Open-Endedness is Essential for Artificial Superhuman Intelligence
Date: 1st July 2025
arXiv Link
Key Points
- Defines Open-Endedness: a system which can generate novel and learnable artefacts.
- This is dependent on the observer. E.g. what is learnable to a human might not be learnable to a mouse
- This depends on cognitive abilities like memory, which influences how learnable a new artefact is.
- Time-horizon dependent: tasks will often plateau in learnability due to limitations of the environment.
- If they can’t remember previous artefacts, they may struggle to understand this new one.
- Foundation models provide a promising method for exploring the huge search space in a meaningful manner. They have
learned intuition about what human find interesting and can therefore probe interesting areas.
- Open-endedness is an inherently experimental process and requires interaction with the learning space.
- Open-endedness requires guidance from an observer as to what we wish to learn. E.g. humans want to solve
interesting human problems.