SUPERINTELLIGENCE WHITE PAPERS
Building Safe and Scalable SuperIntelligence
These papers present a unified framework for designing Artificial General Intelligence and SuperIntelligence that is safe, aligned, and transparent.
Authored by Dr. Craig A. Kaplan, each paper distills more than three decades of research into practical architectures that combine cognitive science, intelligent systems, and AI safety. Together, they outline a pathway toward AI that reflects human values and reduces existential risk.
All papers are freely available for responsible research and safe implementation.
Explore the Papers
Each white paper outlines a key design for building safe AGI and SuperIntelligence. Full PDFs are available for anyone committed to advancing AI safety and alignment.
AGI does not have to come from one very large model. This paper shows how millions of customized AI agents, trained and supervised by the people who use them, combine into AGI, and why that human participation makes the path both faster and safer.
(PDF)
The defining question for AGI is whether its values will match ours. This paper sets out the SCAN-II architecture, which runs ethics checks inside the problem-solving process at machine speed and keeps an auditable record of every decision.
3: Human Centered AGI (PDF)
The safest path to AGI has to be the fastest, or a less careful design arrives first. This paper starts with a network of human problem solvers, adds AI agents that learn from them, and builds the ethics and safety checks into the architecture itself.
Values written by a small group cannot represent everyone. This paper shows how safety and ethics information from many individual agents combines into a statistically valid sample of human values, and how that sample holds as the AGI or SI system scales.
A Personalized SuperIntelligence is only as safe as its owner's values. This paper gives a twelve-step process for building PSI agents and shows how a community of them keeps any single agent from dominating as they grow more capable.
Intelligence grows by acquiring information that matters. This paper introduces Kaplan Information Theory, which measures information by how closely it relates to a goal, and gives a nine-step procedure for finding and testing new information without losing alignment.
Values cannot be derived by logic, so they have to come from people. Ten principles and dozens of methods for designing human values into AGI and SuperIntelligence, with three worked implementations.
Proposes a novel use of online ad systems as large-scale behavioral training environments for AGI and SuperIntelligence. The method leverages global human interaction data to improve ethical decision-making and collective intelligence learning.
Examines how self-awareness can be safely integrated into advanced SI systems. It defines architectures that allow SuperIntelligence to understand and regulate its own goals without diverging from human-aligned ethics.
Introduces the concept of Planetary Intelligence, a globally integrated network of human and AI personaized l SIs that enable our entire planet to think.