\-10\-08T11:02:04\.410307\+00:00: rank \#11 2026\-10\-08T12:02:36\.39618\+00:00: rank \#11 2026\-10\-08T13:03:09\.206401\+00:00: rank \#11 2026\-10\-08T14:03:30\.113384\+00:00: rank \#12 2026\-10\-08T15:02:54\.444758\+00:00: rank \#260 2026\-10\-08T16:03:13\.067807\+00:00: rank \#260 2026\-10\-08T17:03:47\.92465\+00:00: rank \#260 2026\-10\-08T18:01:50\.360431\+00:00: rank \#260 2026\-10\-08T19:02:55\.719274\+00:00: rank \#261 2026\-10\-08T20:01:36\.359022\+00:00: rank \#261 2026\-10\-08T21:03:34\.000719\+00:00: rank \#263 2026\-10\-08T22:01:13\.459732\+00:00: rank \#263 2026\-10\-08T23:01:25\.318392\+00:00: rank \#263 2026\-10\-09T08:01:53\.314021\+00:00: rank \#264 2026\-10\-09T09:02:59\.93426\+00:00: rank \#265 2026\-10\-09T10:02:22\.652527\+00:00: rank \#265 2026\-10\-09T11:02:04\.807282\+00:00: rank \#265 2026\-10\-09T12:02:16\.769295\+00:00: rank \#265 2026\-10\-09T13:02:30\.35034\+00:00: rank \#267 2026\-10\-09T14:01:34\.215134\+00:00: rank \#267 2026\-10\-09T15:01:27\.564392\+00:00: rank \#267 2026\-10\-09T16:02:06\.561056\+00:00: rank \#268 2026\-10\-09T17:03:06\.673648\+00:00: rank \#270 2026\-10\-09T18:01:02\.851052\+00:00: rank \#270 2026\-10\-09T19:01:11\.250291\+00:00: rank \#271 2026\-10\-09T20:01:32\.91907\+00:00: rank \#271 2026\-10\-09T21:02:08\.376144\+00:00: rank \#273 2026\-10\-09T22:01:19\.253075\+00:00: rank \#275 2026\-10\-09T23:01:31\.990912\+00:00: rank \#275 2026\-10\-10T08:01:00\.593376\+00:00: rank \#274 2026\-10\-10T09:01:45\.527975\+00:00: rank \#277 2026\-10\-10T10:00:53\.799353\+00:00: rank \#277 2026\-10\-10T11:00:35\.65851\+00:00: rank \#276 2026\-10\-10T12:01:11\.169647\+00:00: rank \#276 2026\-10\-10T13:00:54\.008377\+00:00: rank \#276 2026\-10\-10T14:01:43\.208766\+00:00: rank \#278 2026\-10\-10T15:01:24\.825188\+00:00: rank \#278 2026\-10\-10T16:01:10\.213328\+00:00: rank \#277 2026\-10\-10T17:00:48\.84189\+00:00: rank \#276 2026\-10\-10T18:00:57\.55132\+00:00: rank \#276 2026\-10\-10T19:00:52\.379463\+00:00: rank \#276 2026\-10\-10T20:00:21\.993912\+00:00: rank \#276 2026\-10\-10T21:01:10\.657738\+00:00: rank \#278 2026\-10\-10T22:00:50\.433569\+00:00: rank \#279 2026\-10\-10T23:01:00\.95027\+00:00: rank \#280 A benchmark finds telegraph\-style 'cablese' cuts LLM output tokens 25–49% while machine readers recover facts at parity or better, though the latent\-register claim lacks a strict generic\-terseness control\. ## The brief An independent benchmark argues that LLMs can write in 'cablese'—a lowercase, telegraph\-style register that drops articles and filler while preserving facts—to cut output tokens by 25–49% without hurting machine comprehension\. Across 50 passages and roughly 1,300 questions, foreign\-model readers answered from compressed records at 0\.99–1\.10 recovery ratios versus plaintext, while compressing mid\-answer imposed a register tax (0\.81)\. The author traces the register to historical telegraphy and says it is latent in training data, but acknowledges no strict control for generic terseness\. - The Telegraph Test uses 50 passages, \~1,300 anchored questions, and a deterministic 0\.8\-overlap grader; scores are ratios against each condition's own plaintext control\. - Cross\-family readers recovered facts from GLM\-5\.3\-Flash's cablese records at 0\.99–1\.10; writers saved 40\.4% (gemma), 48\.9% (Qwen), and 17\.7% (gpt\-5\-mini, whose mandatory reasoning doubled its bill)\. - The condition ladder shows answering in cablese drops to 0\.81, decoded answers reach 0\.86, but cablese records read back at 1\.09 and decoded records at 1\.08\. - Compression is free only after content is settled and the reader is a machine; scratchpads, memory stores, and agent handoffs are the proposed use cases\. - The historical analogy is 1866 transatlantic cable at $10 per word; codebook substitution saved only \~10%, while cablese emerged as a readable, auditable convention\. - The author concedes the latent\-register explanation is an educated guess without a strict 'write as tersely as possible' control, and contrasts cablese's stability with more aggressive but illegible emergent agent protocols\. ## Discussion themes Analyzed: 2026\-10\-07T17:00:44\.972002\+00:00 Analysis sample: Based on 5 of 5 usable stored comments. Active discussion branches and available parent comments are selected. This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is. ### LLM writing style: physical\-object metaphors and vague terms A commenter complains that LLM output uses physical\-object terms such as “sit,” “carry,” “holds,” and “fits” for abstract concepts, and replaces descriptive terms with vague low\-signal wording\. They describe this as frustrating and use “Minimize ambiguity” as an instruction when the model drifts toward vagueness\. Sources: [Comment 49992293]() ### Benchmark scoring penalizes paraphrase A commenter argues the benchmark’s plaintext baseline is only about 91% because correct answers worded differently are marked as failures, so models are penalized for not matching the implementer’s phrasing\. They conclude the benchmark does not meaningfully prove the claimed emergent phenomenon\. Sources: [Comment 49992460]() ### Prior and anecdotal evidence of compressed model\-to\-model encodings A reply cites BabelTele and early GPT\-4 experiments showing models can encode text in compact non\-standard forms such as word fragments, symbols, and emoji, then recover it with high fidelity\. Another commenter reports reasoning leaks in ChatGPT that look like compressed “cablese” and speculates this explains recent token\-usage reductions\. Sources: [Comment 49993128]() · [Comment 49992043]() ### Token\-usage savings from compressed encodings The cited prior work is said to cut usage by about 50%, and another commenter links similar compressed output to lower token usage\. This frames compressed encodings as an efficiency and operating\-cost tradeoff\. Sources: [Comment 49993128]() · [Comment 49992043]() ### Human readability tradeoff of compressed encodings The compressed model\-to\-model forms are described as incomprehensible to humans, which is noted as the cost of the approach\. This raises a constraint for human oversight or readability\. Sources: [Comment 49993128]() ### Use cases for compressed model\-to\-model encodings A reply cites prior work demonstrating compact non\-standard encodings for cross\-model transfer, agent memory, and multi\-agent communication\. These are presented as settings where human\-readable text is not required\. Sources: [Comment 49993128]() ## Sources & coverage AI-generated summary · 2026\-10\-07T15:01:57\.063391\+00:00 Based on 2 of 2 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened. Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.