I recently completed BlueDot Impact's AI Safety Fundamentals course and received the certificate. It is not a major endpoint, but it moved me from simply building useful AI applications toward asking how these systems fail, how they should be evaluated, and how tool-using agents can become safer before people truly depend on them.
What the Course Taught Me
BlueDot is essentially a structured map of AI safety. Before the course, I had already built RAG and full-stack AI projects, but I had not deeply understood how to describe and address their failure modes: incorrect retrieval, missing permissions, over-trust in fluent answers, and similar problems.
The course helped me connect those problems to larger safety questions, including specification, evaluation, deployment risk, interpretability, and scalable oversight. It did not make me less interested in building AI applications, but it changed what I want to focus on. I now want to work more on reliable AI infrastructure.
Where I Am Going Next
The direction I most want to pursue is agent safety evaluations for tool-using and harness-based systems. My current model-eval-system project is a small evaluation suite for daily administrative agents. It tests whether an agent can complete useful tasks while treating retrieved application content as untrusted data.