> The only thing they haven't done is release the weights. Not at all. With GPT-3 they only released a paper roughly describing it but in no way it allowed rep

> The only thing they haven't done is release the weights. Not at all. With GPT-3 they only released a paper roughly describing it but in no way it allowed replication (and obviously no source code, nor the actual NN model, with or without weights). GPT-4 was even worse since they didn't even release a paper, just a "system card" that amounted to describing that its outputs were good.