Meta Galactica

Meta Halts Academic Paper Generator AI Demo After 3 Days

Meta unveiled and retracted a generative text AI trained on and designed to write academic papers in just a few days. The Galactica Large Language Model was only a demonstration, but it quickly drew criticism and plenty of ridicule online over the nonsensical or just problematic attempts at academia, presumably with plans to re-release when it has improved.

Galactica Stumble

Like GPT-3 and other synthetic text engines, Galactica attempts to extrapolate long-form writing from a written prompt. The difference is that it was trained specifically on scientific papers, references, and related texts, including databases of biological and mathematic information. The goal was an AI capable of producing Wikipedia-style articles, solving math problems, and annotating and summarizing other academic literature.

“Information overload is a major obstacle to scientific progress. The explosive growth in scientific literature and data has made it ever harder to discover useful insights in a large mass of information. Today scientific knowledge is accessed through search engines, but they are unable to organize scientific knowledge alone,” the researchers explained in the paper. “We outperform existing models on a range of scientific tasks… We believe these results demonstrate the potential for language models as a new interface for science.”

People found many ways to spotlight Galactica’s more confusing or silly outputs. Users prompted Galactica to proclaim the existence of land squids, dispute global warming’s existence, and declare that 29 is the highest prime number. There were many other examples like these highlights:

The attempt at content moderation was also somewhat piecemeal. Though strict enough to refuse to write articles about naked mole rats, the AI’s filter had gaps wide enough to allow it to generate simple step-by-step instructions for creating a bomb with fertilizer. Three days after the first Twitter announcement, the researchers thanked people for trying Galactica but said that they were pausing the demo, with the models available only to interested researchers who reach out and ask.

Though the comments and jokes went unmentioned, it’s hard to imagine they didn’t have some impact on the decision to pull back on access. Galactica and Meta avoided the descent into racism and bigotry that is the frequent fate of open-source chatbots, but the problems highlighted by experts and inquisitive laypersons have likely given the researchers plenty to work on before it’s available publicly again.

Meta Shares AI Translator for Languages Without Written Forms

Legal Tech Firm Rally Showcases GPT-3 AI Contract Writing Assistant Spellbook

‘The Capture’ Crystallizes Deepfake Fears on TV