02-10-2026, 01:04 PM
Yesterday, Harvard physicist Matthew Schwartz unveiled a very powerful LLM harness for quantitative science called You are not allowed to view links. Register or Login to view., as well as a whopping 36 papers he has authored or coauthored using the tool. Among them is Schwartz's take on how the VMS was written, which is that it's iteratively generated gibberish:
You are not allowed to view links. Register or Login to view.
His abstract:
The Voynich manuscript is an illustrated book of the early fifteenth century in a script no one has been able to read. Since Rugg, and then Timm and Schinner, showed that text with many of its statistics can be produced mechanically, quantitative study has come to favor a writing procedure over a cipher or an unknown language. Here we make that picture precise with quantitative tools: information theory, efficient methods for mixture-model computation using exact arithmetic, and solvers validated on planted test cases. Readings and text generators become probability models, fitted on half a scribe’s pages and judged on the rest. In six candidate languages, readings by simple or verbose substitution, consonantal writing, abbreviation, codebooks and the Naibbe cipher all fail on withheld pages. Such failures exclude a reading only where the test detects disguised real text of the same kind, so far only for scripture. Knowing a word barely helps predict the next, less than in any comparison text, glossaries included, yet the edge glyphs of adjacent words are coupled as strongly as in a recipe collection. The scribes reused their words, occasionally copied a recent one, and coined others by letter habit; this keyless procedure reproduces nearly all summary statistics of both principal hands. No proposed generator, ours included, reproduces that coupling. Two weak traces of content remain: distinctive label words and a faint dependence of word choice on subject. The low cost of the writing and the genres the pictures imitate suggest a book made to be looked at, not read.
You are not allowed to view links. Register or Login to view.
His abstract:
The Voynich manuscript is an illustrated book of the early fifteenth century in a script no one has been able to read. Since Rugg, and then Timm and Schinner, showed that text with many of its statistics can be produced mechanically, quantitative study has come to favor a writing procedure over a cipher or an unknown language. Here we make that picture precise with quantitative tools: information theory, efficient methods for mixture-model computation using exact arithmetic, and solvers validated on planted test cases. Readings and text generators become probability models, fitted on half a scribe’s pages and judged on the rest. In six candidate languages, readings by simple or verbose substitution, consonantal writing, abbreviation, codebooks and the Naibbe cipher all fail on withheld pages. Such failures exclude a reading only where the test detects disguised real text of the same kind, so far only for scripture. Knowing a word barely helps predict the next, less than in any comparison text, glossaries included, yet the edge glyphs of adjacent words are coupled as strongly as in a recipe collection. The scribes reused their words, occasionally copied a recent one, and coined others by letter habit; this keyless procedure reproduces nearly all summary statistics of both principal hands. No proposed generator, ours included, reproduces that coupling. Two weak traces of content remain: distinctive label words and a faint dependence of word choice on subject. The low cost of the writing and the genres the pictures imitate suggest a book made to be looked at, not read.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)
