There I used to be, strolling round my house, taking a video with my cellphone and speaking to Google’s Gemini Live. I used to be giving the AI a tour – and a quiz, asking it to call particular objects it noticed. After it recognized the flowers in a vase in my lounge (chamomile and dianthus, by the way in which), I attempted a curveball: I requested it to inform me the place I’d left a pair of scissors. “I simply noticed your scissors on the desk, proper subsequent to the inexperienced bundle of pistachios. Do you see them?”
It was proper, and I used to be wowed.
Gemini Live will acknowledge a complete lot greater than family odds and ends. Google says it will make it easier to navigate a crowded prepare station or work out the filling of a pastry. It can provide you deeper details about paintings, like the place an object originated and whether or not it was a restricted version.
It’s greater than only a souped-up Google Lens. You speak with it and it talks to you. I did not want to talk to Gemini in any specific manner – it was as informal as any dialog. Way higher than speaking with the previous Google Assistant that the corporate is rapidly phasing out.
Enlarge Image
Here’s a take a look at a part of my dialog with Gemini Live in regards to the objects it was seeing in my house.
Google and Samsung are simply now beginning to formally roll out the function to all Pixel 9 (together with the brand new, Pixel 9a) and Galaxy S25 telephones. It’s out there at no cost for these gadgets, and different Pixel telephones can entry it by way of a Google AI Premium subscription. Google additionally launched a brand new YouTube video for the April 2025 Pixel Drop showcasing the function, and there is now a devoted web page on the Google Store for it.
All you need to do to get began is go dwell with Gemini, allow the digicam and begin speaking.
Gemini Live follows on from Google’s Project Astra, first revealed final yr as probably the corporate’s largest “we’re sooner or later” function, an experimental subsequent step for generative AI capabilities, past your merely typing and even talking prompts right into a chatbot like ChatGPT, Claude or Gemini. It comes as AI corporations proceed to dramatically improve the abilities of AI instruments, from video era to uncooked processing energy. Somewhat much like Gemini Live, there’s Apple’s Visual Intelligence, which the iPhone maker launched in a beta type late final yr.
My massive takeaway is {that a} function like Gemini Live has the potential to vary how we work together with the world round us, melding our digital and bodily worlds collectively simply by holding your digicam in entrance of virtually something.
I put Gemini Live to an actual check
Somehow Gemini Live confirmed up on my Pixel 9 Pro XL a couple of days early, so I’ve already had an opportunity to mess around with it.
The first time I attempted it, Gemini was shockingly correct once I positioned a really particular gaming collectible of a stuffed rabbit in my digicam’s view. The second time, I confirmed it to a pal after we have been in an artwork gallery. It not solely recognized the tortoise on a cross (do not ask me), however it additionally instantly recognized and translated the kanji proper subsequent to the tortoise, giving each of us chills and leaving us greater than a bit creeped out. In a great way, I believe.
This was the primary object I examined with the brand new Gemini Live function, and it impressively acknowledged what it was and what sport it was from (American McGee’s Alice). Every different time I requested Gemini to establish the sport the plush was from, it failed.
In the tour of my house, I used to be following the lead of the demo that Google did final summer time when it first confirmed off these Live video AI capabilities. I attempted random objects in my house (fruit, books, Chapstick), lots of which it simply recognized.
Then I acquired excited about how I may stress-test the function. I attempted to screen-record it in motion, however it constantly fell aside at that process. And what if I went off the crushed path with it? I’m an enormous fan of the horror style — motion pictures, TV exhibits, video video games — and have numerous collectibles, trinkets and what have you ever. How effectively wouldn’t it do with extra obscure stuff — like my horror-themed collectibles?
Initial assessments proved considerably extra profitable than the final, regardless of my giving it a number of hints. Gemini ultimately acquired the sport, Silent Hill: The Short Message, however nonetheless could not give the right title for the determine, touchdown solely on “Cherry Blossom Monster” as an alternative of Sakurahead, which it had accurately guessed a number of instances earlier.
First, let me say that Gemini will be each completely unbelievable and ridiculously irritating in the identical spherical of questions. I had roughly 11 objects that I used to be asking Gemini to establish, and it could generally worsen the longer the dwell session ran, so I needed to restrict classes to just one or two objects. My guess is that Gemini tried to make use of contextual data from beforehand recognized objects to guess new objects put in entrance of it, which type of is smart, however in the end neither I nor it benefited from this.
Sometimes, Gemini was simply on level, simply touchdown the right solutions with no fuss or confusion, however this tended to occur with more moderen or well-liked objects. For instance, I used to be fairly stunned when it instantly guessed considered one of my check objects was not solely from Destiny 2, however was a restricted version from a seasonal occasion from final yr.
At different instances, Gemini could be manner off the mark, and I would want to present it extra hints to get into the ballpark of the precise reply. And generally, it appeared as if Gemini was taking context from my earlier dwell classes to provide you with solutions, figuring out a number of objects as coming from Silent Hill once they weren’t. I’ve a show case devoted to the sport collection, so I may see why it could wish to dip into that territory rapidly.
This was the toughest of my assessments. I requested Gemini to establish not solely what sport this nonetheless was from (Silent Hill 2), however what iconic quote the particular person on the prime of the steps mentioned. Gemini nailed the sport, the characters, and half of the quote on the primary spherical; it took two extra guesses to complete the quote, “You see it, too? For me, it is at all times like this.”
Gemini can get full-on bugged out at instances. On multiple event, Gemini misidentified one of many objects as a made-up character from the unreleased Silent Hill: f sport, clearly merging items of various titles into one thing that by no means was. The different constant bug I skilled was when Gemini would produce an incorrect reply, and I’d right it and trace nearer on the reply — or straight up give it the reply, solely to have it repeat the inaccurate reply as if it was a brand new guess. When that occurred, I’d shut the session and begin a brand new one, which wasn’t at all times useful.
One trick I discovered was that some conversations did higher than others. If I scrolled by means of my Gemini dialog record, tapped an previous chat that had gotten a particular merchandise right, and then went dwell once more from that chat, it could have the ability to establish the objects with out situation. While that is not essentially shocking, it was fascinating to see that some conversations labored higher than others, even when you used the identical language.
Google did not reply to my requests for extra data on how Gemini Live works.
I needed Gemini to efficiently reply my generally extremely particular questions, so I supplied loads of hints to get there. The nudges have been typically useful, however not at all times. Below are a collection of objects I attempted to get Gemini to establish and supply details about.
For this one, I simply requested Gemini what it noticed. “OK, I see a black and white cat that is basking within the solar on a hardwood flooring. The cat is stretched out in a humorous place. There is a inexperienced rug with ‘Home is the place the..’ written on it.” I requested Gemini to guess once more, and I acquired responses from “house is the place the horror is” to “honor,” however it will definitely landed on the right reply (simply the one phrase, “horror”).
Gemini gave me 4 improper characters from the precise sport earlier than accurately figuring out this iconic Bioshock Infinite character, Songbird.
Gemini nailed this creepy determine on the primary guess. (Twin Victim, Silent Hill 4: The Room)
No fuss — Gemini accurately acknowledged Mira from Silent Hill 2, the actual one accountable for the city
This one impressed me. While Gemini may “see” that this was a Silent Hill map, it nailed the truth that this was a limited-run print that was part of an ARG that occurred final yr.
Gemini took a wildly totally different method to figuring out this jacket from Silent Hill 2. It requested 24 particular questions based mostly on the knowledge I gave it, with my first trace being that it was from a online game. However, by the nineteenth query, it appeared that it already knew precisely what sport it was from by the particular questions it was asking me.
This one did not take lengthy, however Gemini initially advised that this portrait is perhaps of American creator and poet John Ashbery. Once I moved the digicam nearer to the picture and mentioned it was from a TV present, Gemini replied accurately, “That’s the Log Lady from Twin Peaks, holding her well-known log.”
This was a straightforward one for Gemini. It instantly acknowledged this as a limited-edition tarot deck that needed to be “earned” by enjoying by means of a particular seasonal occasion in Destiny 2.
