Aspirationally, human-compatible AIs would learn, share what they learn, and collaborate to achieve high standards. They would communicate, establish common ground, learn and read critically, consider the provenance of information, test hypotheses, and collaborate with people.
Creating competent, affordable, and embodied human-compatible AIs would open many long-imagined applications for robots that are not possible using only today’s mainstream AI technology.
These essays in recommended reading order explore the dimensions, research challenges, approaches, goals, and possibilities for creating human-compatible AIs. Below are pre-publication versions of the papers.
The “What and Why” Paper: What AIs are not Learning — 16 pages
- Why do robotic service applications require so much knowledge?
- What are experiential foundation models and why are they potentially better than manual programming or large language models?
- What is a developmental AI approach for creating experiential foundation models?
The “Deep Values” Paper: Towards AIs with Deep “Human-Compatible” Values — 15 pages
- What is the nature of values?
- Why do different groups of people have different values?
- What are “human-compatible” values?
- How do human children and adults (and members of other social species) acquire values?
- How could we create AIs that maintain alignment with human-compatible values?
The “Collaborative AIs” Paper: Roots and Requirements for Collaborative AIs — 24 pages
- Who needs collaborative AI?
- What competences are needed for effective collaboration?
- What is the relationship of AI (artificial intelligence) and IA (intelligence augmentation)?
The “How” Paper: Bootstrapping Developmental AIs — 106 pages
- How do children learn so much so quickly?
- How do people (and animals) acquire competences?
- How does multi-model information fusion work?
- How do early-acquired competences prepare the way for later ones?
- How does a trajectory for acquiring competences work (for humans, animals, machines)?
Publications
Stefik, M. (2024) What AIs are not Learning (and Why). arXiv https://arxiv.org/abs/2404.04267
Stefik, M. (2024) Towards AIs with Deep “Human-Compatible” Values. DropBox Link
Stefik, M. (2023) Roots and Requirements for Collaborative AI. arXiv https://arxiv.org/abs/2303.12040
Stefik, M., Price, R. (2023) Bootstrapping Developmental AIs. arXiv https://arxiv.org/abs/2308.04586