Artificial Minds, Human Values

Artificial Minds, Human Values is the idea, not the résumé line. It is the research and writing project of Ruize Xia, a student at Nanjing Foreign Language School.

The starting claim is narrow and demanding: technical progress is not automatically human progress. Speed, scale, and benchmark scores are easy to talk about. Dignity, access, and the right to disagree with a machine are harder — and more important once a system leaves the lab.

Three ideas organize the work.

Access is a design problem. If a model can generate language or video, the first question is who can enter the conversation. Text2Sign is the current test of that idea: text-to-sign-language video on a single GPU, with public code and a released model.

Capability should survive real limits. A method that only exists on a cluster is not the same as a method someone can run, inspect, and share. The on-device diffusion kernels and the University Ranking app are two different attempts to keep useful systems inside ordinary constraints.

Adaptation needs to stay visible. Fine-tuning can raise a score while changing what a model attends to. The CLIP attention study measures that drift. The notes on human judgment and value-aligned evaluation ask who remains able to contest the result.

The NFLS AI Club is the same argument in a classroom: technical skill, then the question of whether the system should work that way, and for whom.

Contact: xiaruize0911@gmail.com · GitHub · ORCID · Hugging Face