Gwern Branwen - Footnote 32
1 evidence check
Last checked: 4/3/2026
The claim mentions Gwern analyzing datasets like The Pile for model behavior impacts, which is supported by the source, but it overstates his role as an 'archivist'. The source mentions he explored the contents of datasets like The Pile, but doesn't explicitly call him an archivist. The claim mentions Everitt 2018's work on Bayesian history-based RL (superseded by Everitt & Hutter 2019), which is not mentioned in the source. The claim mentions addressing value alignment to avoid Goodhart's law failures like reward hacking, which is not explicitly mentioned in the source.
Evidence — 1 source, 1 check
Note: The claim mentions Gwern analyzing datasets like The Pile for model behavior impacts, which is supported by the source, but it overstates his role as an 'archivist'. The source mentions he explored the contents of datasets like The Pile, but doesn't explicitly call him an archivist. The claim mentions Everitt 2018's work on Bayesian history-based RL (superseded by Everitt & Hutter 2019), which is not mentioned in the source. The claim mentions addressing value alignment to avoid Goodhart's law failures like reward hacking, which is not explicitly mentioned in the source.
Debug info
Record type: citation
Record ID: page:gwern:fn32