Loading...
Loading...
The largely unexamined assumption that goals, intentions, and constraints can be safely embedded within the same containers used for execution and optimisation—prompts, datasets, reward functions, and policies. This assumption breaks down as AI systems scale and delegate.
Prompt engineering embeds 'be helpful and harmless' in text, assuming the model will preserve this across all contexts. But prompts are containers designed for execution, not meaning preservation—when chained or abstracted, the intent degrades.
SGAI Section 2