I have not written any blog for quite some time now.

DevRel @storyprotocol
IP is everything
Alumni @learnweb3 
Ex @bnbchain @flock_io 
Building /creader

crypto-anarchist · onchain social explorer, advisor · spread FC gospel in /chinese community · curate SocialSight.telegram· write 0xluo.xyz · host /luo ·中文 @luo

Got it. That makes sense. Yes, many models undergo some sort of "alignment" during training that isn't publicly disclosed.

I think the black box issue I am trying to discuss here is more on questioning the process. How they train the model, if the model contains more bias, then it should do. For example, scaleAI may have their own data labeling standard, which could intentionally create more bias.

The article mentions that "centralized authorities holding models can lead to black box issues..." However, the "black box" problem in neural networks refers to their lack of interpretability, which decentralization can't solve. Nonetheless, there are already ongoing efforts to explain the functions of each layer in neural networks.