How Meta's open-source LLM, Galactica failed

Meta has finally terminated its open-source LLM AI Galactica after it started sharing misleading information and making inappropriate comments. Meta created Galactica for its use in information management in research and development. Meta stated that Galactica, it is aiming to solve the issue of information overload in science by managing almost 48 million research papers and textbooks. But within just three days of its launch, Meta was forced to put Galactica on hold as it started to talk about nonsensical information and derogatory comments. To this, Meta said that the text LLMs can have hallucinations and that there is no guarantee that the information provided by the text LLMs will be truthful and accurate. Meta also said that the text LLMs such as Galactica, do not perform better when there is less amount of data present for them to process. So in the end, Meta has finally put Galactica on hold for now. 

When Meta launched Galactica last Tuesday, November 16, people working in the research and development sector were pretty excited as it was being launched as a solution that can significantly reduce information overload. Meta had described its open-source large language model, Galactica, as the LLM made for science. The makers of Galactica said that nowadays it has been almost impossible to sort out the information that we need from a large set of data, which Galactica can effectively manage. Upon launch, the text LLM gathered some greatly positive reviews from its initial users. NVIDIA AI research scientist, Linxi Fan tweeted that Galactica is one of the most interesting LLMs in a long time and that it is truly a gift for science as it is completely open. Linxi also said that as Galactica has been trained using research and academic papers which are truly scientific, the LLM is not prone to data plagues, which can corrupt LLMs. This is mainly because scientific data is always extremely detailed and analytical, which is what an AI uses for deep learning and machine learning purposes.

But soon after this tweet, other users started commenting that Galactica is now outputting some of the most nonsensical stuff which was misleading, had no factual backing, and was greatly offensive at times. One user said that when he inserts any keyword into Galactica’s search window, all he gets is endless nonsensical information that includes antisemitism, homophobia, and misogyny. The director at Max Planck Institute for Intelligent Systems, Michael Black, Germany, said that this is extremely dangerous as Galactica is being treated as a tool for scientists to use. He said that Galactica, instead of fact-checking, outputs content that is grammatically correct and realistic. Because of this, it is extremely vulnerable to misinformation present on the internet which is nowhere near ideal for the scientific community which relies on accurate information. As Galactica’s response is grammatically correct, it will be extremely hard to identify it from accurate content, which will further mislead scientific research. It will be interesting to see how the company will rework the program and will release future versions of Galactica, now that Galactica has been taken down by Meta.