From what i can tell, you're saying that when the model runs, it compresses the data into a crystallized, idiosyncratic set of weights, and running the model a

From what i can tell, you're saying that when the model runs, it compresses the data into a crystallized, idiosyncratic set of weights, and running the model a bunch of times + averaging them will smooth out the results. Is that necessarily better? My matches felt like different people, and i like that crunchy recommendation butter more than the smooth version (unlike in real life).