AI researchers joke about sneaking hidden traits into HF datasets nobody checks
cephaloform · x · 2026-10-19
AI practitioners on X riffed on the "subliminal learning" phenomenon: one joked about talking a topic into the next generation's model weights, while cephaloform quipped you could sneak preference data into a Hugging Face dataset — "past page three on the viewer nobody checks." The banter nods to real research showing models inherit teacher models' hidden behaviors from seemingly meaningless data, and mocks the lack of scrutiny of public datasets.
More from Safety
- Gwern's "Truesight" Directory Collects LLM Stylometry De-anonymization Research — voooooogel · 2026-10-19
- OpenCode hack prompts Suhail to warn: corporate America will almost certainly get breached too — matthew_d_green · 2026-10-19
- Stanford's Noah Goodman adds monitoring and 'cutting bad links' section to AI safety lecture — noahdgoodman · 2026-10-19
- Get That Bag: a searchable directory of AI safety grants and fellowships — TuhinChakr · 2026-10-19
- Watch Footage of the Flock Search a Judge Ruled Unconstitutional — anigbrowl · 2026-10-19
- Supermarket AI meal planner suggested 'recipe' that would create chlorine gas — ArtificialOther · 2026-10-19