What’s new with generative AI at Google Cloud

The video script begins with June Yang, Vice President of Cloud AI at Google Cloud, welcoming the audience to Google Cloud Next 2023. She expresses her excitement about being back in person and introduces the topic of generative AI at Google Cloud. June Yang discusses the major shift in AI with the rise of foundation models, which allows for various use cases without ML expertise and minimal training data. The goal of Google's generative AI is to bring enterprise-ready solutions to help organizations innovate faster. She also mentions Google's comprehensive technology stack, from AI infrastructure to tools for building custom models and applications. June Yang emphasizes the integration of security, privacy, compliance, and control across the stack. She invites attendees to learn more about Duet AI and highlights the significant demand from customers for Google Cloud's generative AI capabilities.

In this paragraph, June Yang announces new versions of Google's foundation model families, starting with the text and chat model PaLM 2. She addresses the increased demand for longer input token lengths and introduces the new version with a fourfold increase, allowing for processing of longer documents. June also announces the general availability of 38 languages and private preview for 100 more languages. Upgrades to the text-to-image model, Imagen, are discussed, including improved image quality and the introduction of style tuning. Additionally, the first Cloud provider to enable digital watermarking for generated images is announced, powered by Google DeepMind SynthID. Codey, the text-to-code model, has also been upgraded with a 25% quality improvement. The paragraph concludes with the announcement of the preview of streaming API for code generation and chat, improving the developer experience.

David DeSanto, Chief Product Officer from GitLab, joins June Yang to discuss GitLab's collaboration with Google Cloud. GitLab aims to revolutionize how development, security, and operations teams work together to build secure software. They provide an enterprise DevSecOps platform that is now enhanced with AI for greater efficiency. David addresses concerns about the misuse of software and intellectual property, outlining GitLab's vision for AI use, which includes applying AI across the software development lifecycle, leading with a privacy-first approach, transparency, and providing the best AI experience for specific use cases. GitLab's partnership with Google Cloud is highlighted, emphasizing shared commitments to enterprise readiness and privacy. The integration of PaLM 2 foundation models, including the Codey model family, is discussed, along with the introduction of GitLab Duo, a suite of AI-powered features for the software development lifecycle.

June Yang continues the discussion by introducing model augmentation tools designed to enhance the capabilities of foundation models for enterprise use. Grounding, extension, and tuning are highlighted as key tools. Grounding allows for model outputs based on user-specified source material, ensuring factual and verifiable responses. Extension enables integration with real-time data and actions via APIs, while tuning allows customization of foundation models with enterprise data for more accurate results. June Yang provides examples of each tool and announces the preview of new grounding capabilities in Vertex AI. The paragraph also covers various tuning methods, including prompt design, adapter tuning, reinforcement learning with human feedback, and full fine-tuning. The importance of model evaluation tools and frameworks is emphasized, as is the introduction of Colab Enterprise on Vertex AI, offering a common notebook interface and AI-powered code completion.

Nikita Namjoshi takes the stage to demonstrate the use of Vertex AI and Gen AI Studio to explore and customize foundation models across different data modalities. The demonstration includes using a large language model to extract structured data from text, showcasing the ability to process documents in various languages. Codey, the text-to-code model, is used to generate Dockerfiles and Python code for machine learning models. The demonstration concludes with the use of Imagen to generate images based on text prompts in different styles, highlighting the versatility of foundation models in various data types.

June Yang discusses the introduction of Vertex AI Search and Conversation, formerly known as Gen App Builder, which combines Google's experience in search and conversation with foundation models. The use cases for generative AI in search and chat applications are explored, such as Q&A with internal knowledge bases and chatbots for employees and customers. The orchestration layer's role in understanding user intent and providing answers from various sources is explained. Raman Pugalumperumal from the World Bank shares their experience as an early adopter of Vertex AI Search, discussing the organization's goals and the importance of knowledge management. He highlights the benefits of generative AI-powered search in content analysis, synthesis, and generation, and the potential to elevate the value of development investments through efficient use of the World Bank's extensive data.

Kalyan Pamarthy demonstrates how developers can build and configure a Search and Conversation app on Vertex AI. He walks through the process of creating a search engine, selecting a data source, and indexing data. The demo showcases the capabilities of the search engine in providing relevant results, summarized answers, and citations for verification. The conversational search experience is also demonstrated, highlighting the ability to maintain context and provide follow-up answers. Kalyan emphasizes the customization options and API availability for the system, allowing for integration into various enterprise solutions.

In the final paragraph, June Yang addresses the importance of enterprise readiness in the context of generative AI, focusing on data governance, privacy, security, compliance, reliability, sustainability, safety, and responsibility. She outlines Google Cloud's approach to ensuring data privacy, with a strong emphasis on the customer's ownership of their data. June Yang also discusses the security and compliance measures in place for Vertex AI, including VPC security control, customer-managed encryption keys, data residency, and access transparency. The commitment to responsible AI is highlighted through a suite of tools for safety, bias evaluation, content moderation, and more. June Yang concludes by emphasizing the ongoing journey with generative AI and the exciting potential it holds for the future.