they acknowledged even in the original blog post that any well-funded group of NLP researchers would be able to replicate their work within a few weeks/months (including whatever corporation or state or terrorist group you worry about), and that in terms of methodology it is a natural, incremental improvement over existing techniques. so it's sort of obvious that withholding release can't prevent any harm.
the good faith reason to withhold release is as you say -- start a conversation now about research norms so researchers have some decision making framework in case they come up with something really surprisingly dangerous.
the bad faith reason is it gives them great PR. this was surely part of the motivation, but I bet OpenAI didn't quite expect the level of derangement in the articles that got published, and may regret it a little bit.
Can we even say for certain it's an improvement over say something like TransformerXL? As far as I could see, the changes over GPT were a couple extra and tweak to layer normalizations, a small change to initialization and a change to text pre-processing. Other than for pre-processing, I didn't catch anything on theoretical motivations for these choices nor anything on ablation studies. The only thing that can be said for certain is it used lots of data and a very large number of parameters, trained on powerful hardware and achieved unmatched results in natural language generation.
the good faith reason to withhold release is as you say -- start a conversation now about research norms so researchers have some decision making framework in case they come up with something really surprisingly dangerous.
the bad faith reason is it gives them great PR. this was surely part of the motivation, but I bet OpenAI didn't quite expect the level of derangement in the articles that got published, and may regret it a little bit.