The AI Bot is known as Galactica.
Galactica to Condense Scientific Information
Meta created Galactica, an AI bot, to condense scientific data. The company introduced it last week.
According to Meta, Galactica “Can summarize academic literature, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.”
The company said that the model trained on “106 billion tokens of open-access scientific text and data.” It’s similar to an academic search engine on steroids. Academics and researchers can use it to find the right paper for their research.
It was supposed to be a cheat code. But it provided vast amounts of misinformation. It cited reputable scientists incorrectly. In that case, the scientists’ reputations could be at risk if they are cited erroneously in the text.
The bot’s problem was that Meta promoted it as a way to get facts and information. However, it behaved like a version of a game where you begin with a half sentence, then allow autocomplete to fill the remaining part of the story.
After hours of Galactica going live, users reported racist and inaccurate articles. One user tweeted:
“Shocked SHOCKED that it only took a handful of questions before Meta’s new Galactica text generation model regurgitated racist garbage. I asked it to write about linguistic prejudice.”
It’s fair to say that the site for the demo cautioned not to follow advice from a language model without verifying it first. That’s why when the demo went public, it was easy to understand the reason for such a disclaimer.
The AI also performed rubbish in kindergarten math. It gave incorrect answers.
Galactica is said to generate trash. It has no motive and it does not produce trash. However, because of how it was trained to recognize words, it produces details that may seem authoritative and convincing.
This is concerning because it could deceive humans, despite having a disclaimer. Because of various complaints, Meta paused the demo.
AI spokesperson at Meta told CNET that the bot is not a source of truth. Rather it is a research experiment that uses machine learning to summarize information. It’s exploratory research.
A chief scientist from Meta said that the demo was taken down because the team in charge of building it was “distraught by the vitriol on Twitter.”
Galactica does seem like a good idea. It could solve a huge problem. It was a worthy investment for Meta to create software to condense decades of work and provide it to researchers in an essay way.
However, Meta seemed to have created the most dangerous AI model. Experts believe that the company could have done better by doing user studies first before putting it into public release.
Galactica and other computerized systems can be a boon to humans. The challenge here is the adoption of the algorithm. It could cause catastrophe. For instance, you could tell it to make coffee. But you forget to tell AI when to stop making coffee. As a result, it turns all objects into coffee.
Meanwhile, take a look at these examples of modern AI in action.