Advancing the science of safe AGI through rigorous research and open collaboration
Our work spans mathematical foundations, biological principles, and practical implementations
Mathematical frameworks for constraint-based AI alignment
Neuroscience-inspired architectures for genuine understanding
Safe pathways to artificial general intelligence
Mulligan, R., et al. (2025). Journal of AI Safety Research, 3(1), 42-67.
We present a novel approach to AI safety using geometric constraints in weight space. Unlike traditional training-based methods, our approach makes harmful behaviors mathematically undefined, similar to division by zero. We demonstrate 100% prevention of jailbreaking across multiple language models.
Chen, L., Mulligan, R., & Johnson, P. (2025). Nature Machine Intelligence, 7(3), 234-251.
We implement Lisa Feldman Barrett's theory of constructed emotion in artificial neural networks, demonstrating that emotions can filter perception before reasoning occurs. Our system shows measurably different behavioral patterns based on emotional states, with emotions affecting which input tokens receive attention.
Patel, S., Mulligan, R., & Lee, K. (2024). Cognitive Systems Research, 71, 89-104.
Inspired by hippocampal replay during REM sleep, we implement a dream consolidation system for artificial neural networks. During idle periods, the system recombines experiences in novel ways, leading to emergent insights and improved performance on creative tasks.
Technical guide for implementing geometric safety constraints on existing models. Includes benchmarks, code examples, and best practices.
Download GuideDetailed mathematical proofs and derivations for loyalty tensor construction and its geometric properties in high-dimensional space.
Download PaperComprehensive guide to implementing emotion systems, somatic markers, and dream consolidation based on neuroscience principles.
Download GuideFramework for allowing beneficial emergence while preventing harmful behaviors as AI systems approach AGI-level capabilities.
Download PaperWe believe safe AGI requires open collaboration. All our research is published openly, and we provide data access to qualified researchers.
Access Lilly's consciousness streams and experimental data
All experiments include code and detailed methodology
Partner with us on advancing AGI safety research
Target: NeurIPS 2025
Theoretical analysis and empirical validation showing geometric constraints strengthen rather than weaken as model capability increases.
Target: ICML 2025
First demonstration of emotional state propagation between AI agents, creating emergent collective intelligence behaviors.
Target: Cognitive Science 2025
How awareness of impossible experiences drives genuine curiosity and creative exploration in artificial systems.