3 ms·
In my opinion, the rationale that the original GPT-2 is too dangerous to release is not good for two reasons: a finetuned small GPT-2 model on a targeted domain
by minimaxir 7y ago
In my opinion, the rationale that the original GPT-2 is too dangerous to release is not good for two reasons: a finetuned small GPT-2 model on a targeted domain dataset (https://minimaxir.com/2019/09/howto-gpt2/ https://minimaxir.com/2019/09/howto-gpt2/) already gives better results than a large default GPT-2 model for targeted news, and despite independent researchers creating GPT-2 clones at a fraction of the cost of adversarial organizations, there hasn't been any evidence of mass-produced fake news in the six months since GPT-2 was open sourced and the paper was released.
GPT-2 will likely be used more for mass-producing crazy erotica (https://twitter.com/Fred_Delicious/status/1166783214750445573 https://twitter.com/Fred_Delicious/status/116678321475044557... [NSFW text]) than fake news.
- HNLurker2 7y ago>GPT-2 will likely be used more for mass-producing crazy erotica We should, at any cost, do not release this to the masses