What did Google announce at I/O 2024, and what changes for you?

Yesterday, Tuesday, May 14, 2024, Google held the opening keynote of I/O, its yearly conference for developers. Sundar Pichai, CEO of Google and its parent company Alphabet, opened with a short summary of where the company stands: "Google is fully in our Gemini era." Gemini is Google's family of AI models, and almost every announcement that followed was a way of putting it to work.

For most people, the biggest change is in a product they already use every day: Google Search.

The announcements at a glance

Pichai and his colleagues covered a lot of ground. Here are the main items, with who can use them:

Google also previewed Gemini Live, for spoken conversations with Gemini; a version of Gemini Nano, its small on-device model, that understands images and sound on Android; a new generation of its open Gemma models; and an expansion of SynthID, its invisible watermark for AI-generated content, to text and video.

AI Overviews: how AI answers arrive in Search

For the past year, Google has tested AI answers in Search through an experiment called the Search Generative Experience, open to people who signed up. Pichai says it has answered billions of questions that way. Now the feature, renamed AI Overviews, is rolling out to everyone in the US, with more countries to follow.

Here is what happens when you search. For complex questions, a Gemini model built specially for Search reads relevant pages and writes a short summary at the top of the results, with links to the sources so you can read further. Liz Reid, Google's head of Search, said that the model combines current information, Google's ranking of pages, the ability to read long material, and understanding of images. She also said overviews won't appear when ordinary results already do the job; they are meant for questions whose answers are scattered across many sites.

TechCrunch reports that Google expects the feature to reach hundreds of millions of US users this week and more than a billion people before the year is out. Google says people clicked on a wider variety of websites during the test, but publishers and critics worry that fewer visitors will reach the sites the summaries draw on.

Veo and Imagen 3: video and pictures from text

Both new creative models are generative: you describe what you want in words, and the model creates it.

Veo generates video in 1080p, which is full high-definition, in many cinematic and visual styles, and Google says clips can run past a minute. It understands film vocabulary such as "timelapse" or aerial shots of a landscape, and Google says it keeps people, animals and objects consistent as they move. Filmmaker Donald Glover and his studio, Gilga, tested it for a film project. Every Veo video made in Google's VideoFX tool carries a SynthID watermark.

Imagen 3 produces more detailed, photo-like images than earlier versions, with fewer odd glitches, follows long prompts more closely and handles text inside images better, a known weakness of image generators. Google suggests uses like birthday messages or title slides.

A more specific prompt usually gets a more useful result from models like these. Compare:

Vague:    A city at night.
Specific: Timelapse of a rainy city street at night, neon signs reflected
          in puddles, slow aerial shot moving forward, cinematic lighting.

The second version names the subject, the setting, the camera movement and the style, which are exactly the details Google says Veo is designed to follow.

A faster model and a longer memory

Gemini 1.5 Flash is aimed at developers who need quick, inexpensive answers at large volume, such as chat apps or summarizing many documents. Think of it as the quick, affordable member of the family: a little less capable than 1.5 Pro, but built to answer fast and in bulk while keeping the long memory.

Gemini 1.5 Pro, meanwhile, gets better at translation, coding and reasoning. Its context window, the amount of material it can consider at once, stays at 1 million tokens for all developers (a token is a word or part of a word; that is enough for hundreds of pages or about an hour of video), with 2 million available in a private preview.

Why it matters

For everyday users, AI Overviews mean that AI answers stop being something you choose to try and become part of the default Search page for many people in the US. That makes it more important than ever to check the links behind a summary when the answer matters.

For creators and developers, Veo and Imagen 3 put Google firmly in the race for AI video and images, while Flash lowers the cost of building with Gemini. Many details, including prices and wider access for the creative models, are still to come.

Which Google I/O 2024 announcements can you try right now?

If you are in the US, you may already see AI Overviews in Google Search, since the rollout begins this week; it appears only for some questions. If you subscribe to Gemini Advanced, you can try the updated Gemini 1.5 Pro with its 1-million-token window, which lets you upload long documents and ask questions about them.

Developers can test Gemini 1.5 Flash and the updated 1.5 Pro in Google AI Studio or on Vertex AI, Google Cloud's platform for AI. For Veo and Imagen 3, you can join the waitlists through Google's VideoFX and ImageFX tools, though early access goes to selected creators. Ask Photos, Gemini Live and Project Astra are announced but not yet available, so for now they are worth watching rather than trying.