Google Cloud Intensifies Focus on Generative AI with Vertex AI Updates
Recent findings from a joint survey by Fortune and Deloitte reveal that over half of global CEOs are exploring AI for generating text, images, and diverse data types. Simultaneously, a report by McKinsey indicates that about a third of organizations are using generative AI “regularly” across at least one business function. With the expanding market potential, Google Cloud is doubling down to maintain its competitive edge.
At the recent Cloud Next conference, Google unveiled enhancements to Vertex AI, its cloud platform aimed at building, training, and deploying machine learning models. Vertex AI now boasts upgraded AI models for text, image, and code generation, alongside new partnerships with startups like Anthropic and Meta. These enhancements include extensions that empower developers to integrate proprietary company data and perform actions on behalf of users.
June Yang, VP of Cloud AI and Industry Solutions at Google, highlighted the company’s commitment to an open ecosystem. “We’re collaborating with a wide range of partners to offer our customers more choice and flexibility,” she stated. “Our approach to generative AI focuses on enterprise readiness, emphasizing data governance, responsible AI, and security.”
In terms of model updates, Google has made significant improvements to its Codey code-generating model, achieving a 25% enhancement in quality for major programming languages. Unfortunately, details on this metric were not extensively clarified. The Imagen model has also been improved, enhancing image quality and introducing Style Tuning, which enables customers to generate brand-aligned images with as few as ten reference images.
Additionally, the PaLM 2 language model can now understand 38 languages in general availability, and over 100 in preview, while featuring an expanded context window of 32,000 tokens. This context window allows the model to consider substantial amounts of text, translating to roughly 25,000 words or around 80 double-spaced pages before generating further content.
While this context window isn’t the largest—Anthropic’s Claude 2 boasts a 100,000-token context window—Google’s Nenshad Bardoliwalla emphasized that their choice of 32,000 tokens balances flexibility, cost, and performance. “Our customers are navigating the trade-off between the flexibility of large models and the costs associated with inference and fine-tuning,” Bardoliwalla explained.
To cater to diverse customer needs, Google has introduced third-party models to Vertex AI’s Model Garden, which includes Claude 2, Meta’s Llama 2, and the open-source Falcon LLM from the Technology Innovation Institute. This strategic move positions Google in direct competition with Amazon Bedrock, which allows users to build generative AI apps using pretrained models.
Google is also enhancing Vertex AI with Extensions and data connectors—akin to AI model plugins from OpenAI and Microsoft. Extensions enable developers to link models in the Model Garden to real-time data and third-party applications, while data connectors can seamlessly import data from platforms like Salesforce and Jira.
In a related upgrade, Vertex AI now supports Ray, an open-source compute framework that scales AI and Python workloads, in addition to incorporating Google’s TensorFlow.
However, Google has been reticent in addressing ethical challenges in generative AI, particularly around copyright. AI models like PaLM 2 and Imagen are trained on extensive datasets, often scraped from public, copyrighted sources, raising questions about content ownership. Bardoliwalla mentioned that Google conducts “broad data governance reviews” to ensure compliance but failed to clarify whether users retain ownership of AI-generated content.
Vertex AI Search and Conversation Features
Embracing the trend of AI-powered chatbots and search tools, Google has launched Vertex AI Search and Vertex AI Conversation. These platforms simplify the process for developers to create generative search and chat applications grounded in specific business data.
Now generally available, these tools allow developers to build customizable search engines and chatbots for use cases such as order placement, banking assistance, and customer service automation. New features like multiturn search enable seamless follow-up questions, while summary tools condense search results and chat dialogues.
The introductory Playbook feature for Vertex AI Conversation allows users to define natural language responses for voice and chatbots, adding a persona, goal, and detailed steps for task completion.
Moreover, Vertex AI Search and Conversation can leverage model extensions and data connectors, alongside a new grounding capability to ensure the model’s answers reference company data.
Google plans to enhance Vertex AI Search with enterprise access controls to ensure information is appropriately surfaced and increase user confidence through relevance scoring. However, given the propensity of generative AI models to fabricate information, caution is warranted. There are risks of malicious prompt injection attacks, along with concerns about potential bias in outputs.
Bardoliwalla acknowledged that while grounding tools alone won't completely resolve the issues of hallucination and toxicity in generative models, they are a significant step toward improved reliability.
Google claims each API call to its generative models is assessed for safety attributes, including toxicity and violence. Vertex assigns scores and, in some cases, blocks responses based on these evaluations.
As generative AI becomes increasingly sophisticated, its ongoing interpretability will be an essential area of focus for Google Cloud and its customers.
Ranking
