Meta AI and Papers with Code unveil Galactica, an open-source LLM for generating literature reviews, wiki articles, lecture notes on scientific topics, and more
The Galactica large language model (LLM) is being trained with millions of pieces of academic content.
The DecoderMatthias Bastian
Context & Ripple Effects
Galactica is Meta AI and Papers with Code's bid to make an LLM out of the scientific record itself, trained on millions of pieces of academic content and pointed at literature reviews, wiki articles, and lecture notes. The bet is that a curated academic corpus produces more trustworthy text than web-scale training — a direct answer to the hallucination problem that has dogged general-purpose models.
Scientists and students get immediate access to a tool that generates polished scientific prose — literature reviews, wiki articles, lecture notes — but the corpus-driven output still asserts falsehoods, putting the burden of verification on every user.
Second-order effects
The backlash forces Meta AI and Papers with Code to pull the demo within days, trading public visibility for damage control and pushing Meta's next open release, LLaMA, behind a researcher-only gate instead of a public web demo.
Third-order effects
If the pattern holds, open scientific-text models get released to credentialed researchers rather than as public tools, and 'trained on authoritative sources' stops being accepted as a substitute for factuality — shifting the burden to retrieval and verification layers around the model.
The trend: Meta's open-AI strategy is recalibrating from public demos of corpus-trained models toward researcher-gated releases, with Galactica's fast pullback as the cautionary first data point.
🪐 Introducing Galactica. A large language model for science. Can summarize academic literature, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more. Explore and get weights: https://galactica.org/ https://twitter.com/...
A Large Language Model trained on scientific papers. Type a text and https://galactica.ai/ will generate a paper with relevant references, formulas, and everything. Amazing work by @MetaAI / @paperswithcode https://twitter.com/...
Today a 120B model called “Galactica” is open-sourced by @paperswithcode. It's capable of writing math notations, citations, code, chemical formula, DNA, etc. Here's why I think Galactica is a huge milestone in open foundation models, scientific automation, and responsible AI: 🧵 …
#Galactica AI does not disappoint. Look how short that proof is, and it saved me time generating my own LaTeX error and false claim about projected gradient. https://twitter.com/...
Pretty impressed so far with initial results from GALACTICA developed by @MetaAI, a scientific LLM that can write academic papers and wiki articles. Here is an example of an AI generated wiki article on LI-RADS with citations https://galactica.org/...
Trying out #Galactica, the new open source AI model from @MetaAI that has been trained on humanity's scientific knowledge. It's still not perfect but a very interesting tool to search about new topics, it also provides you the knowledge source of its results. https://twitter.com/…
The new language model for science https://galactica.org/. Upon few quick tries, it seems to generate professional text in the areas I am familiar with. And 7 years ago we were *joking* about ML writing papers!
“Write a wiki article about the golden nosed dolphins of the Caribbean.” — https://galactica.org/ vs. GPT-3 Safe to say, GPT-3 understood the task better, was factually rich (even if the “facts” are made up), and was more creative. https://twitter.com/...