Training on prompts very similar to the evaluation can reduce blackmail rate significantly, but it did not improve performance on our held-out automated alignment assessment.
这是提纲第8题「刷题不泛化」的原文出处，但原文比转述更微妙：贴近评测的训练确实显著降低了目标指标（blackmail rate），只是没能迁移到留出集。也就是说「刷题」对被刷的那门考试是有效的，失效的是泛化。提纲写成「连机器都因为刷题而无法泛化」会让人误以为刷题连本科目都提不动——恰恰相反，这才是应试教育难以证伪的原因。
Parametric Polymorphism
"Inheritance" is run time poluymorphism.
"Generic" is parametric poluymorphism. poluymorphism at complie time.
for example, comments and identifiers
Some better illustrated examples can be found in UBCx: SoftConst2x - Software Construction: Object Oriented Design's course lecture on Coupling.
If one object is part of another object, then we use a diamond at the start of the arrow (next to the containing object), and a normal arrow at the end.
Another way of thinking of this is, if the original owner (source) object and the owned (target) object share the same life cycle -- that is, the owned exists only when the owner does -- we say that the owner aggregates owned object(s). They share a whole-part relationship.
What I did like very much about the video, was when the instructor pointed out that there's a small fallacy: aggregation, in OOD, does not really imply that owned object(s) must be a list.