Proposal: solve alignment of the current generation before training the next
Start here
A proposal on the AI Alignment Forum suggests that labs must adequately resolve major alignment issues in their current-generation models before training the next generation. The proposal requires multiple independent alignment organizations to review and red-team the latest models at each training checkpoint, and to monitor them in restricted deployments with trusted users; if major alignment issues arise, the model should be taken down and a root-cause analysis conducted. The author states this is only a heuristic minimum standard, and questions such as how to define a "generation of models" and when to stop model development remain unresolved.
Generated by AI from reports · Updated 9 hr ago
Progress timeline
- AI Alignment ForumAI Alignment Forum 提案：先充分解决当前代模型的对齐问题，再训练下一代
A proposal on the AI Alignment Forum suggests that labs must adequately resolve major alignment issues in their current-generation models before training the next generation. The proposal requires multiple independent alignment organizations to review and red-team the latest models at each training checkpoint, and to monitor them in restricted deployments with trusted users; if major alignment issues arise, the model should be taken down and a root-cause analysis conducted. The author states this is only a heuristic minimum standard, and questions such as how to define a "generation of models" and when to stop model development remain unresolved.
Related discussion
Attention trend
New sources per day
- Oct 12New1 sources (1 outlets,0 accounts)
- Oct 13New0 sources (0 outlets,0 accounts)