Google's AI is getting smarter

Project Astra's interactive interface. Photo: Google .

At the I/O 2025 event held in the early morning of May 21 (Vietnam time), Google announced a number of new features of Project Astra. First launched in 2024, Project Astra uses a multimodal AI model to exploit and analyze images and visual information in real time. This is Google's great ambition with the goal of making AI more useful in life.

After a testing period, some Project Astra features like voice feedback, memory and basic calculations will appear on Gemini and Google apps.

During the event, Google demonstrated some of the ways Astra can help with everyday life. In addition to answering questions and using the camera to remember where you put your glasses, Astra can now proactively complete tasks without you having to ask.

“Astra can choose when to speak based on what it sees. It constantly observes and responds accordingly,” said Greg Wayne, a researcher at DeepMind.

According to The Verge , the feature could change the way AI works. Instead of waiting for users to point their cameras and ask for help, Astra can continuously monitor, listen, and automatically assist if necessary. Currently, Google is focusing on developing Astra for smartphones and smart glasses.

Wayne gives some examples of how Astra works. While doing homework, the tool can monitor and immediately alert users when mistakes are detected, rather than waiting for the user to finish and ask for a check.

Demis Hassabis, CEO of DeepMind, said that training Astra to run tasks on its own was one of the most difficult technical skills. He called it “situation reading,” because the AI ​​needed to determine when to help, how to help, and when to stay quiet. These are actions that humans do quite well, but are difficult to quantify and study.

Astra’s next upgrade is to mine user information on the web and Google products. According to The Verge , the tool can know the schedule to notify when to leave the house, read the content of emails to extract phone numbers when checking in at hotels…

Project Astra's Highlighting feature, which helps scan objects and compare information extracted in other applications. Photo: Google .

The final piece of Astra’s puzzle is its ability to learn how to use an Android phone. Demonstrating at the event, Bibo Xiu, a representative of DeepMind’s product division, pointed the phone’s camera at a pair of Sony headphones and asked Astra to identify the model.

After Astra replied that the headphones could be Sony WH-1000XM4 or WH-1000XM3, Xiu continued to ask the tool to find the user manual, then explained how to connect to the phone. After Astra instructed, Xiu interrupted and asked the chatbot to connect for him. Astra immediately opened the settings and did it himself.

The process wasn't entirely seamless as Xiu had to allow Astra to monitor the phone screen. In the future, Google is aiming for the chatbot to be able to monitor screens automatically.

In another Google video , a person asks Astra to find a Huffy bike manual, then navigate to a specific page with instructions for fixing a bike part. Astra can even find a YouTube video on how to remove a stripped screw, then contact the nearest store to find a replacement part. It all happens seamlessly in the demonstration.

Although Astra has not been widely released, Google said that the Gemini Live image analysis feature will be released to all iOS users (previously only supported Android).

In the future, Google's AI won't just act on command, but Astra will help "understand what can and cannot be seen at the moment," Xiu said.

Source: https://znews.vn/ai-cua-google-ngay-cang-thong-minh-post1554615.html