DeepMind Unveils 280B Parameter Gopher Model Alongside Ethical Risk Study
DeepMind releases three new papers exploring large language models, headlined by Gopher, a 280 billion parameter model evaluated across numerous tasks alongside a dedicated study on ethical risks.
DeepMind introduces Gopher, a massive 280 billion parameter transformer language model designed to advance natural language processing capabilities. The research team trains a series of models ranging from 44 million to 280 billion parameters to understand how scale impacts performance across various tasks. This exploration highlights the potential of large language models to safely summarize information, provide expert advice, and follow complex instructions.
The findings reveal that increasing model scale significantly boosts performance in specific areas such as reading comprehension, fact-checking, and the identification of toxic language. However, the researchers also note that simply making the model larger does not improve results in logical reasoning and common-sense tasks. This nuanced understanding helps map the exact strengths and weaknesses of scaling up artificial intelligence systems.
Alongside the technical model release, DeepMind publishes a dedicated study examining the ethical and social risks associated with large language models, as well as a paper investigating a more training-efficient architecture. This interdisciplinary approach emphasizes that developing beneficial AI requires collaboration across varied fields to anticipate and address the challenges of training algorithms on existing human datasets.