A new benchmark, FACTS Grounding, evaluates the accuracy of large language models (LLMs) in grounding their responses in provided source material and avoiding hallucinations.

Source: https://deepmind.google/discover/blog/

Source: https://deepmind.google/discover/blog/

Source: https://deepmind.google/discover/blog/

Veo 2 generates videos without modifications, acknowledging key research and engineering contributions from multiple individuals.

Source: https://deepmind.google/discover/blog/

Google and Alphabet CEO Sundar Pichai announced Gemini 2.0, a new AI model that builds on the success of Gemini 1.0 and 1.5, which enabled multimodality and long context understanding across various inputs and outputs.

Source: https://deepmind.google/discover/blog/

Project Astra is a research prototype aimed at prioritizing safety and security in technological development. It is being tested by a limited group of trusted individuals. The project is supported by a team of exceptional contributors.

Source: https://deepmind.google/discover/blog/

A new enhanced version of the popular MMLU dataset, called Gemini 1.5 Pro, has been released with questions covering multiple subjects and increased difficulty levels. It includes a code-generation component and a held-out dataset not available online.

Source: https://deepmind.google/discover/blog/

Gemini has released an enhanced version of their MMLU dataset with increased difficulty and questions from various subjects. They also introduced Code Natural2Code, a code generation tool that supports Python, Java, C++, JS, and Go. The Gemini 1.5 Pro version shows improved performance in flash tests, and a held-out dataset, HumanEval-like, is available for testing.

Source: https://deepmind.google/discover/blog/

Gemini Ultra emphasizes the importance of responsible AI development and usage to benefit society.

Source: https://deepmind.google/discover/blog/

DeepMind launches DeepMind Health to address societal challenges with technology.

Source: https://deepmind.google/discover/blog/