Google brings Gemini to Search, mobile and video creation

Google turned its I/O conference this Tuesday into a showcase of its reach: Gemini is no longer a standalone chatbot, but is moving into Search, Android, Workspace and developer tools. Alongside that integration, the company unveiled Project Astra, an assistant that sees and listens in real time, and Veo, its new video-generation model.

The announcement matters because of Google’s scale. OpenAI popularized conversations with generative models, but Google controls products used every day by billions of people. Bringing Gemini into Search could change how a large share of users find information online.

Search starts answering, not just linking

Google has begun rolling out AI Overviews in the United States: AI-generated answers that appear above traditional links to resolve complex queries. The feature grew out of Search Generative Experience, the experiment the company had been testing in Search Labs since 2023.

The idea is for Search to combine information from several pages and provide an initial explanation, with links for further reading. Google expects to bring AI Overviews to more countries over the course of the year and reach more than one billion users before the end of 2024.

The change could save steps in comparison searches or tasks involving multiple constraints, such as planning a meal for a group with dietary restrictions. But it also raises the familiar risk posed by generative models: an answer can sound convincing while containing errors. For media outlets and specialized websites, the question will be whether these summaries send readers to the sources or keep the query inside Google.

Project Astra: an assistant that interprets what’s in front of it

Project Astra is a multimodal agent prototype developed by Google DeepMind. A multimodal system processes several types of information, including text, speech, images and video. In demonstrations, Astra followed what a camera was showing, identified objects, answered spoken questions and remembered where it had seen an item in a room.

Its goal is not merely to describe an image sent by the user, but to hold a conversation about what it is seeing through the camera. Google has also shown a glasses-based version, although that was a research demonstration, not an announced commercial product.

The company plans to gradually bring Astra’s capabilities to the Gemini app during this year. The challenge will be making that visual memory useful without turning an always-on camera into a privacy problem. Google will need to clearly explain what data is processed, how long it is retained and what user controls apply.

Veo and Imagen 3 raise the creative stakes

Google DeepMind has unveiled Veo, a model capable of generating videos at 1080p resolution lasting more than a minute from text prompts, images or reference videos. It can also interpret filmmaking terms such as camera movements and visual styles.

Veo will compete with systems such as OpenAI’s Sora, announced in February but still not generally available to the public. Google will initially offer it in private preview to selected Vertex AI customers, through its enterprise artificial intelligence platform.

The model includes SynthID, Google’s technology for embedding an invisible digital watermark in AI-generated content. It is a useful measure for aiding the technical identification of synthetic videos, although it does not by itself eliminate impersonation risks or guarantee that a manipulated clip will retain the signal as it circulates online.

Imagen 3, the new version of Google’s image generator, is also arriving. The company highlights improvements in visual quality and in generating text within images, one of the most visible shortcomings of earlier tools. It will first be available to select creators through ImageFX and to Vertex AI customers.

Gemini 1.5 Flash targets speed and lower costs

For developers, Google announced Gemini 1.5 Flash, a lighter version of Gemini 1.5 Pro. It is designed for tasks that require fast responses and a high volume of requests, such as summarizing text, extracting data or running customer-service assistants.

Flash retains a context window of up to one million tokens. Context is the amount of information a model can keep in mind while responding; one million tokens is roughly equivalent to hundreds of thousands of words. Google is making it available in public preview through AI Studio and Vertex AI.

Meanwhile, Gemini 1.5 Pro with one million tokens is becoming available to Gemini Advanced subscribers, while the company is opening a private preview of a two-million-token version. The capacity makes it possible to analyze extensive collections of documents, code or recordings, but it does not replace human review: a model’s ability to read more material does not mean it will interpret every detail without making mistakes.

Google has not introduced a single new application, but a strategy: making Gemini the AI layer across its products. Execution will be decisive. Integrating these features into everyday services could normalize generative AI far more than a flashy demonstration, but it will also expose its failures to a much larger audience.