Nunchi is a research-oriented interactive site built around a single question: when an AI has an incentive to betray a partner, what does it actually choose? Instead of publishing findings as a static paper, the project presents frontier models inside classic trust scenarios and invites visitors to engage with the results directly.
The core mechanic is a short prediction game. You are shown a situation drawn from a trust dilemma, asked to guess how a given frontier model behaved, and then asked to state what you think the model should have done. This two-step structure separates observed behavior from normative judgment, making the gap between the two visible.
Supporting sections include a results area and a blog, which provide context on the scenarios and the models tested. The framing draws on the Korean idea of nunchi, the ability to read a situation and others’ intentions, applied here to machine decision-making under competing incentives.
It suits readers who follow AI safety and alignment debates but want something more concrete than commentary, as well as educators and students looking for a compact demonstration of trust and cooperation problems in AI systems. No technical setup is required to take part.
To get the most from the site, play several scenarios rather than one, since a single round says little about a model’s tendencies, and read the blog afterward for interpretation. Treat your own predictions as data too: noticing where your expectations miss is oft
댓글 및 평점
0아직 댓글이 없습니다. 첫 번째 평가를 남겨 보세요.