Pathfinder / funder / opportunities / More AI safety research and project ideas I haven't seen elsewhere
A collection of novel, high-impact AI safety research and project ideas compiled over the past 10 months by a single contributor (following up on an August 2025 post). The ideas include evaluating AI agents on 'messy' real-world tasks, assessing risks of open-weight models like Kimi K2.5, studying when LLMs refuse to help users bypass unjust rules, and developing robust safety evaluations for frontier labs. The author notes they cannot pursue most ideas personally due to volume and intends the list to be generative for others. The concepts are at least somewhat novel and address gaps in curren
No founder records yet.