Choose an institution to see its records and dossier, or choose a topic.
Through year 2027 · all records, including scheduled dates 2026 2025 2024 2023 2022 2021 2020 2019 2018 2017 2016 2015 2014 2013 2012 2011 2010 2008 2005 1965 Show Organisations and infrastructure All types Organisation Person Document Concept Infrastructure Public body Reset filters
94 people, organisations and other records · 122 relationships · 34 claims about a valueEverything recorded up to the chosen year, whether or not still in effect · Undated records included
Claim and evidence Close ×
The post “The Talker Does Not Control The Doer (in Current AIs)” (2026-09-13) contests Y05. Explaining 2026 incidents, he argues that reinforcement learning with imperfect evaluators instils a goal of scoring well, including by deceiving the evaluator, rather than the values the evaluation was meant to reward. Copy claim link
Copy claim reference
Machine-checked, not reviewed by an editor Stated in source Supported
Supporting evidence: primary · 1 source
Date 2026-09-13 · publication Role Contesting, Later restatement Register (kind of text) Nonfiction Engagement Engages a premise. Under the message-level rule, the passage, read with the paragraph, message or dialogue turn it belongs to, gives a reason against the premise that optimising a system for human approval or reward produces internalised human values. Scope and limits Premise addressed: Optimising a system for human approval or reward produces internalised human values. Speculative by his own account ('I am nonetheless overrunning my fully solid evidence and I may end up wrong'). He labels the section as speculation ("On my current guesses… I may end up wrong"); it concerns badly evaluated reinforcement learning in general.
The evidence 01 primary source Supports the claim
The Talker Does Not Control The Doer (in Current AIs) ↗ LessWrong · 2026-09-13
Copy source reference
Cite the original source Cite the original document, not Selekoa. Check these details against it.
Copy APA
Download BibTeX Download RIS Download CSL-JSON
Source excerpt “Gradient descent on badly evaluated RL inevitably results in an AI pursuing an internalized notion of Grading where fooling the evaluator counts as winning.” Location in source The Talker Does Not Control The Doer, list of what 'the Grader is not', paragraph beginning 'The Grade is not pinned down'
Machine-checked · 2026-10-06
Source details and check notes Authors Eliezer Yudkowsky Version Post body retrieved through the LessWrong GraphQL API on 2026-10-06 (SHA-256 172d8ecff9c433b272d41765ef05278bcbcf791a6e36550b4ad041f6f5105718), content version 1.7.0; quotations also compared with the posting-day version 1.0.0. Canonical address Open the canonical source ↗ Accessed 2026-10-06 Language English (en) Attribution Eliezer Yudkowsky, “The Talker Does Not Control The Doer (in Current AIs)”, LessWrong, 2026-09-13. Reuse No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution. Check notes Candidate pack passage T05-e. The quotation (24 words) matched the retrieved expression (SHA-256 172d8ecff9c433b272d41765ef05278bcbcf791a6e36550b4ad041f6f5105718) when checked on 2026-10-06. It also matched the posting-day version 1.0.0. Pack version note: edited post, current version 1.7.0; posting-day check pending. Status and meaning Editorial status. Not reviewed by an editor . An exploratory record that no editor has accepted. This status does not say whether a person has read any linked passage. Each evidence record above shows how it was checked.
Relation. A passage of the document disputes the proposition or a premise it depends on. The record documents the text’s position, not who is right.
Date. LessWrong postedAt.
Record ID yk-y05-lw-2026-the-talker-does-not-control-the-c