Google CEO Sundar Pichai on stage in Mountain View, CA at Google I/O. Screenshot: Google I/O livestream
Google on Tuesday announced a slew of new AI-related products, including new ways it will use its Gemini model to improve products such as Google Search and Google Photos.
Why it matters: Google is using its I/O developer conference to highlight AI advances across the company as it looks to demonstrate it is keeping pace with OpenAI, Microsoft and others while defending its stronghold in search.
Here's what Google has announced so far:
- Google DeepMind head Demis Hassabis unveiled Project Astra, Google's vision for a much more responsive AI-powered assistant that can converse naturally, without lag, while taking in a great deal of information. In one impressive example, Astra was able to help a user find her glasses because they'd shown up in recent video clips on her phone.
- It's bringing some improvements to the paid version of the Gemini app, including a more conversational tone and the ability to interrupt the assistant (features that OpenAI announced for ChatGPT on Monday.)
- Gemini Advanced users will also get the ability to create custom chatbots, known as Gems, somewhat akin to OpenAI's custom GPTs.
- Hassabis also announced Gemini 1.5 Flash, a lighter-weight model designed to be faster and cheaper than the original Gemini 1.5.
In Search, Google is launching its AI overviews — which summarize search results — across the U.S. this week, a move that has publishers worried about an inevitable decline in traffic from the search giant.
- In Photos, Gemini will allow users to ask questions and get answers from their photo collections starting this summer. In a separate demo, Google showed the ability to use a video as part of a search query, in this case to get help using a record player.
- In Gmail, Google showed how AI can help corporate and small business users not only organize their inbox, but also pull out data, such as putting receipts into a spreadsheet. That task can then be automated going forward.
- Google also unveiled a new Veo engine for creating AI-generated videos, which will be available to select creators in the coming weeks; and Imagen 3, an update to its text-to-image engine that can better understand prompts and more capably render text.
- On the Android front, Google said it is making Gemini the mobile software's new built-in assistant and making AI-powered search more prominent, among other enhancements coming this year. It's also expanding a "circle-to-search" feature to help with math and other tasks. And it is testing a feature that would use on-device AI, monitoring phone calls in real time, to alert to potential scams.
- For the data center, Google announced Trillium, its sixth generation Tensor processor chip, available to cloud customers in late 2024. Google also said it will be among the first to offer Nvidia's Blackwell chips, starting in early 2025.
Zoom in: Many of Google's announcements focus on increasing the size of the "context window" — the amount of information the model can use in conjunction with a user's prompts.
- Google already allows developers to use 1 million tokens — roughly hundreds of pages of text or an hour of video — and said it will expand that to 2 million tokens, starting with select developers.
- It's also making the 1-million token context available for subscribers of its paid Gemini Advanced service.
Another big focus for Google is transforming AI chatbots into agents that can take action, albeit with human supervision.
- CEO Sundar Pichai shared an example of how someone moving to a new city could use an agent to get information on services in their new neighborhood, change their address across multiple web sites and more.
- "We are thinking hard about how to do it in a way that is private and secure and works for everybody," Pichai said.
Google also unveiled a new watermarking technique for text to identify if it was AI-generated without altering the quality or meaning of the writing.
- Hiding watermarks in text has been a longstanding problem as compared to images, audio, and video where it is easier to add undetectable signals to show that content has been created by AI.
- Google says in its announcement that the new features, which are being added to its existing SynthID tool, will work best with prompts for longer and more varied text, like writing an essay, a theater script or "variations on an email."
The big picture: Google's launches come amid a furious pace of AI announcements across the major tech companies.
- OpenAI on Monday announced its faster, less costly GPT-4o model and a more conversational chatbot.
- Microsoft holds its Build conference next week in Seattle, while Apple is expected to unveil about its generative AI strategy at its WWDC event June 10.
This event is still ongoing. Check back for frequent updates.