Google on Tuesday unveiled several updates to its Vertex AI platform and an updated version of its text-image model, Imagen 2.
At its Google Next ’24 conference in Las Vegas, the cloud provider revealed that its new extended language model (LLM), Gemini 1.5 Pro, is available in public preview on Google’s enterprise AI platform. Google. AI Summit.
Imagen 2, the new version of its image generation model, can now also create 4-second live images from text prompts and has new image editing capabilities.
Vertex AI also now has new grounding features, including the ability to ground answers with Google Search. It also offers new rapid management and evaluation services for large models. Grounding is the extra step of making sure the data is accurate and based on reliable answers and something other than the data a model was trained on.
With these new updates and advancements, Google continues the pattern it established earlier this year of advancing its GenAI technologies despite growing competition.
With its latest updates in Vertex around Gemini 1.5 ProWith the latest version of Google’s Gemini LLM family of models unveiled in February, Google aims to provide businesses with tools to build around Gemini, said Rowan Curran, an analyst at Forrester Research.
“The ability to have a model with such a large context window changes the type of use cases and applications that you can work around,” Curran said, referring to Gemini Pro’s big million dollar deal. pop-up window option.
It’s difficult for companies to build on a model like Gemini if they don’t have tools to quickly manage or test new responses, he added.
“This is the set of tools that they are advancing and introducing specifically to support generative AI,” he said.
The different tools and capabilities support the Gemini family of models and open up a new set of possibilities for how businesses can apply and develop generative AI, he continued.
Updates in Vertex AI
Google is providing support in this regard by including new prompt management and rating services for large models such as Gemini 1.5 Pro in Vertex AI.
The new services allow users to organize, track, and edit prompts for machine learning models.
“This benefit streamlines the process of creating, editing and managing prompts,” said Futurum Group analyst Paul Nashawaty.
Additionally, for businesses looking to build GenAI applicationsrapid management and assessment services will be crucial because of the ability to assess previous prompts and responses, Curran said.
“You need to have that ability to save those queries and responses because there’s no way you’re going to have to accurately regenerate them in the future,” Curran added.
Meanwhile, with new grounding features in Vertex AI, now in preview, users can ground LLM answers with Google search or enterprise data sources using augmented recovery generation (CLOTH). RAG helps optimize the output of LLMs.
The new grounding feature promises to reduce LLM hallucinations, Nashawaty said.
“This is the main breakthrough,” he said. “By doing so, businesses can increase their use of LLMs with confidence. »
Grounding allows companies to ensure that what AI systems understand or interact with in the real world is accurate, said Sid Nag, an analyst at Gartner.
“It’s a bridge between the abstract concepts of AI and practical, tangible results,” he said.
It provides real-world accuracy but adds human sentiment analysis that helps avoid errors that could arise due to simulated data, he added.
The core technology comes as more companies commit to using search technology to support the foundation of generative models, Curran said.
“More and more companies just want to anchor the answers of big language models into their data,” he said.
RAG’s popularity has led to recent developments for Google’s competitors, like Microsoft, which recently unveiled changes to Azure AI Search which allow customers to run RAG at any scale.
Businesses looking to build AI applications in the future will need predictive AI to understand the likelihood of customers taking action, generative AI to understand customer intent based on natural language or images, and search to help them retrieve or find the correct information. he said.
“I see the future as a nice trio or tripod, built on predictivity, generation and research,” he continued.
Other updates to Vertex AI include Gemini 1.5’s professional ability to process audio streams, including voice and audio portions of videos.
Google also revealed that the Anthropic Claude 3 family of models is available on Vertex AI.
Open models such as Llama 2 Mistral 7B and Mixtral 8 are also available in Vertex AI.
Imagen 2 new abilities
Introduced in preview, Imagen 2 includes live text-to-image capabilities that allow marketing teams to generate GIFs and video loops from text prompts. Imagen 2 also has advanced photo editing capabilities.
Imagen 2 comes after Google suspended the new image generation feature for its conversational app Gemini, formerly known as Bard, after the app generated inaccurate images of historical figures.
The benefits of Imagen 2’s live image capabilities include a faster creation process, reduced human error and the ability to minimize tedious tasks, Nashawaty said.
Disadvantages include a steep learning curve for users and the cost of implementation, he said.
And there are similar models on the market.
“This is not a technological breakthrough,” Curran said. For example, Open source tools such as Text2Live the tools have similar functionality.
Stability AI Stable broadcast 3 also has similar abilities,
Text-to-live images can also pose privacy concerns, Nag said.
“I don’t know the limitations or the use of the text-to-live feature from a privacy perspective,” Nag said. “If this feature is limited to certain types of enterprise workloads, that’s a good thing.”
Google also revealed new partner infrastructure and partner news.
- Cloud TPU v5p, the vendor’s next-generation accelerator for training GenAI models, is now available to everyone. It is powered by Nvidia H100 Tensor Core GPUs.
- Nvidia Blackwell GPUs are also coming to Google Cloud.
- Nvidia and Google are collaborating to help startups build GenAI apps and services. Nvidia Inception members now have a path to use Google Infrastructure.
All of Google’s updates show that the race for GenAI continues to heat up.
For companies, it’s more important to focus on GenAI’s applications than the newest product launches, Curran said.
“It’s really important to understand how to build things and what the best practices are emerging,” he said.
Esther Ajao is a news editor at TechTarget and host of podcasts covering artificial intelligence software and systems.
