ITBusiness.ca

Google upgrades Vertex AI, foundation models, adds Meta’s Llama 2, Anthropic to portfolio and more

Source: Google Cloud

Today, at Google Cloud Next 2023, Google announced that it is expanding the capabilities of Vertex AI, upgrading its own foundation models and is adding new third-party models to its Model Garden, including Meta’s Llama 2, Anthropic Claude 2 and more.

Upgrades to Vertex AI

Vertex AI, Google’s cloud-based platform for building, training and deploying machine learning models, received a series of upgrades to let it keep up with the generative AI boom.

This comes after the company made generative AI support generally available on Vertex AI in June, giving customers access to Google’s large generative AI models and tooling, including PaLM 2, Codey APIs, Text Embedding API and Imagen.

Today, it announced that Vertex Search and Conversation is now generally available,  allowing developers to ingest data, add customizations, or build a search engine or chatbot “with a few clicks”.

This product will have additional features including multi-turn search, which supports follow-up questions without starting the interaction over, and conversation and search summarization, as well as grounding, which roots generative AI outputs in enterprise data. 

The addition of Vertex AI extensions to Search and Conversation can retrieve information in real time and act on behalf of users across Google and third party applications, while Vertex AI data connectors help ingest data from enterprise and third-party applications such as Salesforce, Confluence, and JIRA.

Further, Google introduced two new products for Vertex AI focused on data science and machine learning engineering.

First is Colab Enterprise, announced in public preview today, enabling data scientists to accelerate AI workflows with access to Vertex AI platform capabilities, integration with BigQuery for direct data access, and even code completion and generation.

Second is a new MLOps Framework for predictive and generative AI to help customers navigate challenges around enterprise-readiness. It includes features that allow developers to tune models based on their specific needs, evaluate the quality of their models, as well as avoid data duplication and preserve data access policies.

Upgrades to Google’s foundation Models

Google announced upgrades to PaLM 2, Google’s large language model. That includes the general availability of 38 languages in PaLM and the possibility for customers to ground responses with their own enterprise data. 

The new version of PaLM 2 for text and chat also supports longer question-answer chats, with 32,000-token context windows, enough to include an 85-page document in a prompt. 

Further, Google announced that output from its code generator model, Codey, will be improved in quality by 25 per cent in major supported languages, for code generation and code chat.

Also, Google improved the visual appeal of its text-to-image foundation model, Imagen, and its recently added capabilities such as image editing, captioning, and visual questions and answering.

Notably, Google announced it is testing a digital watermarking feature to give customers the ability to verify AI-generated images produced by Imagen. This new experimental feature is powered by Google DeepMind SynthID, which, Google says, embeds the digital watermark directly into the image of pixels, making it invisible to the human eye, and very difficult to tamper with, without damaging the image. 

Additions to the Model Garden

Google is bringing Meta’s Llama 2 and TII’s Falcon into its Model Garden, and pre-announcing Claude 2 from Anthropic. Adding these models, Google says, will let enterprises match models to their needs and access full transparency into a model with Meta’s and Falcon’s open-source options.

Nenshad Bardoliwalla, who leads the product teams for Vertex AI at Google Cloud, announced during the opening keynote that Meta’s new Code Llama, unveiled last week, will also be available to all customers and partners in the Model Garden.

Developers and data scientists will also be able to tune these models with their own enterprise data with Colab Enterprise.

Exit mobile version