On November 15, Meta unveiled a new language model called Galactica, designed to help scientists. But instead of being welcomed by the scientific community as Meta had hoped, the Galactica project died after three days of intense reviews.
Yesterday the company download the public demo that promoted everyone to try the language.
Meta's mistake – and its hubris – show once again that the big companies technology they have a blind spot about the serious limitations of large language models. There is a large body of research that highlights the flaws of this technology, such as its tendency to reproduce biases and present fake news as fact.
Galactica is a large language model for science, trained on 48 millions examples of scientific articles, websites, textbooks, lecture notes and encyclopedias.
Meta promoted its model to researchers and students. In the company's words, Galactica “can summarize academic papers, solve math problems, create Wiki articles, write scientific code and much more”.
Absolutely.
— Grady Booch (@Grady_Booch) November 17, 2022
Galactica is little more than statistical nonsense at scale.
Amusing. Dangerous. And IMHO unethical. https://t.co/15DAFJCzIb
But the treasure turned out to be coal. Like all language models, Galactica is a mindless robot that cannot distinguish fact from fiction. Within hours, scientists began to discover its biased and erroneous results.