> Two. Releasing work is generally beneficial for the world. I can't think of a single model release that has ever harmed the world. Let's wait until one actually does before we freak out about whether it might.

I like a lot of the idealism in this comment, and learned a few things too, so thanks for that. I'm not an ML researcher, which may cloud my views a bit, but I think you would be hardpressed to think that the release of ML models hasn't harmed the world. Perhaps we haven't reached AGI, but we don't need to for the release of ML models to harm the world.

I'll give two examples: 1) image detection - Now that image detection is good enough, surveillance in dictatorships can be run more efficiently and scalably in a way KGB couldn't have dreamt of. Even if the release of a model like YOLO saves an hour's worth of time of an ML researcher working for the dictatorship on their surveillance project, this can cause a lot of harm to the oppressed ppl living in those countries.

2) troll bots based on GPT-2. You gave an example of high-quality, human-quality propaganda, but what troll bots lack in quality can be made up by quantity. If you run a lot of troll bots and can sway the dominant viewpoint on all the forums you want to target (which you can since these bots are infinitely scalable), you've achieved your purpose. I also think you are overestimating the quality needed to influence the average person's worldview. For example, we saw from the Cambridge Analytica news that all they had to do to affect a few swing voters' behavior was target them with a few ads. I personally read through some of the example output from GPT-3 and if I were browsing a forum, I wouldn't be all that confident whether they were a bot or not.