In a newly published FHI Technical Report, “MDL Intelligence Distillation: Exploring strategies for safe access to superintelligent problem-solving capabilities”, Eric Drexler explores a general approach to separating learning capacity from domain knowledge, and then using controlled input and retention of specialised domain knowledge to focus and implicitly constrain the capabilities of domain-specific superintelligent problem solvers.
The report argues that superintelligent problem-solving systems for relatively broad domains might be implemented though modular architectures that combine well-defined, specialised problem-solvers in well-defined ways. As an example, the report outlines an architecture for conversational engineering-design systems that would omit content and processes of the kind that might otherwise enable strong, risky AI agency, while providing broad, superintelligent competence in engineering domains.
Although the report explores an approach to expanding the scope of safe applications of superintelligence, the potential range of safe applications remains an open question. The report outlines a range of directions for relevant research. For the full report, please see here.