Occasionally, the AI research field becomes heated, not due to new launches but because of debates about who is wrong or who is undermining each other’s research. This time, it was Yann LeCun pointing fingers on Gary Marcus, Grady Booch, and Michael J Black for killing Meta’s Galactica.
It all started with the launch of Sakana AI’s AI Scientist, which the company claims is the first AI system for automating scientific research paper and open-ended discovery. This seems very similar to Galactica, an LLM trained on scientific knowledge for the scientific field. However, it was taken down because researchers were concerned about its outputs, which could be unreliable and prone to hallucination.
“Will the AI negativists who killed Galactica by dousing it with vitriol kill this one too?” asked LeCun, further questioning Booch, Marcus, and Black, and others, “will rescind their prophecies of LLM-powered doom for the scientific community?”
Just weeks after Galactica, ChatGPT was released. “Contrary to their predictions, the scientific publication system was not hurt by the availability of LLMs, let alone destroyed. If anything, it was actually helpful,” said LeCun.
What are these ‘Prophecies of Doom’?
Black, who is the director of Max Planck Institute for Intelligent Systems, had raised questions about the reliability of Galactica and similar products during the launch.
Back in November 2022, Black posted a thread of all the problems with Galactica and said, “I applaud the ambition of this project but caution everyone about the hype surrounding it. This is not a great accelerator for science or even a helpful tool for science writing. It is potentially distorting and dangerous for science.”
Black replied to LeCun saying that his criticism was genuine. “I’m sorry but the truth is that *YOU* killed it, not me,” said Black. He mentioned that LeCun promoted Galactica as, “Type a text and galactica.ai will generate a paper with relevant references, formulas, and everything,” which according to Black is a big claim to make.
Thus, when Black conducted the test, Galactica fell short of his expectations. “It made up realistic looking references. And, it did so in a way that sounded authentic,” he added, saying that Meta did the right thing by taking it down as it did not live up to the hype around it.
Marcus, a Professor of Psychology and Neural Science at NYU, is a popular critic of deep learning and AGI, also took to X to state that Galactica got his birthday, education as well as research interests wrong. Nearly 85% of the results presented by Galactica about Marcus were not true.
Similarly, Booch had said, “Galactica is little more than statistical nonsense at scale,” adding adjectives like amusing, dangerous, and unethical.
As for the ‘prophecies of doom’, Black had done a thorough analysis of the problem with LLMs then, and said there should be new ethics rules for scientific publication. He highlighted issues that there can be a flow of fake scientific papers, in an industry which completely relies on peer reviews and public trust. He said that he also uses LLMs for research and there should be further research on LLMs instead of slowing it down and building safeguards to defend the public trust in science.
LeCun doesn’t agree. He replied to Black saying that the negative reaction from a scientist led to shutting down of Galactica. “In simple terms, *YOU* killed it,” said LeCun, asking him if the Galactica demo would have made the world a better place.
So, Who Killed Galactica?
Black believes that Meta, as an organisation, should have been more careful when releasing Galactica, even as a demo, and LeCun should accept the scrutiny faced for promotion of the product.
“You will likely say that this was just a “demo” and not a “product”. Do you think the public really differentiates these when it comes from Meta, has a slick website, and is heavily hyped?” asked Black. Which makes sense. Since Meta released it as a software and not a research publication, it puts Meta under the spotlight, specifically when the claims were not adding up.
Though Black appreciates LeCun and Meta’s efforts towards science and open source AI, he said that Meta is not a university, thus receives a lot of attention from the world, not just academia, which makes it riskier.
Today, the weights of Galactica are still up for academic research, but the product is not available for direct use. “That was a good decision. It shifted the focus away from Galactica being a Meta “product” to it being a research prototype,” said Black.
Maybe that is also why LeCun had invested in Perplexity AI, which often sells itself as a research focused search engine.
Probably, Black is right. With the advancements in LLMs and people getting accustomed to them in the real world, shutting down Galactica back then sounded like a decent step to take. He didn’t kill Galactica, his feedback was strong enough that Meta had to kill it.
However, now seems like a good time to finally bring back Galactica.
