“The safety concerns about releasing AI models, such as those raised about GPT-2 six to seven years ago, have historically proven to be overblown.”