– The failed launch of Galactica, Meta’s scientific generative AI
On November 15, Meta released a demo version of Galactica, an artificial intelligence developed for scientific research. Two days after its launch, he was taken down for talking nonsense.
Emily Turrettini – journalist specializing in new technologies
Galactic is a large-scale language model (LLM), trained on 48 million pieces of scientific data comprising 120 billion parameters. has for Assignment to “summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and much more.”
As soon as it hit the web, many of the answers to questions asked by users were found to be confusing, nonsensical, or wrong.
For example, artificial intelligence invented the existence of software Gaydar to find homosexuals on Facebook and proposed a study on the benefits of consuming crushed glass. In other research findings, many references and quotes have been made up and attributed to real scientists.
According to Carl Bergstrom, professor of biology at the University of Washington, it is – pardon the expression – “a random generator of bullshit”.
According to Dan Hendrycks, an artificial intelligence security researcher at the University of California, Berkeley, interviewed in C/NetworkMeta should have probed its own AI for this kind of drift before releasing it and points out in passing that “Meta’s AI division doesn’t have a security team, unlike its counterparts, DeepMind, Anthropic, and OpenAI.”
It is incomprehensible that the Web giant has made public this version as imperfect as it is dangerous. In addition, each generated result was accompanied by the warning: “The results may be unreliable. Language models tend to invent”.
Many studies show that the shortcoming of generative AI – a known and recognized fact – is its tendency to hallucinate the material it presents in the context of factual information.
Always check the facts stated by an AI writing software. Using this type of tool requires a critical mind, because with each new prompt or prompt, the AI charges for a new text, sometimes contradicting the previous one.
Galactica’s goal is commendable in wanting to help scientists process the pharaonic mass of published scientific information, impossible for a human brain to assimilate.
But this is the second time in a few months that a Meta AI disappoints. The chatbot Blenderbot launched in September, it was supposed to allow users to chat with him to learn more about him, but the level of interaction was simply poor. An incomprehensible timing, because this launch followed the controversy around the MDA in June, the Google Chatbot, whose words were so sensible and profound, as to convince an engineer to have reaches the stage of consciousness.
The fervor is big around generative AI, it’s described as the “BIG BANG of Silicon Valley” and “the invention that will define the decade to come“.
By launching this premature version of Galactica, Meta has brought these software into disrepute. It shouldn’t be taken into consideration. A real revolution is underway and all sectors of the economy will be affected.