grantmaking.ai Launch Round
Lighthouse is a research project investigating whether many conversational AI failures share underlying behavioral dimensions.** The project explores whether those dimensions can be independently measured and governed through controlled experimentation.
To investigate these questions, I worked with a senior AI engineer to build a research testbed and completed an initial round of controlled experiments. Rather than validating a finished solution, those experiments raised new questions about how different behavioral dimensions interact and helped define the next phase of the research.
I’ll continue to lead the research by designing the experiments, defining the behavioral dimensions, and analyzing the results. I’m working with a senior AI engineer with extensive experience in AI infrastructure and distributed systems who built the research testbed and continues to provide technical guidance.
The project will produce an experimental research platform, a framework for evaluating conversational behavior, and empirical findings about whether these behaviors can be independently characterized and governed. The research will help determine whether failures like sycophancy, false certainty, persuasion, and over-engagement stem from underlying behavioral dimensions or require fundamentally different approaches.
Minimum ($15,000):
Funding would support continued development of the existing research testbed, additional senior AI engineering time, cloud compute and infrastructure, and a focused set of controlled experiments investigating the project’s core research hypothesis. This would allow me to continue developing and evaluating the behavioral framework while generating additional empirical findings.
Ideal ($34,000):
Funding would support additional senior AI engineering time, expanded behavioral experiments across a broader range of conversational scenarios, expansion of the behavioral evaluation framework and supporting evaluation tools, and documentation and publication of the research findings. This represents funding for the next defined phase of research. Exact costs will depend in part on what the research reveals.