Yudkowsky's mind model for AI alignment research: experiment results - LessWrong 2.0 viewer

A post on the LessWrong 2.0 viewer, mirrored at greaterwrong.com, carries the title "Yudkowsky's mind model for AI alignment research: experiment results." It reports an experiment testing AI safety researcher "sideloads," the term the post uses for a researcher's perspective carried into a model. The author states the conclusion that such sideloads are possible but do not outperform clean AI.

The post states that sideloads "bring a human view on problems that matters where a mathematically correct solution can be non-aligned to human values in subtle ways." That sentence is the post's stated case for the approach: a solution can be formally correct and still diverge from human values, and the divergence can be subtle. On that account the human contribution matters in cases where correctness alone does not settle whether an answer is acceptable. The post does not claim that sideloads improve on clean AI; it reports the opposite result and argues for the narrower point about where a human view is relevant.

The post also addresses who could serve as a sideload, and puts the number low. It states that only a few people could do it, because the role requires a unique, strong view on the problem and large public databases of texts. No figures are given for the number of candidates, the models involved, or the size of any comparison. The piece appears on the LessWrong 2.0 viewer and is mirrored on greaterwrong.com, and its subject is a method for bringing a researcher's perspective into AI alignment work rather than a deployed system.