Provably Safe Artificial General Intelligence via Interactive Proofs
نویسندگان
چکیده
Methods are currently lacking to prove artificial general intelligence (AGI) safety. An AGI ‘hard takeoff’ is possible, in which first generation AGI1 rapidly triggers a succession of more powerful AGIn that differ dramatically their computational capabilities (AGIn << AGIn+1). No proof exists will benefit humans or sound value-alignment method. Numerous paths toward human extinction subjugation have been identified. We suggest probabilistic methods the fundamental paradigm for proving safety and between disparately autonomous agents. Interactive systems (IPS) describe mathematical communication protocols wherein Verifier queries computationally Prover reduces probability deceiving any specified low (e.g., 2−100). IPS procedures can test behavior control incorporate hard-coded ethics value-learning methods. Mapping axioms transformation rules system finite set prime numbers allows validation ‘safe’ via number-theoretic Many other representations needed various properties. Multi-prover IPS, program-checking probabilistically checkable proofs further extend paradigm. In toto, provides way reduce ↔ AGIn+1 interaction hazards an acceptably level.
منابع مشابه
Artificial General Intelligence via Finite Covering with Learning
This position paper claims that the combination of solutions to a finite collection of problem instances and an expansion capability of those solutions to similar problems is enough to achieve the artificial general intelligence comparable to the human intelligence. Learning takes place during expansion of existing solutions using various methods such as trial and error, generalization, case-ba...
متن کاملSelf-Regulating Artificial General Intelligence
This paper examines the paperclip apocalypse concern for artificial general intelligence. This arises when a superintelligent AI with a simple goal (ie., producing paperclips) accumulates power so that all resources are devoted towards that goal and are unavailable for any other use. Conditions are provided under which a paper apocalypse can arise but the model also shows that, under certain ar...
متن کاملRisks of general artificial intelligence
The papers in this special volume of the Journal of Experimental and Theoretical Artificial Intelligence are the outcome of a conference on the ‘Impacts and Risks of Artificial General Intelligence’ (AGI-Impacts) that took place at the University of Oxford, St Anne’s College, on 10 and 11 December 2012 – jointly with the fifth annual conference on ‘Artificial General Intelligence’ (AGI-12). The...
متن کاملFacets of Artificial General Intelligence
We argue that time has come for a serious endeavor to work towards artificial general intelligence (AGI). This positive assessment of the very possibility of AGI has partially its roots in the development of new methodological achievements in the AI area, like new learning paradigms and new integration techniques for different methodologies. The article sketches some of these methods as prototy...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Philosophies
سال: 2021
ISSN: ['2409-9287']
DOI: https://doi.org/10.3390/philosophies6040083