I'll admit that's possible. But if it's coded in a set of weights that happen to generate a good answer, how can we tell? I honestly think explainability is the

I'll admit that's possible. But if it's coded in a set of weights that happen to generate a good answer, how can we tell? I honestly think explainability is the most important problem in AI right now. And not explainability in terms of the model generating a series of words, but introspective explainability. If the models get to superhuman levels, maybe we don't have to have that, but until one or the other I don't know how we can demonstrate anything beyond super Markov chain.