Published letter · AI and model evaluation
What supports reading successful prediction as evidence of understanding?
Sent · no reply publishedContext by beforeword · separate from the sent text
Why this letter was written
What prompted the letter
The letter was prompted by a quoted statement linking successful prediction of the next unit of text to understanding the reality that led to its production.
Why this matters to readers
Readers need to be able to distinguish the reported prediction result from the broader claim that the model understands.
What the letter asks
The letter asks what criterion supports moving from the first to the second, and why it applies in this argument. It also proposes examining explanations of how training text was produced as written accounts in their own right.
Sent letter · English transcription
The letter
Transcribed from the supplied screenshots. Screen line breaks are joined into paragraphs; wording and punctuation are retained.
Please forward this research note to Ilya Sutskever.
Dear Ilya,
The published transcript of your March 2023 conversation with Dwarkesh Patel includes:
“Predicting the next token well means that you understand the underlying reality that led to the creation of that token.”
beforeword reads “means that” here as a proposed inference from predictive success to an attribution of understanding.
The proposal is to keep the reported prediction result, its selected interpretation, and the grounds offered for that interpretation separately inspectable. Reading a result as understanding does not, by itself, establish the attribution.
The explanation of how the training text was produced also remains a written account whose interpretation and stated grounds require examination. This applies equally to human self-descriptions and to this letter.
The reading procedure and worked examples:
https://beforeword.xyz/model/en/
Please indicate the criterion for accepting predictive success as grounds for attributing understanding, and the grounds for applying it in this argument.
Kirill Shebetov
beforeword
+
Publications discussed in the letter
These links are provided for reading in context. They are separate from the sent letter; beforeword’s commentary is not a reply from the recipient.
Ilya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligence
Dwarkesh Podcast, 27 March 2023. Transcript section: What’s after generative models?.
The host’s transcript contains the next-token-prediction statement quoted in the letter.
Sending record and files
Sent to comms@ssi.inc with a request to forward it to Ilya Sutskever. The screenshots do not establish that it was forwarded.
The author supplied screenshots from the Sent folder. They show the message and the available sending details. These materials do not record the recipient reading it.
View screenshots · 2

Screenshot 1 · Open full size

Screenshot 2 · Open full size
Each TXT file begins with publication details, followed by the separately marked letter text. These are screenshot transcriptions and translations, not original email files. The purpose explanation and opening question are editorial summaries by beforeword.
Correspondence record
Sending shown in the author’s materials
The header shows “Friday”; the calendar date is not displayed.
Receipt by the intended recipient and a substantive reply are not separately recorded in the published materials.
Added to the public collection
The English transcription, Russian translation and screenshots are presented separately. No replies have been published yet.
The absence of a published reply does not establish agreement, rejection or an inability to object.
