The Problem

AI Safety research is currently an "ivory tower" domain. While models are being deployed to millions, the technical understanding of how they fail (alignment, reward hacking, deceptive behavior) is restricted to a small circle of PhDs. We lack the scalable safety infrastructure to educate the developers and policy-makers who will govern these systems.

The Solution

I have already launched aiunderstanding.org, a live platform with 130+ technical guides. I am now building "Safety Playgrounds": a suite of interactive, LLM-powered simulations. Unlike static text, these playgrounds require users to interactively "jailbreak" or "misalign" sandboxed models to understand their failure modes in real-time.

Funding Breakdown

• Minimum Funding ($50,000): This covers 6 months of my full-time development stipend, core infrastructure scaling, and the launch of the first 10 "Safety Playground" modules.

• Funding Goal ($100,000): This allows us to scale significantly: hiring a part-time technical researcher to ensure modules match the latest frontier research, and providing massive API/Compute credits to keep the labs free and open to a global audience.

Milestones

• Phase 1 (Months 1-2): Launch simulations for Reward Hacking, Prompt Injection, and Goal Misgeneralization.

• Phase 2 (Months 3-4): Reach 10,000 monthly unique users; launch "Safety Certification" for developers.

• Phase 3 (Month 6): Open-source the simulation framework to allow other labs to contribute their own demos.