Solomet

> 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.

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.

I saw someone describe LLM hallucinations and inaccuracies as "compression artifacts." This thing has seen all the data that can possibly be fed to it and stored it in a much smaller size, so it's not crazy to think there will be compression artifacts. I have noticed that the quality of its answers seems to go down significantly when I am asking about a topic about which there isn't much information on google.

But it makes you wonder if future LLMs aren't going to suffer from the JPEG of a JPEG of a JPEG effect where it will be impossible to train an LLM off of new data that isn't generated by another LLM.