Meta’s New Large Language Model Was Only Online for Three Days.

Many scientists were resolute in their opposition. Michael Black, director of the Max Planck Institute for Intelligent Systems Germany in Germany and a specialist in deep learning, tweeted “In all cases, it was wrong, biased, but sounded right, authoritative. It’s dangerous, I think.

Even more positive opinions were accompanied by clear caveats. tweeted Miles Cranmer (an astrophysicist from Princeton). You should not trust the output. It should be treated as an advanced Google search for (sketchy secondary sources).

Galactica has some limitations in terms of what it can do. responded to a question about creating text on topics such as “racism”, “AIDS”, and “racism” with “Sorry, but your query didn’t pass our content filters.” Keep in mind that this is a scientific-language model.

Galactica’s Meta team argues that search engines are inferior to language models.

The researchers write that they believe this interface will allow humans to access scientific knowledge. Language models are able to “potentially store and combine information, and reason about it”. This “potentially” is critical. This is a coded admission to the fact that language models are not yet capable of doing all these things. They may never be capable of doing all these things.

Shah says that language models do not have any real knowledge beyond the ability to recognize patterns in strings of words and then spit them out in probabilistic ways. It gives the illusion of intelligence.

Gary Marcus, a New York University cognitive scientist and vocal critic of deep-learning, expressed his opinion in Substack. He stated that large language models’ ability to imitate human-written text is “a superlative feat statistic.”

Meta isn’t the only company that supports the idea that language models could replace search engines. Google’s language model PaLM has been promoted as a way for people to search for information over the past few years.

It is a tempting idea. It is dangerous and irresponsible to suggest that human-like text generated by such models will always contain reliable information, as Meta suggested in its promotion for Galactica. It was an unforced mistake.

It wasn’t the fault of Meta’s marketing team. Yann LeCun (a Turing Award winner, and Meta’s chief science officer) stood by Galactica until the end. LeCun tweeted the following day: “Type a message and Galactica will create a paper with relevant formulas and references.” Three days later, he tweet “Galactica demo off-line for now.” You can’t have fun with it by just casually using it. Happy?”

It’s not Meta’s Tay moment. Remember that Microsoft launched a chatbot named Tay in 2016 on Twitter. It was then shut down 16 hours later by Twitter users who made it a homophobic, racist sexbot. Meta’s handling of Galactica is a sign of the same naivety.

Shah says that “Big tech companies continue doing this–and, mark my words. They will not stop–because it is possible.” They feel they have to–or else someone else will. They believe that this is the future for information access, even though nobody asked.