About Telegrapher
Telegrapher is an independent, unfunded, unincorporated research collaboration.
The work started with Telegraph English, a structured way of rewriting a prompt so that a model reads fewer tokens while the facts stay intact. Measuring what that rewrite kept and lost pulled the group toward a wider question: where do errors come from as a model generates, and why do some tasks degrade gracefully while others fall apart? That question became the error-accumulation trilogy.
The collaboration has no institution, funding or company behind it. Members work on the papers together, and the author list of each paper records who did what. Everything on this site is also published in machine-readable form for agents: see llms.txt.
Lines of work
The name
Telegrams were paid for by the word, so senders learned to keep only the words that carried meaning. Telegraph English, the project the group started with, applies the same economy to the prompts a model reads. The dots and dashes on this site are International Morse code, sent at twenty words a minute. The mark beside the name spells TE. Each paper page opens with a short telegram, the paper and its venue (TE NEURIPS), and each member's page with the initials of their name as a call sign. Sections are divided the way a telegrapher divides a message, with the prosign BT, and a post ends with AR, end of message. Hover any strip to see what it says.