The Meta Galactica AI project, which was supposed to support the preparation of scientific papers, was launched last week, but is already unavailable, but with the help of users, it quickly turned into a joke.
Meta (Facebook) after just three days https://twitter.com/ylecun/status/1593293058174500865 for a demo of an AI system called Galactica AI published on November 15th. Galactica is a large language model (LLM), which is said to have been trained to produce similar texts by processing nearly 50 million scientific articles, but in practice it only produced formally plausible, but in fact nonsensical, sometimes out-of-the-box writings.
After prodding pundits and social media users in the field, Galactica produced a series of fabricated nonsense about a Soviet bear sent into space or that the current Russian-Ukrainian war actually started with the Ukrainian military’s attack in February. Chief Artificial Intelligence Scientist at Meta, Yann LeCun announced the shutdown of the demo in an offended Twitter message, citing that the system was being misused by many people.
Although many people recognize the long-term potential of the model, according to critics, the quick shutdown of the project is equivalent to admitting that the software robot bot was really revealed to the public too early. A day earlier, LeCun had posted about articles and new interesting scientific content whose authors were using Galactica to write them, but it soon became clear that the model was not only unreliable, but also prone to “inventing” certain things.
Is this not going to work out?
It obviously didn’t help judging the matter that Meta’s imager immediately began to scold the users for making a joke out of his great device. They didn’t have too much trouble doing this, though: Galactica spit out a random mix of truth and nonsense even for reasonable input texts, while LeCun recently announced the demo by saying that “you type in a text and it generates a paper with relevant references, formulas , with everything.”
Futurism.com reports on it all Gary Marcus cites a neuroscientist and artificial intelligence expert who called the system’s outputs simply nonsense and noted that Galactica follows in the footsteps of OpenAI’s GPT-3 text generator, which is capable of producing grammatically correct texts, but how much meaning the texts have is another question. To this extent, the project fully reflects one of the oddities of contemporary AI applications.
According to Marcus, large language models such as GPT-3 and Galactica are like bulls lost in a china shop: very powerful, but equally reckless, which significantly increases the challenges caused by misinformation in terms of their field of application. Others point out that using LLMs practically as a database is doomed to failure, while they are much more suitable as a cognitive engine or for coordinating tasks.
Tags: Soviet space bear Metas cherished artificial intelligence
