This is certainly a reasonable idea, and the OpenAI release statement did include a risk assessment attempting to determine who would be able to replicate their work.
It's sort of hard for me to understand how this lead to the decision they made, though. OpenAI's original aim was democratizing access to AI tools so they wouldn't be abused by a handful of players, but releasing a scaled-down model seemed to primarily limit its utility for hobbyists. Assembling and training on a larger corpus is precisely the sort of step that's out of reach for individuals, but extremely approachable for governments or corporations with lots of computing power and expert-hours.
That said, OpenAI also noted that their staged-release model was something of an experiment, so it's not entirely clear to me how much danger they thought GPT-2 actually posed. It seems to have been a useful experiment in gradual release and collaboration with other researchers, and a test of 'inoculating' people via a weak model. GPT-2 isn't a great tool for tricking people or disrupting conversations, but publicizing awareness of AI text generation before turning it loose has probably further decreased that risk.
It's sort of hard for me to understand how this lead to the decision they made, though. OpenAI's original aim was democratizing access to AI tools so they wouldn't be abused by a handful of players, but releasing a scaled-down model seemed to primarily limit its utility for hobbyists. Assembling and training on a larger corpus is precisely the sort of step that's out of reach for individuals, but extremely approachable for governments or corporations with lots of computing power and expert-hours.
That said, OpenAI also noted that their staged-release model was something of an experiment, so it's not entirely clear to me how much danger they thought GPT-2 actually posed. It seems to have been a useful experiment in gradual release and collaboration with other researchers, and a test of 'inoculating' people via a weak model. GPT-2 isn't a great tool for tricking people or disrupting conversations, but publicizing awareness of AI text generation before turning it loose has probably further decreased that risk.