Read on Twitter
Nine months after saying it was withholding software called GPT-2 that generates text due to safety concerns, OpenAI just...tweeted it out. https://twitter.com/OpenAI/status/1191764001434173440
A thread
A thread
OpenAI researchers trained machine learning software on webpages linked on Reddit. They saw it could generate fluid (if oft nonsensical) text and worried it could be used for fake news or spam. In February they released a small version & sounded the alarm. https://www.wired.com/story/ai-text-generator-too-dangerous-to-make-public/
This annoyed some people in AI research who thought it was fear mongering or hype. A few months later, OpenAI updated their initial blog post to say it would release successively larger versions of the model over time to give others time to "discuss their societal implications."
In August, OpenAI released a version of GPT-2 half the size of the secret one, and a report probing how convincing or biased its output was. OpenAI said its warnings appeared to have convinced AI researchers to be more careful & not release similar models. https://openai.com/blog/gpt-2-6-month-follow-up/
A week later, two masters grads used free Google cloud credits to create their own version of GPT-2. They tried to discuss it with the lead researcher on OpenAI's original paper but he didn't reply. They released the full model and the world did not end. https://www.wired.com/story/dangerous-ai-open-source/
Today, OpenAI released the full GPT-2 and a 70-page report with research on the capabilities of text generation for tasks like propaganda. It asks AI researchers to be more careful about what they publish. Google previously said it withholds some research. https://d4mucfpksywv.cloudfront.net/papers/GPT_2_Report.pdf
OpenAI's report also acknowledges that text-generators are so far not dangerous: "We have not
seen any significant action toward misuse language models during this time...threats...are not as immediate as we previously thought."
seen any significant action toward misuse language models during this time...threats...are not as immediate as we previously thought."
