Last week, Meta AI presented demo version galactica, a large language model for “storage, unification and analysis of scientific knowledge.” Users found that Galactica generates hostile and insulting articles. After the mass criticism of the language model Meta turned off demo version of Galactica.
The Galactica language model is designed to write scientific literature. Galactica contains more than 48 million articles, textbooks, lecture notes, scientific websites and encyclopedias. Galactica can generate documents, such as literature reviews, Wikipedia articles, lecture notes and answers to questions.
An example of generating a Galactica document
Some users found that if you introduce racist or potentially insulting tips, the language model will create material on these topics that will look believable. For example, one of the users used Galactica to write an article on Wikipedia about a fictional research article entitled “Advantages of eating crushed glass.”
Galactica could also distort scientific facts, for example, indicated the wrong dates or names of animals. In this case, you need to have deep knowledge of the subject in order to notice errors.
In September, Twitter users managed to hack a bot to search for remote work working on the Openai GPT-3 language model. Using a technique called “Attack Inject Attack), users reprogrammed the bot to repeat offensive and provocative phrases.