Meta Galactica Author Breaks Silence on Model's Turbulent Launch

Cast your mind back a year, a fortnight before OpenAI would unleash ChatGPT. One AI model from Meta came, caused a ton of controversy, then disappeared. Now, one of the original authors has spoken out.

Ross Taylor co-founded the Papers with Code platform that houses research papers for machine learning models. He was one of the original authors of a paper outlining Galactica, a large language model trained on scientific papers.

However, the demo for the model caused controversy – Galactica generated citations to papers that didn’t exist. Just days after it was published, Galactica was promptly pulled.

In a post on X (Twitter), Taylor broke his silence, saying the team behind the model was “overstretched and lost situational awareness at launch by releasing [a] demo” without checks.

“We were aware of what potential criticisms would be, but we lost sight of the obvious in the workload we were under.”

The model outperformed the likes of PaLM from Google and DeepMind’s Chinchilla, but it was the demo that let it down.

Taylor said the idea for the demo was to understand the distribution of scientific queries that people would use for LLMs. He added: “Obviously this was a free goal we gave to journalists who instead queried it outside its domain. But yes, we should have known better.”

Related:Update: Meta’s Galactica AI Criticized as ‘Dangerous’ for Science

‘Really stupid’ commentary

The author blamed users for Galactica’s failure, saying that people saw it as “a product.”

“We put our vision etc. on the site, which misled about expectations. We definitely did not view it as a product! It was a base model demo,” he said.

While he thought the drama around the demo “could have been avoided” by the research team, he said fears of the model generating “fake science” were “very ridiculous.”

At the time, Michael Black, director of the renowned Max Planck Institute for Intelligent Systems said Galactica could “usher in an era of deep scientific fakes.”

And Grady Booch, part of the trio that developed the Unified Modeling Language, described Galactica as “little more than statistical nonsense at scale.”

Taylor wrote that the anti-Galactica commentary was “really stupid, however, we should not have allowed that to even happen if we had launched it better.”

‘Would do it all again’

Taylor said on X that he sticks by the Galactica research – even the demo decision, which he added was “unprecedented openness for a big company with a large language model at the time.”

“[The demo] was just misguided given the attack vectors it opened for us,” he added.

Meta has since stuck to its open source strategy, leading to the release of the popular large language model LLaMA and the subsequent, more powerful Llama 2, an outcome Taylor refers to as fortuitous as lessons and work from Galactica went into most recent of those models.

Related:In a Rare Outburst, Meta’s LeCun Blasts OpenAI, Turing Awardees

“Despite all the above, I would do it all again in a heartbeat,” he admitted. “Better to do something and regret, than not do anything at all. Still hurts though!”