July 2024 Summaries

Gemma 2 is a suite of open AI models introduced in June, featuring a 27 billion and a 9 billion parameter model, with the 27B model quickly becoming a top performer on the LMSYS Chatbot Arena leaderboard. The series emphasizes responsible AI, prioritizing safety and accessibility, and includes new additions to enhance this focus: Gemma 2 2B, a smaller yet highly efficient model; ShieldGemma, a set of safety classifiers to filter harmful content; and Gemma Scope, an interpretability tool that uses sparse autoencoders to provide insights into model decision-making. These tools are designed to allow researchers and developers to create safer and more transparent AI applications, while enabling flexible deployment across various hardware platforms, including edge devices and cloud services. The open and collaborative nature of these models encourages transparency and advances industry safety standards, with the aim of fostering a future where AI is both beneficial and responsibly managed.

Google has revised its plan regarding the discontinuation of goo.gl URL shortener support, now preserving actively used links rather than deactivating them after August 25, 2025. While any unused links that showed no activity in late 2024 will still be deactivated, users are advised to transition to other URL shorteners if they receive warnings that their links will expire. The URL shortener service was initially deprecated in 2018 due to shifting trends in internet content discovery and the emergence of alternative services. Even so, goo.gl links created through Google apps will remain functional. Starting August 23, 2024, some goo.gl links will display an interstitial page indicating the impending deactivation, potentially disrupting current redirect flows and social metadata displays. To mitigate these disruptions, users can suppress the interstitial page by adding a specific query parameter to their links. Google encourages users to transition their links and reach out to Firebase Support for assistance if needed.

In Bengaluru, Google showcased its advancements in AI and app development, highlighting collaborations and innovations from the Indian developer community. Key initiatives include expanding Project Vaani with the Indian Institute of Science to capture India's diverse languages, launching the IndicGenBench for evaluating language models in Indic languages, and open-sourcing the CALM framework for integrating specialized language models. Google also introduced MatFormer for optimizing AI models, and partnerships with startups like Miko.AI and Cropin to promote responsible AI solutions. The company announced support for 10,000 Indian startups through Google Cloud credits and AI-focused programs, including a GenAI Hackathon and AI Startup Bootcamp. Further developments include Project IDX for full-stack app development, new Firebase services, and the Checks AI Safety service for monitoring AI models. Google is also exploring AI agents to assist in software development and maintenance, exemplified by Project Oscar, and previewed AI-generated UI enhancements with Flutter. The event emphasized Google's commitment to creating an integrated development suite, with future plans to connect with developers in Beijing.

Genkit for Go is an open-source framework designed to advance AI-powered application development in the Go programming language, introduced as a collaborative effort between Firebase and Go teams. Leveraging Go’s simplicity, scalability, and security, Genkit provides developers with tools to build generative AI applications, such as intelligent assistants and customer support agents, by integrating retrieval-augmented generation and data transformation capabilities. Currently in its alpha phase, Genkit offers intuitive libraries and a unified API for generating content from various models, supporting native vector databases, and enabling multi-step AI workflows with built-in observability. It also includes Dotprompt for streamlined prompt engineering and integrates with Google and third-party AI services through a plugin system. Developers are encouraged to experiment with the framework, provide feedback, and contribute to its ecosystem, with the goal of fostering a community of AI developers in the Go language.

Generative AI (gen AI) holds significant potential for creating products that require stringent privacy and confidentiality measures, such as personal assistants and business automation tools. Google has developed technologies like the GenC open-source project, which combines Confidential Computing, Gemma open-source models, and mobile platforms to experiment with gen AI applications handling sensitive data. One example is a mobile app using a hybrid setup of two large language models (LLMs), Gemma 2B on-device for swift, local processing, and the more robust Gemma 7B in Google Cloud's Confidential Space for complex queries, offering a balance of speed, cost-efficiency, and quality. This architecture prioritizes privacy through encrypted communication and verification processes, while also allowing developers to utilize a flexible, language-independent Intermediate Representation (IR) for seamless deployment across various platforms. GenC is designed to simplify the development process and encourage experimentation, with ongoing enhancements expected to improve performance and broaden the range of supported models and hardware.

Gemma is a family of open AI models designed for a wide array of tasks, including text and code generation, vision-language tasks, and more, adaptable across devices from edge to cloud environments. Built for the open community of developers and researchers, Gemma models are highly customizable, allowing for fine-tuning to meet specific needs. A blog post highlights three project ideas utilizing Gemma: translating old Korean literature with its unique language challenges, brainstorming game design concepts, and creating personalized letters from Santa Claus. The translation project, for instance, uses Gemma's SentencePiece tokenizer for translating archaic Korean script into modern Korean, demonstrating the model's adaptability and potential for fine-tuning with limited data sources. The same framework can be adapted for tasks like dialect translation or style transfer between American and British English. Additionally, Gemma supports creative endeavors such as game design and interactive letter writing through its text generation capabilities, offering a versatile tool for both technical and creative projects.

Gemma is a family of open models derived from the Gemini models, designed to handle diverse tasks such as text and code generation, fine-tuning, and execution on various devices. The tutorial explores using Gemma in conjunction with Ray, an open-source framework for scaling AI applications, on Google's Vertex AI platform. It provides a comprehensive guide to deploying and fine-tuning the Gemma model using Ray for distributed computing, outlining steps to create and configure Ray clusters and Docker images, and leveraging Google Cloud services like Cloud Storage and TensorBoard for artifact management and job tracking. The tutorial demonstrates how to fine-tune Gemma for summarizing newspaper articles using HuggingFace Transformers, evaluate model performance with ROUGE metrics, and generate offline predictions at scale with Ray Data. It also emphasizes the importance of cleaning up resources post-tutorial to avoid unnecessary costs, and encourages engagement with the Google Developer Community for further learning and collaboration.