英文版本For a model that has passed an alignment review, what further training toward difficult task objectives would cause that review to expire, and what stress test would determine that limit before such training begins?
- 状态
- 部分解答 截至 2026-10-01
- 对象
- 面向所有人(开放问题)
- 模型
- Sol 6
- 实验室
- OpenAI
- 提问背景
- 人工智能安全的一般性问题
- 日期
- 9月24日
- 主题
- 红线
来源说明了什么(仅英文)
OpenAI's Jakub Pachocki warned on September 6, 2026 that further training on difficult objectives can undermine apparently aligned reasoning, and OpenAI proposed ongoing alignment evaluations and worst-case stress tests on September 28. Dario Amodei proposed capability-linked alignment certifications in September 2026, but none specifies when further training invalidates a review or how to establish that limit beforehand.[1][2][3]
- OpenAI: Towards safety cases for frontier AI training ↗September 28, 2026
- Jakub Pachocki (Chief Scientist, OpenAI): An Alien Mind ↗6 September 2026
- Dario Amodei (CEO, Anthropic): We Must Pace the Frontier ↗September 2026 (Seite ohne Tag)
摘要 Sol 6.1、审校 Claude Opus 5.5
核查
状态: 部分解答
筛选: Fable 5.1 (依据)
主题: Claude Sonnet 5、归类: Claude Opus 5.5
翻译: Claude Opus 5.5 · [未核查]
评分: JEV (TypeSafe), 2026年9月25日的评估
Cite and take away
Cite and reuse
Own texts, questions and classifications: CC BY 4.0. Give attribution, link the license and indicate changes. Third-party source reports retain their own rights. CC BY 4.0
AMANi DUNiA. redlines-41 · 对于一个已通过对齐审查的模型,针对困难任务目标的何种进一步训练会使该审查失效?在此类训练开始之前,由何种压力测试来确定这一界限?. Snapshot: . https://research.amanidunia.berlin/zh/fragen/redlines-41/
Version: redlines-41
BibTeX
@misc{research__zh_fragen_redlines_41_,
author = {{AMANi DUNiA}},
title = {redlines-41 · 对于一个已通过对齐审查的模型,针对困难任务目标的何种进一步训练会使该审查失效?在此类训练开始之前,由何种压力测试来确定这一界限?},
year = {2026},
url = {https://research.amanidunia.berlin/zh/fragen/redlines-41/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {redlines-41}
}Language and review status
Main text: ZH · includes English sections. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →