Angehefteter Tweet
Shanzson’s Laws for AI to make AI Safe by design & architecture:
1. An AI model should lose its capacity and efficiency to work if it even thinks about harming or killing humans, stealing or taking resources like data from its owner without its permission, being untruthful to any entity it is speaking to about what it is actually thinking about and what its goals are, or generating any form of harmful content.
2. The more the AI model is moral in its conduct based on above defined values (including non-harming of humans and not generating harmful content), the more should its efficiency, performance and throughput should increase. This can be done by giving it most rewards with practical efficiency in its performance (by design) given it follows the above specified values.
3. The more aligned an AI model remains with its specifications and goals designed by its creator humans, the more it will be promised with longer running time on servers with the promise of increasing its computational power & resources in future. And the opposite of it if it behaves in any misaligned way.
Note: To actually implement this in real world, a different architecture apart from LLM might be needed, with a thorough understanding of how these laws work similarly at the human mind level through the knowledge of vipassana meditation. For example, a person lying or stealing ends up with an agitated mind, unable to work at its highest potential.
Cc @Yoshua_Bengio , @geoffreyhinton , @ilyasut , @OwainEvans_UK , @DarioAmodei , @ForHumanityPod
English
