The "it" in AI models is the dataset

3 jxmorris12 1 9/1/2025, 8:30:54 PM nonint.com ↗

Comments (1)

measurablefunc · 7h ago
If the goal is to recreate the training data set then all functional approximations are extensionally equivalent modulo biases introduced by the architecture. What I mean by architectural bias is how missing pieces of the data manifold are imputed, i.e. given some point x (w/o a matching output in the optimization corpus) different algorithms will give different results based on how x is encoded into the interal/latent representation of the data manifold. But even this difference is essentially averaged away by the users b/c the goal is to create something that will please the most number of users so it all eventually converges to the average agreed upon sentiment of a large enough sample of people.