Transcription
[Music] Google started rolling out Project Astra in Gemini Live. The move brings real-time visual AI, allowing multilingual conversations about anything seen and heard via a phone's camera or screen sharing. It's coming to Pixel 9 and Samsung Galaxy S25 devices, and it's already shaking things up with promises of advanced camera-based analysis and expanded AI-driven interactions.
What's happening behind the scenes is that Samsung and Google have joined forces, letting owners of the Galaxy S25 experience Gemini Live's real-time visual AI without an additional subscription to Gemini Advanced. There's some speculation about whether it's a straightforward permanent perk or a free trial that just happens to be bundled with new devices, but the bottom line is that S25 owners get to open their camera, point at anything, and watch Gemini Live interpret it on the spot. It could be as simple as analyzing how many fingers you're holding up or identifying what street you're on as your car passes through a city.
The immediate feedback from early users is that the feature behaves more like a snapshot tool than a continuous live feed. If you ask Gemini Live how many fingers you're showing, it'll take a photo at the moment you query and answer based on that static frame. If you move your hand around afterward, it won't keep updating the answer. Still, it's an impressive step forward in how your phone might one day blend a lens-like object recognition system with a more human-sounding AI conversation.
Some folks with Pixel 9 devices are getting a similar roll-out, though those who want the advanced real-time camera mode may need Gemini Advanced. It's not entirely clear if there's a separate promo code or if certain Pixel 9 owners are automatically granted full access, but either way, the lineup is expanding.
There's also a new dimension of screen sharing. You can hop into the Gemini app, switch on share screen, and let the AI see whatever you're looking at. Maybe you're reading a news article and want the AI to summarize it or point out more details, or you're browsing an online store and want immediate feedback on product comparisons or discount codes. You can even switch to your camera while screen sharing so the AI sees a live feed of your surroundings. The minute you lock your phone or pause the session, the camera shuts off. You can easily toggle it on again, but there's a built-in safeguard so you don't walk around broadcasting everything by accident.
This concept is tied to Google's new AI mode in its search platform, which was first limited to Google 1 AI premium subscribers but has since expanded to more Labs users in the US. The search giant says the feedback so far is largely positive and that people are issuing longer, more nuanced queries than they do with classic text-based searches.
One factor behind that surge might be the introduction of multimodal capabilities. Google's new approach, fueled by a version of its Gemini model, can analyze text and images at the same time, letting you either take a photo or upload one to get a meaningful response. Think about showing it a screenshot of a new sneaker or of a complicated pie chart for work. AI mode works in tandem with Google Lens, which identifies objects in the image and then runs what Google calls a "query fan-out technique." Essentially, it breaks the image down into multiple subqueries. If you're showing it a shelf of books, the system identifies each individual title, queries data about them, and combines the results in a cohesive answer, like a curated reading list with direct purchase links.
Google hopes this deeper, more contextual approach will turn the AI into something that feels like a personal assistant, not just a search box.
While this is exciting for end-users, it's also sparking concern among independent publishers. Some site owners say traffic's dropping off since Google's AI can provide thorough summaries right on the spot, meaning fewer clicks heading back to the original content. One person with a home improvement site noticed that the AI sometimes suggests products that don't actually exist or recommends questionable safety practices. Google claims there's no definitive proof that AI overviews are the reason for a broader traffic decline, arguing that search algorithms shift all the time, users' interests evolve seasonally, and many other factors can also cause hits to site visits. Still, the worry remains that the more Google's AI refines its ability to deliver direct answers, the more small sites will see their organic traffic dwindle.
Competition for talent is another hot topic. DeepMind, which Google acquired, is one of the top-tier AI divisions on the planet. Recently, it's been reported that they use strict non-compete agreements in the UK. Some employees can't jump over to rivals like OpenAI or Microsoft for up to a year. The employees aren't entirely powerless because they're still getting paid, but they're also essentially sidelined during that period, which can be frustrating if they want to keep pushing AI boundaries. The FTC in the US cracked down on non-compete agreements, so you wouldn't see this practice as often stateside, but in the UK, it's still fair game, and that can slow down the movement of AI experts who are eager to try out new roles elsewhere.
Meanwhile, Google is doubling down on building out even more robust AI technologies under the Gemini umbrella. Gemini 2.5 Pro just entered a wider public preview with higher usage limits via the Gemini API and Google AI Studio. Pricing is tiered to attract developers in search of large-scale computing. If you're under 200,000 tokens per month, you pay $1.25 for every million input tokens and $10 for every million output tokens. Once you exceed that cap, the rates jump to 2.5 and 15 respectively. Tokens can refer to bits of text or other data such as image or audio inputs, and the output tokens reflect the complexity of the reasoning. Google insists this is still cheaper than what some big-name competitors charge, pointing out that Anthropic's Claude 3.7 sonnet or OpenAI's GPT models can quickly become more expensive. Google hasn't made the new Gemini model available in Vertex AI just yet, but it's on the roadmap.
Despite its aggressive pursuit of market dominance, Google does openly acknowledge that some forms of advanced AI, especially AGI, could pose a real threat if misused. DeepMind published a research paper admitting that AGI might lead to substantial harm, dividing the risks into misuse, misalignment, mistakes, and structural flaws. It also outlines a variety of defenses like restricting access to very high-level AI capabilities, adopting extra security measures, and building in constant monitoring. Demis Hassabis, who leads DeepMind, has said that if we do build powerful agentic systems, they could be transformative for science, medicine, and pretty much every field. However, that power also means they need major oversight. He believes it's essential for society, economists, and governments to think carefully about how we want the future to look once AGI technology is fully developed.
All these new features—camera feed analysis, screen sharing, and smart search—are slowly turning your phone into a real-time assistant. You can point your camera at something in a store, share your screen while browsing, or snap a photo of a landmark, and Gemini will try to help instantly. For Galaxy S25 users, it all works without needing a Gemini Advanced subscription, at least for now, which might push other brands to follow. Of course, some worry that AI summaries will replace clicks to real websites or that smaller creators will get buried, but Google's pushing forward, expanding AI mode and rolling out Gemini 2.5 Pro with cheaper rates to attract developers. And with Apple, OpenAI, and others in the game, this AI race isn't slowing down.
Project Astra is now live inside Gemini, and while the camera feed isn't fully continuous yet, it's already a big step toward a more aware AI. It still has quirks, but if you've got a Pixel 9 or Galaxy S25, you can try it right now and see where this all might be heading.
That's it for this one. Hope you enjoyed it. If you haven't already, hit that subscribe button. Thanks for watching, and I'll see you in the next video.