Google Cloud has announced a set of end-to-end AI updates across its platform as it builds on its Vertex and Gemini offerings to help customers achieve “AI everywhere.”
A major offering touted at Google Cloud Next, the company's annual conference held this year in Mandalay Bay, Las Vegas, has been AI 'agents' accessible through its Vertex AI platform.
Developers can set custom tasks and capabilities for agents using natural language in Vertex AI Agent Builder, a new no-code solution for implementing generative AI assistants. Rooted in Google's Gemini family of models, agents can then complete simple or complex tasks as businesses see fit.
In a conversation with assembled media, Will Grannis, vice president and chief technology officer of Google Cloud, described Vertex AI Agents as “at the top” of Google Cloud's sophisticated artificial intelligence ecosystem.
“Moving from natural language processing robots sitting at the top of a website to intelligent agents that can take the context of an organization, its brand guidelines, its knowledge of its own customers and use them to create new experiences, workflows. more efficient work, significantly changing customer service. , agents as a manifestation of all that underlying innovation… is a world away from a year ago.”
At the event's keynote, attendees were shown demos where Vertex AI Agent could identify products within a company's stock based on a text description from a video, or answer detailed questions from employees about the plan. of your company's healthcare.
Through API connections, the agent could recommend specific orthodontic clinics near the user and get Google reviews for each one.
Agents are just one way to connect AI with customer data, and Google Cloud has repeatedly stated its goal of offering as many options as possible to its customers. Another can be found in the advances with Google Distributed Cloud (GDC)
An 'AI anywhere' approach through Google Distributed Cloud
Through GDC updates, Google Cloud has doubled down on its desire for customers to connect AI to their data wherever it lives.
With generative AI search capabilities for GDC, businesses will be able to use natural language prompts to find business data and content. Google Cloud said that because the service is based on Gemma, Google DeepMind's open, lightweight model, all queries and results will remain on-premises to keep data secure.
Enterprises will be able to search data at the GDC edge and in sandboxes using AI, much like Vertex AI agents operate in other parts of their stack.
It's another avenue by which companies can unlock AI in every corner of the workday, even when that means using AI to track sensitive data that never interacts with the public cloud. The feature is expected to go into preview in Q2 2024.
Through its managed approach, Google Cloud also aims to bring its cloud services to its customers' data, even at the edge. This means that, if desired, customers can use GDC's cross-cloud networking functionality to search for isolated data using natural language prompts as easily as they could search for data in the public cloud.
It uses Vertex AI models and a preconfigured API for optical character recognition (OCR), speech recognition, and translation.
Two more announcements that focus on data mining with AI include Gemini on BigQuery and Gemini on Looker, which allow workers to use prompts to discover insights for data analysis.
Both run on Gemini 1.5 Pro, the latest version of Google Cloud's internal Gemini family of large language models that is now generally available on Vertex. The model will feature a sizable 128k token prompt window, along with the capacity for a million jumbo tokens for specific circumstances.
At the keynote, Google Cloud showed a demo where Gemini on Looker could answer questions about the correlation between sales and data and produce charts based on this data. As Gemini 1.5 Pro is multi-modal, the model can also extract information from images.
In the main demo, the user could find information about product sales and supply chain by requesting all results that looked similar to a specific product.
Similarly, Gemini in BigQuery can be used to perform powerful data analysis on all of a company's data using natural language inputs. As part of their commitment to achieving “AI anywhere” without the need for companies to change or sanitize their data sets, BigQuery and Looker can connect to Vertex AI natively to allow models to connect directly to data business.
An open, mix-and-match approach to models
A key selling point of the Google Cloud platform in Google Cloud Next so far has been its pursuit of the widest possible range of models, architecture options and training paths. Several models are newly available on the platform, including all iterations of Anthropic's Claude 3, Google's latest imaging model, Image 2, and a new lightweight open source completion model called Code Gemma.
Grannis described Vertex as a platform aimed at providing customers with “almost any” model out there and empowering them to train, tune and base it on a company's unique data.
Companies will be able to use Google Search to base model results, which Google Cloud says will help ensure that models align with facts and provide verifiable information more reliably. Customers who want to use their own data to inform models will be able to connect agents directly to information from platforms such as Salesforce or BigQuery.
“One of the deepest things in that layer is grounding,” Grannis said.
“So Google search is native with the Gemini API, as well as authoritative search of your own documents, your unstructured data, structured data, all part of Vertex AI. So taking a model and making it really useful in the context of an industry or use case.”
“You don't get real results with the first shot,” Grannis said. “The real results come from iterative development and exploring what these models can do, what they don't do, what data you do and don't have, and the ability to combine that information into your development workflow.”
“Think of Vertex as an opinionated MLOps platform that democratizes experimentation and generative AI,” Grannis added.
