English · Русский
27 letters · English texts · Russian translations · sources and correspondence records
Company letters · Researcher letters · Replies and status · Using the archive
What justifies describing someone as “unreliable”, saying that an AI “understands”, or reading a sentence as evidence of a person’s beliefs?
This archive brings together letters from Kirill Shebetov / beforeword: 24 to researchers and three to OpenAI, Anthropic and xAI. The research letters ask specific questions about published arguments. The company letters propose changes to response rules and the reporting of model evaluations.
The shared proposal is to examine the written record, the interpretation added to it and the grounds for applying that interpretation separately. Learning a word does not by itself justify applying it to a particular person or acting on it. People should be able to question a description or a decision based on it without first accepting the description as who they are. An explanation can supply grounds; it does not turn the description into what it describes.
Each entry separates the sent text, its translation, an explanation of why it was written, and the correspondence record.
-
How does a description become a basis for a decision? A proposal to OpenAI: show the grounds for the description and the action, while preserving the ability to challenge both.
-
Same words, different conclusions? The letter to Amanda Askell proposes keeping a sentence unchanged and changing only its source label: human or AI.
-
What connects prediction with understanding? The letter to Ilya Sutskever asks for the criterion supporting that inference from a result.
Proposals about response rules, the grounds for decisions and how to evaluate those rules. Sending date reported by the author: 4 October 2026.
-
OpenAI — Show the grounds for applying a description and acting on it; preserve the ability to challenge both.
-
Anthropic — Show the grounds for attributing understanding, values or identity, and for decisions based on those descriptions.
-
xAI — Show the grounds for descriptions, decisions and conclusions drawn from evaluations of Grok’s responses.
Choose a question. Each letter includes an English transcription, a Russian translation, its purpose, links to the publications discussed, and the author’s screenshots.
-
Amanda Askell — What changes in a judgment of sincerity when only the source label changes?
-
Chris Olah — What supports each claim in an explanation of a model’s behavior?
-
Murray Shanahan — What supports different readings of “I believe this” from a human and an AI?
-
Yoshua Bengio — What supports admitting a written record as an observation for AI training?
-
Ilya Sutskever — What supports reading successful prediction as evidence of understanding?
-
Yann LeCun — What supports reading action-recognition results as progress toward understanding the world?
-
Demis Hassabis — What justifies reading model output as an account of its reasoning?
-
Stuart Russell — What supports reading a passage as evidence of a human preference?
-
David Chalmers — What supports reading a written sentence as a report about consciousness?
-
Geoffrey Hinton — What supports saying that two formulations “say the same thing”?
-
Douglas Hofstadter — What supports treating two instances of a pattern as the same self?
-
Thomas Metzinger — What supports reading a report as evidence of “pure awareness”?
-
Wojciech Zaremba — What supports reading a sentence as an explanation of perception?
-
Nick Bostrom — What supports the move from described behavior to an attribution of experienced pain?
-
Steven Pinker — What supports attributing the same idea to two people?
-
Robert Brandom — What supports treating a writer as committed to justifying a claim?
-
Michael Tomasello — What supports attributing mutual knowledge in a shared viewing task?
-
Andy Clark — What warrants calling a revision of text a correction of thought?
-
Dan Sperber — What is added to the wording when inferring what the writer meant?
-
Judea Pearl — What supports reading an experimental report as evidence for a causal assumption?
-
David Deutsch — What distinguishes changing the words from changing the explanation?
-
Timothy Williamson — What supports saying that a reader knows a claim is true?
-
Michael Levin — What supports describing tissue’s bioelectrical state as memory of the body’s shape?
-
Joscha Bach — What claim favors a consciousness interpretation over an alternative explanation?
xAI: a Safety Team acknowledgement is published. No substantive response to the proposal is published. The archive contains no replies to the other letters.
Statuses describe the available records. Sending, delivery, reading and a substantive reply are different events. A request to forward a letter does not establish that it was forwarded. No published reply does not mean agreement or rejection; naming a recipient does not imply their support for beforeword.
-
Read and compare. Every page begins with an EN/RU link. Context and source guidance are separate from the letter; sending records and a plain-text version are also available. No GitHub account is needed to read or download the archive.
-
Inspect the record. Research-letter transcriptions retain the wording of the supplied screenshots. The sent editions of company letters are preserved as separate files. Missing dates and hidden addresses remain unknown.
-
Cite a specific version. Use a GitHub permalink to the file at a chosen commit. The history lets readers compare revisions; changes to context or translation must not replace the sent wording.
-
Use the data. The JSON catalog lists all 27 records with links to texts and detailed data. Archive maintenance explains how the pages are built and checked.
beforeword reading procedure · Website · Download ZIP
Author: Kirill Shebetov / beforeword · [email protected]
The repository’s existing MIT license is retained. It does not grant rights in included third-party material, such as quotations or incoming correspondence.