Iliad does research and fieldbuilding to develop AI safety as a theoretical and experimental science.
On the research side, Iliad incubates technical research bets. We are interested in funding theoretical projects that develop concrete predictions that can be tested experimentally, as well as methodologically mature, theoretically-informed experiments whose results can be used to advance theory. If you are working on a project that fits these specifications, contact us.
Iliad runs the Iliad Intensive to train researchers in the principles of AI safety research, hosts conferences and a fellowship, and will be launching an Alignment Journal in July 2026.
Incubated Bets
-
Principia builds theoretical model organisms to capture the complex phenomena we observe in neural networks. Using tools from physics, it derives analytical insights that yield testable predictions about the inner workings of intelligent systems, advancing our understanding of the principles that govern intelligence and alignment.
-
XOR Labs develops a principled understanding of how agentic behavior emerges in AI systems, and what capabilities and risks result from this emergence. This work spans both individual AI systems and systems of interacting agents, with the goal of identifying how these risks can be effectively mitigated.
-
Stormglass studies how design choices shape model behaviors, uncovering the causal mechanisms of training itself through rigorous theory and practical safety tools. Its work enables researchers, labs, and policymakers to anticipate risks and detect unsafe behaviors earlier —to ensure future AI systems are built on a robust foundation.
-
AIXI is the leading mathematical model of artificial superintelligence, representing the theoretical upper limit of AI capabilities. By modeling AI risk factors and safety mitigations as AIXI variants and translating them to real AI agents, it enables rigorous testing of both the risks and the interventions designed to address them.
-
A peer-reviewed journal for a foundational approach to AI safety. Launching July 2026.