Meta’s New AI Galactica Model Is Gone – It Makes Stuff Up

A new Meta AI Galactica demo of its LLM, Large Language Model, was published on Tuesday for public use and was shut down within two days. Meta trained an AI with 48 million science papers, but it extrapolated like a dog, making stuff up.

Visitors to the Galactica website could type in prompts to generate documents such as literature reviews, wiki articles, lecture notes, and answers to questions. Unfortunately, users quickly learned they could manipulate the data “we know about the universe” into gibberish.

Users could write anything. Racist and inaccurate scientific literature took no time at all to conjure up.

Galactica is a large language model (LLM) to “store, combine and reason about scientific knowledge.” It was supposed to speed up the writing of scientific literature. Unfortunately, the models are simplistic with no conceptualization ability.

One user suggested that Interpolative [alter or corrupt something, like text] Search Engine – ISE would be a far better acronym for LLMs.

The developers thought the high-quality data would end in high-quality output. Simplistic!

It could spit out inaccurate instructions on how to make napalm in a bathtub; a fake wiki entry on the benefits of suicide; a wiki entry on the benefits of being white; and research papers on the benefits of eating crushed glass.

It could be racist. After a handful of questions, META’s new Galactica text generation model put out racist garbage. The user asked it to write about linguistic prejudice, and it put out prejudice. The user, Rikker Dockum wasn’t even slightly shocked.

It could become offensive. [Gaydar is a colloquialism referring to the intuitive ability of a person to assess others’ sexual orientations as homosexual, bisexual or straight.]

It could be used to write believable fakes. It’s offline and the chief scientist was not happy.

It didn’t live up to its hype. It’s a dog that makes stuff up.

PowerInbox