Accelerate your generative AI projects with new tools on Vertex AI

The session begins with Jason Gelman welcoming the audience to a discussion on Vertex AI and its generative capabilities. He introduces himself as the manager of several AI projects at Google Cloud and his colleague Eric, the vice president of data science at Estée Lauder. The agenda includes an overview of Vertex AI, its integration into Google Cloud's suite of products, and a focus on generative AI capabilities. Eric will share how Estée Lauder utilizes Vertex AI for business solutions, followed by demos and a Q&A session. Google Cloud's technology stack is briefly explained, highlighting its infrastructure, Model Garden, AI platform capabilities, and the importance of responsible AI tooling.

Eric Higgins from Estée Lauder Companies discusses the company's use of Google Cloud and Vertex AI to enhance consumer experiences in the luxury beauty sector. He provides a business perspective, mentioning the various brands under Estée Lauder's umbrella and the challenges faced in the market, such as clienteling, digital experiences, and keeping up with trends. Eric emphasizes the company's guiding principles and the formation of a new data science team in 2021, coinciding with the launch of Vertex AI. The team's mission is to deliver data science and analytics solutions across all brands globally. Eric also discusses the importance of a DevOps culture reinforced by Chris Curro, a machine learning leader at Estée Lauder, and how Vertex AI and Google Cloud tools have enabled their efforts.

The paragraph focuses on how Vertex AI has been instrumental in delivering business value for Estée Lauder. The company's data science team, formed concurrently with the launch of Vertex AI, has progressively integrated more AI solutions into their system. Early successes included A/B testing frameworks, advanced recommendation systems, and consumer lifetime value models, which led to cost savings and increased marketing efficiency. Eric details the role of Vertex AI in their data science platform, highlighting its importance in maintaining focus on business deliverables and development partnerships with Eviden for staff augmentation and knowledge sharing. The architecture of their current system is also discussed, with Vertex AI playing a central role alongside BigQuery and other Google Cloud tools.

Jason Gelman delves into the Model Garden, Google Cloud's collection of curated AI models, emphasizing its role in providing a starting point for AI journeys. The Model Garden integrates Google's own, open-source, and partner models with Vertex AI's tooling, allowing for quick starts and end-to-end journeys. It includes a variety of large foundation models and task-specific ones, with the collection growing continuously. The integration of models with MLOps tooling is highlighted, as well as the commitment to providing choice to customers. The paragraph also touches on the new OSS models, partner models like Meta's Llama 2 and Code Llama, and the pre-announcement of Anthropic's Claude 2. The Model Garden's commitment to providing more than just links to model repositories, including responsible AI tooling and model evaluation, is emphasized.

Jason demonstrates the Model Garden interface and its functionalities, focusing on a marketing use case. He explains how generative AI can aid in content creation for marketing, such as generating text and images for a healthy snack product line. The use of PaLM 2 for text generation is showcased, with detailed steps on navigating the Model Garden, selecting models, and customizing settings like temperature and token limit. The concept of streaming and grounding is introduced, with grounding being a novel feature to customize model responses with enterprise data. Despite a technical issue preventing the completion of the live demo, Jason provides insights into how Model Garden can be utilized for various applications, including image generation and fine-tuning models with custom data within an enterprise environment.

The discussion continues with Jason highlighting the ease of deploying custom solutions using Model Garden, specifically focusing on Meta's Llama 2 model. He demonstrates how to access model cards, utilize the new 'Open Notebook' feature for model tuning and deployment, and integrate custom data for fine-tuning. The enterprise version of Google Colab is introduced, allowing for seamless integration with Vertex and full control over data. Jason also explains the process of reinforcement learning with human feedback tuning (RLHF) and its benefits. The paragraph concludes with an emphasis on Vertex's enterprise controls, commitment to global AI regulations, and responsible AI tooling, including content moderation and bias evaluation, ensuring the safe and ethical deployment of AI models.

In the final paragraph, Jason discusses the application of AI for improving product search on e-commerce websites. He introduces the concept of text embeddings for generating vector representations of product descriptions and highlights Google's textembedding-gecko model. The combination of embeddings with Vertex AI's Matching Engine is proposed as a solution for robust product search. Jason also mentions the recent release of an image embeddings model, suggesting a multimodal search approach that includes both text and image-based searches. The capabilities of Model Garden and the potential of AI in enhancing customer experiences by aiding product discovery are emphasized.