kaal:claim:4855607-017

Balancing helpfulness against harmlessness is an inherent tension in Safe RLHF rather than a tuning problem that can be resolved once.

Source quote, verbatim

Balancing the dual objectives of helpfulness and harmlessness remains an inherent tension in Safe RLHF.

From

Wulf A. Kaal, How AI Models are Optimized Through Web3 Governance (2024), Model Overview: Reinforcement Learning Through Human Feedback (RLHF), p. 30
https://ssrn.com/abstract=4855607 · source PDF

Cite as

Wulf A. Kaal, How AI Models are Optimized Through Web3 Governance (2024). SSRN: https://ssrn.com/abstract=4855607

Holds when

Safe RLHF designs that optimize helpfulness and harmlessness jointly

Classification

failuresupport: evidencedfailure: Helpfulness harmlessness tensionfamily: ai-oversight-and-alignment-gapai-and-agents

Related claims

extended_by kaal:claim:6607458-030 The AI alignment problem is, within this framework, a foundational policy problem of Computative Economics rat...

Verify

The quote above is an exact substring of the source PDF, whose sha256 is eb0b3e62374b45a8fa888c6bde9725e606bcb46cf4b5e74a6e851d9f25099113. Extraction method: pdf-text-layer.
Attestation record: colloquium/attestations/cde141bf7ae6f5c9...json
Verify the binding yourself: curl -s https://wulfkaal.github.io/claims/4855607-017.md | sha256sum