Introducing Google's Latest Investment in Open Models: Gemma 2.0

Hi, I'm Kathleen Cane, Research Engineer at Google DeepMind, and I'm excited to be here today to unveil the latest advancements in our open-source AI models, Gemma!

  • 1. Kathleen Creel is a research engineer at Google DeepMind and technical lead of the Gemma team.
  • 2. Gemma's North Star was to empower and accelerate the open-source community's work.
  • 3. Since launching their first models in February, Kathleen has been amazed by projects built on top of Gemma.
  • 4. Google has published key research in AI and ML for over a decade, including on transformers, BERT, and more.
  • 5. Google DeepMind continues this tradition, sharing research for the world to validate, examine, and build upon.
  • 6. Google's support for the open-source community extends beyond research, with work on hardware breakthroughs like TPUs and ML frameworks such as TensorFlow and Jax.
  • 7. Gemma is Google DeepMind's family of open-source, lightweight, state-of-the-art models built from the same research and technology used to create the Gemini models.
  • 8. Gemma models are responsible by design, with safety being a top priority from day one.
  • 9. Data sets are manually inspected to ensure high quality and safety, and models are evaluated for safety throughout development.
  • 10. Final models undergo rigorous state-of-the-art safety evaluations before deployment.
  • 11. Gemma models achieve unparalleled breakthrough performance for their scale and outperform significantly larger models.
  • 12. Models are highly extensible and optimized for TPUs, GPUs, and local devices.
  • 13. They support many frameworks, including TensorFlow, Jax, Caris, PyTorch, and ONNX.
  • 14. The real power of Gemma models comes from their open access and open license.
  • 15. Models are available on various platforms, allowing developers to use them with preferred tools and when needed.
  • 16. Since the initial launch in February, several variants have been added to the Gemma model family, including Gemma 1.0 (foundational LLMs), Code Gemma (fine-tuned for code generation and evaluation
  • 17. These models have been updated since their initial release, with new features like better instruction following, improved chat capabilities, enhanced code performance, and larger parameter sizes.
  • 18. The latest releases include P-Gemma, which combines the Siegel Vision encoder with the Gemma 1.0 text decoder for various image-text tasks and capabilities, and Gemma V2, available in 9 billion an
  • 19. P-Gemma comes with pre-trained weights for specific tasks, fine-tuned variants for object detection and segmentation, and transfer checkpoints for academic benchmarks.
  • 20. Gemma V2 models are performant, easy to integrate into existing workflows, and designed for efficient downstream fine-tuning.
  • 21. The 27b model is available in Google AI Studios, allowing developers to experiment with prompts right away.
  • 22. Extensive human evaluations show that Gemma models are consistently preferred over other open models, including larger ones.
  • 23. The Gemma cookbook on GitHub contains recipes for using the Gemma models, and it accepts pull requests to share projects with the community.
  • 24. Developers can apply for gcp credits to accelerate research using Gemma 2, access significantly improved documentation, and engage with the Gemma team on Discord or other social media channels.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!