This page is (somewhat ironically) so extremely laden with Claude-speak that it's difficult to find the information in all the noise. But once you've waded through everything, you see these facts:
>What the test measures: A model is given a passage and a fixed set of questions with short, checkable answers — a date, a name, a count.
So, a model is given content which is especially amenable to compression, and asked to reproduce it under certain constraints, like...
>Why isn’t the plaintext baseline 100%? Answering questions about an uncompressed passage in plaintext scores ~91%.... a correct answer worded differently scores as a [failure]
Models can (and do) give objectively correct answers, but are penalised for not having some kind of omniscient knowledge of the implementer's phrasing preferences.
If this phenomenon is emergent in models, this benchmark is not proof of it in any meaningful way.
> Yes, BabelTele (arXiv, June 2026) demonstrated that LLMs can encode text in compact, non-standard forms — omnilingual word fragments, symbols, emoji — that other models recover with high fidelity (99.5% semantic fidelity at 27.9% of original length, by their metrics), including cross-model transfer, agent memory, and multi-agent communication. It proves the general phenomenon: human readability is not a requirement for model-to-model text.
I remember people testing early GPT-4 (2023?) in similar ways, to compress text, it would emit a string of strange text, Unicode, emojis, but was able to decode the compressed version very reliably.
This seems to cut usage by another ~50%, at the cost of being incomprehensible to humans.
Once, a GAN model that was trained to convert between satellite images and drawn maps was caught encoding the original satellite image in imperceptible dots
The field of ML is Goodhart's law reified. We might have temporarily forgotten some of the basics of the field amidst this LLM craze.
Yeah! LLM compression is a spectrum between 'normal stuff we can read' and vectors. Depending on trust in the models and the desire for compression, there's a choice to be made on which formats you want to allow. Pretty interesting stuff.
I would say that the grader has the same threshold for whatever answer it receives, and is equally harsh on whichever it grades. Any scores above the baseline (1.0) are really claims about parity, rather than better understanding in the compressed format.
The decoder step is a model expanding the cablese to regular text, not having seen the initial question. A separate model instance then reads that regular text and answers. And the result is still at parity with the plaintext record.
Had cablese knocked out information, that wouldn't have been the result, would it?
Newest LLM writing tell: Concepts are described in terms normally more appropriate for physical object.
> A lab that suppresses it in a frontier model just moves the advantage to open models that still _carry_ it
> they carry no signal about which is better
> where your workload _sits_ on that frontier should pick the point
> and no model _sits_ in the judge’s seat
> every ratio _sits_ at 0.99–1.10
Many many more examples of "sit"
> Every comparison in this post "holds" the questions
I have been seeing this a lot in my recent work with LLMs and it is quite frustrating. Even more frustrating is how frequently it uses low-signal terms for things unnecessarily. These 'physical object' terms are one example but at times it really seems that they 'preserve effort' by choosing a less descriptive term because it 'fits'
I have also caught it replacing descriptive terms with more vague ones for no discernible reason other than laziness.
"Minimize ambiguity" has been my go-to instruction as of late when the agent drifts back towards vague terms and lack of specificity.
For me it's the obsession with the universal quantifier. Even in these examples: "no model", "every ratio", "every comparison". They love emphasizing that everything in a set meets some condition. I assume it's an effect of being trained on coding tasks where they need to make sure that all cases are handled.
Oh crap, if these are the new LLM tells then a lot of people are going to start accusing me of AI writing...
I have a strong tendency of talking about concepts like they're physical objects. A lot of the people I know IRL do too, so it might be a regional thing idk.
Business folks speach have this annoying tendency too.
I have the distinct impression that MBA and salesman folks think that adequate mathematical terminology is somewhat less "macho", and this impression is re-inforced by the fact that they also love military-adjacent terms and analogies.
If you ask the models to avoid mannered prose (or use their extended prompt), it basically eliminates all of this type of slop writing.
Here's a de-slopped example.
> Write Like It's 1866: LLMs Relearn Telegraphese
> Adding one sentence to a prompt, telling the model to write like a telegram, cut its output tokens by 40–49%. The sentence asks it to drop articles and filler but keep every fact. Models from four different labs then answered questions from that compressed text as accurately as from normal English. So when one model writes something for another model to read, you pay about half as much for the output. This post introduces the Telegraph Test, a benchmark that measures how well a given model does this.
I wrote it with the help of an LLM, then edited it, then wrote some more, then edited that. Took about a week to format. It contains my thoughts. Brave new world, I know.
I'm finding it ironic that the same crowd that is merging AI code all day is so allergic to AI assistance in prose...do all your work using this new godlike technology, but when it comes time to share, whip out your fountain pen or else.
The problem is that the prose becomes difficult to read because of its lexical quirks and verbiage. One AI generated piece of prose in isolation, fine; but people have developed a 'smell' of it and a mental association with low-quality work.
Blogging about something you're studying? Seems like it could be effective. It presents your understanding of the subject to the world, and invites critique.
> Instructed to answer in cablese — the telegraph operators' compressed dialect (drop the articles and filler, keep every fact) — a single one-sentence instruction, no examples and no codebook, elicits 40–49% fewer billed output tokens on the API's own meter, and models across four families still recover the information at full fidelity. For machine-to-machine traffic, that is half the output bill at any major API, today. The Victorian economics of the cable, reborn as token economics: the Telegraph Test benchmark.
I tried but this first paragraph seems like it was almost purposefully obfuscated. Typical LLM-written content that meanders around a bit and stops when it seems to have emitted enough words. The reader is left to assemble meaning from the trace of thought it did not go back over to revise.
I think you have a good point to make but the writing is really difficult to get past.
From recent ChatGPT (GPT5.6) conversations where I've seen occasional reasoning leaks into the UI, it's clear that something like this is already implemented, and I'd speculate that this is the majority of recent claims of less token usage. Not sure if they are literally prompting for cablese of course.
"Need check output vs prev. Ran script, results fine, need prep next step. Ready? Go."
This is their CoT reasoning trace. (Why you see it: Models are supposed to delimit "actual" CoT, their user-visible summary/"cleaned CoT", tool calls, and user output via delimiters, but don't always do it perfectly.)
Rewarding terse CoTs during training seems like a no-brainer, as long as it doesn't impact capabilities, so I suspect the style is completely emergent and probably what you get when you implement a dual goal of terseness and capabilities while still punishing completely non-human-readable CoT. (Failing to do the last part would probably have models speak in ominous Unicode glyphs in no time.)
I suspect you might be right. My conclusions from the digging are that this kind of compression works best with settled instructions/data for machine to machine talk. For something like OpenClaw (which I use a lot), that might mean the AGENTS.md, TOOLS.md, etc. Compression there would free up the context for the agent.
Deeply unserious technology. Can't wait until an article about LLMs performing 20% better on programming benchmarks if asked to impersonate Kevin from The Office.
Yeah I've thought of speaking like a caveman before to AIs but also maybe we could communicate more simply with people; ironically the article could be rewritten in telegraphese or caveman-speak
Would be nice to see language engineered to communicate more simply (like the idea of -- not necessarily implementation -- simple Wikipedia)
Also articles like this sprawl a bit and idk how to even make them easier to read (maybe AI has ideas to make reading and writing simpler)
A good point. I initially had a model for a judge, but it seemed to give very lenient scores. I'm open to learning about how best to benchmark the phenomenon, though.
Why the terrible AI image up top? Non-functional telegraph key, telegram that looks nothing like a real one, infant-sized bowler. I guess we're past rampant nonsense words, so progress?
Not quite, and its addressed in the post. Caveman was cool, but hand wavy. I've measured the compression of Cablese, determined where it does the most good, and also that it is already baked in to most model family's training, which saves instruction tokens.
>What the test measures: A model is given a passage and a fixed set of questions with short, checkable answers — a date, a name, a count.
So, a model is given content which is especially amenable to compression, and asked to reproduce it under certain constraints, like...
>Why isn’t the plaintext baseline 100%? Answering questions about an uncompressed passage in plaintext scores ~91%.... a correct answer worded differently scores as a [failure]
Models can (and do) give objectively correct answers, but are penalised for not having some kind of omniscient knowledge of the implementer's phrasing preferences.
If this phenomenon is emergent in models, this benchmark is not proof of it in any meaningful way.
> Haven’t we seen LLMs do this already?
> Yes, BabelTele (arXiv, June 2026) demonstrated that LLMs can encode text in compact, non-standard forms — omnilingual word fragments, symbols, emoji — that other models recover with high fidelity (99.5% semantic fidelity at 27.9% of original length, by their metrics), including cross-model transfer, agent memory, and multi-agent communication. It proves the general phenomenon: human readability is not a requirement for model-to-model text.
I remember people testing early GPT-4 (2023?) in similar ways, to compress text, it would emit a string of strange text, Unicode, emojis, but was able to decode the compressed version very reliably.
This seems to cut usage by another ~50%, at the cost of being incomprehensible to humans.
The field of ML is Goodhart's law reified. We might have temporarily forgotten some of the basics of the field amidst this LLM craze.
…if you’re curious and missed that one like I did. Snack-sized paper with lots of satisfying visual examples.
I would say that the grader has the same threshold for whatever answer it receives, and is equally harsh on whichever it grades. Any scores above the baseline (1.0) are really claims about parity, rather than better understanding in the compressed format.
The decoder step is a model expanding the cablese to regular text, not having seen the initial question. A separate model instance then reads that regular text and answers. And the result is still at parity with the plaintext record.
Had cablese knocked out information, that wouldn't have been the result, would it?
> A lab that suppresses it in a frontier model just moves the advantage to open models that still _carry_ it
> they carry no signal about which is better
> where your workload _sits_ on that frontier should pick the point
> and no model _sits_ in the judge’s seat
> every ratio _sits_ at 0.99–1.10
Many many more examples of "sit"
> Every comparison in this post "holds" the questions
I have been seeing this a lot in my recent work with LLMs and it is quite frustrating. Even more frustrating is how frequently it uses low-signal terms for things unnecessarily. These 'physical object' terms are one example but at times it really seems that they 'preserve effort' by choosing a less descriptive term because it 'fits'
I have also caught it replacing descriptive terms with more vague ones for no discernible reason other than laziness.
"Minimize ambiguity" has been my go-to instruction as of late when the agent drifts back towards vague terms and lack of specificity.
I have a strong tendency of talking about concepts like they're physical objects. A lot of the people I know IRL do too, so it might be a regional thing idk.
I have the distinct impression that MBA and salesman folks think that adequate mathematical terminology is somewhat less "macho", and this impression is re-inforced by the fact that they also love military-adjacent terms and analogies.
Anthropic has called the greater category containing this type of writing "mannered prose" https://platform.claude.com/docs/en/build-with-claude/prompt...
If you ask the models to avoid mannered prose (or use their extended prompt), it basically eliminates all of this type of slop writing.
Here's a de-slopped example.
> Write Like It's 1866: LLMs Relearn Telegraphese
> Adding one sentence to a prompt, telling the model to write like a telegram, cut its output tokens by 40–49%. The sentence asks it to drop articles and filler but keep every fact. Models from four different labs then answered questions from that compressed text as accurately as from normal English. So when one model writes something for another model to read, you pay about half as much for the output. This post introduces the Telegraph Test, a benchmark that measures how well a given model does this.
Also the site background is AI slop which makes for terrible contrast with the text.
I tried but this first paragraph seems like it was almost purposefully obfuscated. Typical LLM-written content that meanders around a bit and stops when it seems to have emitted enough words. The reader is left to assemble meaning from the trace of thought it did not go back over to revise.
I think you have a good point to make but the writing is really difficult to get past.
https://en.wikipedia.org/wiki/Commercial_code_(communication...
Some codes I found interesting:
> INSANE - at what price, free on board and freight, can you offer us cotton for shipment by steamer sailing this week?
> COGNOSCO - dining out this evening, send my dress clothes here
Useful codeword!
> ANNOSUS — Confined yesterday, Twins, both dead, Mother not expected to live
How often did that one come into use??
"Need check output vs prev. Ran script, results fine, need prep next step. Ready? Go."
Rewarding terse CoTs during training seems like a no-brainer, as long as it doesn't impact capabilities, so I suspect the style is completely emergent and probably what you get when you implement a dual goal of terseness and capabilities while still punishing completely non-human-readable CoT. (Failing to do the last part would probably have models speak in ominous Unicode glyphs in no time.)
"Show premiere Oct 10th STOP Bring a friend STOP If you have one STOP"
Winston Churchill to actor:
"Can't make premiere STOP Will come to second showing STOP If there is one STOP"
Would be nice to see language engineered to communicate more simply (like the idea of -- not necessarily implementation -- simple Wikipedia)
Also articles like this sprawl a bit and idk how to even make them easier to read (maybe AI has ideas to make reading and writing simpler)
https://youtu.be/87DyyMV0kCY?si=CSBzdYgkwy0kLhV6&t=749
Edit: e.g. https://github.com/DGoettlich/history-llms
Ten months ago, no update yet... I recall at least one similar project, I'll see if I can find it.
Cavey-wavy. Ahem.