{
  "video_id": "chs2Xx4JbBs",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "AI Dev 26 x SF | Paige Bailey: What's New and What's Next in AI",
  "duration_seconds": 2558,
  "url": "https://www.youtube.com/watch?v=chs2Xx4JbBs",
  "upload_date": "20260520",
  "transcript": "My name is Paige. I lead engineering for\ndeveloper relations at Google DeepMind.\nI'm based in the Bay Area. Um, and uh, I\ndon't think it's a secret that Google\nhas been a little bit busy over the\ncourse of the last year, especially over\nthe course of the last couple of months.\nIt feels like we've been releasing a new\nmodel, new products, new features every\nsingle day of the week. Um, most\nnotably, Gemini 3.1 Flash Live, which\ngives you the ability to talk to the\nmodel in real time. Um, Gemini 3.1 Pro,\nwhich is our largest model in\nFlashlight, which is our smallest um\nthat's available via the API. Um, but is\nvery, very powerful. Um, Nano Banana 2,\nwhich gives you the ability to create\nimages, edit images, create images and\ntext interled. Embeddings. Our new\nembeddings model allows you to use video\nand audio and images and text all in the\nsame embedding space. Liria 3 for music\ngeneration. Genie3 for world model\nbuilding. Our full stack runtime for AI\nstudio which incorporates OOTH and\ndatabase support for generating apps.\nGemma 4 for open models, VO3.1 Light for\ngenerating realistic and uh video with\naudio at a fraction of the cost of our\noriginal VO3.1 model and Gemini 3.1 text\nto speech which gives you the ability to\nhave really really hyper realistic\nsounding um text to audio outputs uh for\na variety of languages and a variety of\nstyles. Um, so all of this all of these\nthings that you see on this slide have\nbeen released over the course of the\nlast few months. Um, and I think I'm\nparticularly impressed with how it spans\nnot just the gambit of kind of large\nlanguage models, but also multimodal\nmodels, uh, and generative models in\nterms of images and video and audio as\noutput. Um, so it's really wild to see\nthe team cook like this. I think it's\nit's uh notable. It's notable certainly\nin my tenure at uh Brain and at Deep\nMind. For folks in the audience who\nhaven't had a chance to experiment with\nthe Gemini models so far, um they're\nnatively multimodal, which means that\nthey can understand video and images and\naudio and text and code, all of the\nabove all at once. But they can also\noutput multiple modalities. So they can\noutput text, they can output code, kind\nof similar to other models that you\nmight have worked with, but they can\nalso output audio tokens. Um they can\noutput uh frame by frame uh images that\nyou can stitch together into a video. Um\nthey can output images and text interled\nand they can also do a broad spectrum of\nother tasks when you incorporate\nfunction calling um as well as some of\nour native tools like code execution and\ngrounding with Google search. Um, so I\nthink it's a lot more fun though to to\nkind of show rather than tell. And I'm\nuh I'm hoping that the the Wi-Fi gods\nare appeased.\nUm, so I'm going to head over to Google\nAI Studio and we're going to take a look\nat some of these models and what they're\nable to do. Um, I'm also going to be\nshowing a broad spectrum of models\ntoday. So both our Gemma open model\nfamily which you can ping via API and AI\nstudio but you can also download the\ndownload the models themselves from\nhugging face. I'll show a Gemini 3.1\nflash live which is the model that gives\nyou real-time interactions. Um Gemini\n3.1 flashlight\nand then also do some head-to-head\ncomparisons.\nUm so how many folks have played with\n3.1 flashlight before? Few hands. If you\nhaven't, um, you can kind of select it\noff to the right in AI Studio. You can\naccess AI Studio with just your personal\nGmail account for free using a.dev or a\nstudio.google.com.\nYou can turn on and turn off different\ntools off to the right. So, you can see\nthings like code execution, which gives\nGemini a sandbox Python environment in\norder to do its work. things like\nfunction calling, Google search\ngrounding, grounding with Google Maps,\nURL context, which is kind of like\nretrieval for free based on certain\nnumber of URLs. And if I click this plus\nsign, um you can see that I can also add\na whole bunch of different data types.\nSo things from drive, I can upload files\ndirectly, I can record audio, add\nimages, add YouTube videos, add sample\nmedia. And for this, let's go ahead and\ntry adding a YouTube video. Um, so I\nlove dinosaurs.\nUm, so I'm gonna look for a dinosaur\nvideo. Um, maybe try to find one. An\neducational dinosaur. Uh, TV for kids.\nThis looks fun.\nUh, the carnivorous.\nUm, I'm going to head back over to uh to\ngrounding. I'm going to paste in the URL\nfor this YouTube video. It's about an\nhour long, which is pretty lengthy. So,\nI'm going to set the start time to be at\nthe very beginning and set the end time\nto be about 5 minutes in. It defaults to\none frame per second. I'm going to hit\nsave. And then I'm going to say um\ncreate a table with timestamps\nfor all of the dinosaurs\nthat you see in this video. Um, make\nsure to include a fun fact about each\ndinosaur. Uh, and then I'm going to hit\nrun. And you can see that the first five\nminutes of this video sampled at one\nframe per second. So every single frame,\nall of the audio tokens associated with\nthat first 5 minutes. um the the kind of\nmetadata associated with the video ends\nup being around 39,900\ntokens which was well within the 1\nmillion token level um for the Gemini\n3.1 flashlight model. You can see the\ndinosaurs getting added on the left um\nthe timestamp in which they were seen,\nthe fun fact about each one, and then\nalso the cited sources for each one of\nthose dinosaurs. Since I turned on\ngrounding with Google search, um, I'd\nset the thinking level to minimal, so\nthe the output happened relatively\nquickly for this 5 minutes long of video\nfootage. And then if I hit get code, it\ngives me all of the code that I would\nneed to replicate whatever I just did in\nAI Studio. So you can see the model that\nI selected. You can see that I turned on\nsearch grounding as a tool, which was\njust a oneliner. You can see that my\ngeneration config was set to minimal.\nUm, and as I change and modify these\nfeatures within AI Studio, um, all of\nthose changes are expressed in this get\ncode category. So the TLDDR is that if\nyou get it working in AI studio, you can\nget the code to use it for your project\nin Python and TypeScript and Java or\nwhatever your favorite language might\nbe, which is pretty wild. Um, you can\nalso toggle on the outputs. So if you\ndid want to have kind of the the outputs\nthat the model had it expressed or kind\nof the the system instructions or some\nof the parts of the conversation, you\ncan toggle that on. You can also if\nyou've got it selected to be Python, you\ncan export everything directly to Collab\nor you can also uh create and invoke a\nGemini API key to use for your project.\nUm,\nif you uh kind of hit this little\nnavigation menu off to the left, you can\nalso see that you can generate API keys\ndirectly from AI Studio. So, you don't\nhave to kind of like wrangle around with\nthe the AI consoles\num in various cloud service providers in\norder to do more or less the same thing.\nUm, so if you're looking for a fast path\nto getting an API key, including an API\nkey that you can use for free, um, AI\nStudio is a good place to go. Um, so\nthis is this is AI Studio. I also really\nlike um, one of the under uh underlod\nfeatures in AI Studio is this thing\ncalled compare mode where you can kind\nof go in, you can compare two different\nmodels on the same task. Um, I'm going\nto go ahead uh and move to the next tab,\nwhich is showing Gemini 3.1 flashlight\npreview. Um, I'm going to make sure that\ncode execution is turned on for both it\nand Gemini 3 flash preview. Um, and then\nI'm going to go ahead and try to look\nfor\num just a picture of Lego bricks and\npieces\non Google search.\nLet's see if I can find a good one. This\none looks really nice.\nI'm going to copy it.\nGo back to um our UI, paste it in. And\nthen just say something like draw\nbounding boxes\naround all of the green Lego bricks\nusing Python.\num make sure to uh and then maybe\ndisplay the image with bounding boxes.\nI'm going to hit run and we immediately\nsee the two models go head-to-head to\naccomplish this task. Gemini 3.1\nFlashlight got it right kind of right\nout of the gate. Um which is pretty\ncool. But I think the even cooler thing\num as we watch Gemini 3 Flash kind of\nkeep thinking through the task is that\nif I hover over the cost to do this um\nit ends up being only a very very small\nfraction of a penny in order to do the\nbounding boxes around these objects. You\ncould also say like draw bounding boxes\naround all of the cars, draw bounding\nboxes across all of the like Nissan\nUltimas. um draw bounding boxes around\nall of the Ford Mustangs. Um try to\nthink of how many cars I know so my\nrelatives won't disown me. Um but uh but\nyou can kind of incorporate the model's\nreasoning capabilities to uh to kind of\nanalyze these input images and it ends\nup being a small small component of the\ncost if you incorporate these tools like\ncode execution with a smaller lighter\nweight model. Um, I think it's really\nreally nice as well that you can do\nsegmentation masks. So you could instead\nsay like um do a segmentation mask for\nall of the green pixels, do a\nsegmentation mask around the persons in\nthe image um or similar. And just like\nwe did with that uh that dinosaur\nexample, if I hover over the tokens for\nthe dinosaur example, you can see\nanalyzing that 5 minutes of YouTube\nvideo ended up being around a penny and\na half, which is wild, right? Like you\ncan analyze a full video with\nunderstanding the frames, understanding\nthe audio, understanding the background\ncontext for just pennies on the dollar.\nAnd you could have asked for it to\ntranscribe all of the text. You could\nhave asked for it to be segmented into\nchapters. You could have asked any\nnumber of questions and it would still\nbe around the same number of tokens and\nput it out. Um so strongly strongly\nrecommend playing around with Gemini 3.1\nFlashlight, Gemini 3 flash um and using\nthem uh using them for your work. The\nother thing that I want to show is\nGemini 3.1 flashive. Um how many folks\nhave heard of flash live?\ncouple of hands. So, what this does is\nit gives you the ability to have a\nreal-time conversation with the model,\nto share your desktop, to share an input\nvideo feed, um, and then to incorporate\nthese tools like reasoning,\num, as well as grounding with Google\nsearch, um, function calling, etc. off\nto the right. So, I'm going to go ahead\nand turn on grounding with Google\nsearch. We have some different voices\nthat you can select. So, Puck um\ngot a project in mind.\nWhat do you want to explore?\nReady to make something amazing?\nI kind of like Zephr, though. So, we're\ngoing to stick with him. Um and then you\ncan also modify system instructions,\nshare your screen, um and the likes. So,\nI'm going to share my screen really\nquick\nand select one of my Chrome tabs. So\nmaybe this uh maybe the bounding box is\none because it's got some nice code on\nit.\nAnd I'm going to go back to AI Studio\nwhere my screen is being shared and say,\n\"Hey Gemini, what do you see on the\nscreen?\nI see snippets of Python code for\ndetection and bounding box drawing,\nlikely for an image processing project\ninvolving green Legos. There are also\ncomparisons between Gemini 3.1 Flash and\nGemini 3 Flash preview. What specific\naspect would you like to know more\nabout?\nAnd so you can see that the model was\nable to take the input image, describe\nit, kind of sus out what the general\ncode was about, um, and then also give\nme an output response. Um, but it also\nworks in languages that aren't English.\nSo, does anybody in the audience speak a\nlanguage that is not English?\nI'm gonna pick somebody close to the\nfront.\nHey, Gemini, could you tell me the same\nthing that you just said? Um, but could\nyou tell me in Spanish, please?\n3.1 flashi\nflash.\nOkay, excellent. And then uh another\nlanguage from somebody in the audience.\nDo you want to\nHindi? Yep.\nUh, and Gemini, could you tell me what\nthe weather is like in San Francisco\ntoday, but could you tell me in Hindi?\nSan Francisco.\nAnd was that was that okay? Cool. Cool.\nCool. Excellent. So, uh, you can ask for\ndynamic swapping between different\nlanguages. You can share your screen,\nbut you can also share a video feed. So,\nas an example, if I wanted to share a\ncamera source,\num, it would pick up on me and then I\ncould say something to the effect of,\n\"Hey, Gemini, could you tell me um how\nmany fingers I'm holding up?\" And also\nexplain what you see on the screen.\nYou're holding up two fingers, like a\npeace sign. Behind you is a screen\nshowing what looks like a live video\nfeed of that. Are you testing something\nor was there something specific you\nwanted me to notice on the screen?\nWell, it didn't have to be so snippy,\nbut the uh but the uh it's great to be\nable to to kind of have this flexible\nthis flexible way of speaking. And then\nalso, if you wanted to modify the system\ninstructions,\num you would be able to do that as well.\nSo say you didn't want the model to\nrespond in any arbitrary language. You\nonly wanted the model to respond to the\nuser\num in German. Um then you could say\nsomething to the effect of\nhey Gemini, could you tell me about um\nthe AI dev conference that's happening\nin San Francisco and then also a little\nbit about deep learning.ai AI\nthe AI dev conference in San Francisco\nconference\ndeep learning.ai\nIt accent is a lot better than mine.\nYeah. And you can also see the cited\nsources because I had turned on Google\nsearch grounding. Um, and again, just\nlike uh all of the other things in AI\nStudio, if I click get code, it gives me\nall of the code that I would need to\nreplicate what I just did. Um so the\nspecific model that I selected the uh\nthe kind of config settings the oneliner\nto turn on Google search grounding and\nthen also the system instructions that I\nhad added into the into the content\nstream as well. So again tlddr if you\ncan get it working in AI studio you can\nget the code that you would need in\norder to incorporate it into your\nproject which is pretty rad. the um so\nwe've done Gemini 3.1 flashlight, we've\ndone 3.1 flashlight preview. I'm also\ngoing to talk a little bit about one of\nmy most favorite things\num that we've released just recently\nwhich is Gemma 4 um which uh our open\nmodel family and hopefully everybody in\nthe audience has open source near and\ndear to their hearts. Um, Gemma is the\nlatest in our open family of models. It\nhas four sizes. So, a two billion\nparameter version, a four billion\nparameter version. The two billion\nparameter version is actually small\nenough to run on mobile devices, which\nis pretty rad. Um, 4 billion parameters\ncan fit well on your laptop. Um, and\nthen we also have a couple of um, uh,\nkind of moderatesized versions that are\navailable within AI Studio that you can\ninvoke via the Gemini APIs or you can\njust download themselves and use them\nfor your company. Um but they're Apache\n2 licensed which is uh which hooray like\nif you uh if you saw the license for\nprevious iterations of Gemma um it was\nnot Apache 2 but now we've got it into a\nstate where it can be uh you know easily\nused for for building out your building\nout your companies. Woo. Thank you.\nThank you for Yeah. Open source. Um I\nowe my entire career to open source. So\nI I really really love that. Uh I really\nreally love that Google is investing in\nit. Um the model performance is actually\nquite good. So you can see that our 26\nbillion mixture of experts model and our\n31 billion uh dense model are performing\nvery very well in the chatbot arena. Um\nand they're also really modest in size.\nSo the GPU footprint required in order\nto run one of these models is something\nthat's tractable for most companies as\nopposed to, you know, if you see a model\nthat's got a trillion plus par\nparameters, it's going to be uh you\ncan't really like get enough GPUs in\nyour closet to uh to actually run it\neffectively. So modest modest size in\nterms of GPUs. Um they also support\naudio, video, images similar to how you\nwould interact with the Gemini family of\nmodels. Many of the same training um\nprocesses are invoked by the by the\nGemma family. Um I've selected the\nmixture of experts here off to the right\njust so you can uh see it in action. And\nif I add one of these sample images,\nmaybe a couple,\num, you can say something to the effect\nof describe both of these images.\nUh, and then also write a limmerick\nabout them. Um, turn the thinking level\nto minimal so it runs a little bit more\nspeedily.\nWe can see a tabby kitten,\na small fluffy puppy, and then a kitten\nwith stripes, etc. Um, same as always,\nif you click get code, it gives you the\ncode that you would need to run it\nyourself. And, uh, developers on the\nfree tier get 10,000 or more requests\nper day to the Gemma models via API. Um,\nso though we don't advertise that a ton.\nUh, so, uh, so I'm advertising it to\ny'all. Consider yourselves, uh, consider\nyourselves aware. Um, you get a whole\nbunch of free tier API calls to the\nGemma family of models, which are solid,\nwhich can do all of these multimodal\ntasks. Um, and it's zero dollars to you,\nrelatively fast. Um, and, uh, we're\nreally excited to see how people how\npeople use these models.\nWe also have a new feature within AI\nStudio called build which allows you to\ncreate an app um uh just via voice. So\nyou can kind of describe the app that\nyou want to see in your world. Um you\ncan also click I'm feeling lucky uh and\nhave uh and have them uh you know Gemini\nkind of create an example app for you.\nAnd then we do also have uh or we just\ncreated tab completion. So if you can't\nthink of additional ideas,\nyou can just have Gemini create them for\nyou. Um\nand I've been trying to push the limits\nof how long it will actually do this.\nAfter enough time, it finally it tells\nme like gives me keywords like finally\nfinally page this is all you should be\nincluding. But the um but if I hit build\num immediately we start seeing um kind\nof this uh this UI where you can add\nkind of features to the app. It gives\nyou styles and kind of um images that\nyou can invoke. Um, and you can also see\noff to the left the different thinking\ncomponents um that are that are being\nincorporated into the app as well. So,\nI'm going to select this design.\nUm, we see Gemini 3.1 Pro Preview is the\nmodel that's selected, but you can also\nselect different ones here off to the\nright. Um, so Gemini 3 Flash, I believe,\nis the default. Gemini 3.1 Flash Light\nis also one of the models that you can\ninvoke. Um the settings feature also has\num a kind of different microphone\nsources, different system instructions.\nSo if you wanted the model to use\nspecific libraries or specific kinds of\narchitectures, you can. Um we have a\nplace where you can add secrets. So you\ncan add secrets for the Gemini APIs or\nother services. You can see versions,\nwhich if you've ever used AI Studio\nBuild before, you know is very very um\nimportant being able to to version\ncontrol your projects. There's also a\nGitHub integration, so you can see that\num OOTH, which enables Google signin. Um\na place where you can publish, so you\ncan share it with your friends. Um and\nthen also as the code is being written,\nyou can see it populated in the app\nitself. So, you can see it's\nincorporating all of those things that I\nhad put into my description of a brand\nidentity generator, um kind of a brand\nbible dashboard. Uh the app should allow\nusers to select from a variety of style\ntemplates, etc., and an AI content\ngenerator um for uh for post captions.\nI'm going to go over to this gallery\nfeature as well, just so you just so\nyou're able to see it. Um there are\n[clears throat] many many different\nkinds of apps that you can draw\ninspiration from. So things like\nmultiplayer experiments. Um things that\nare grounded with Google maps so you can\nsee different maps and incorporate them\ninto your app. Um things that are using\nLIA 3 for music generation which we\nmight take a look at in a second. Um\ndesign with Nano Banana 2. Um things\nlike Gemini Slingshot which is also\nincorporating in media pipe. um VO\nmodels, um landing pages, things like\ncomic book generation, and a whole bunch\nmore. Um so, if I zoom over to the app,\nit looks like it's still cooking. Um,\nfor that I'm going to go ahead and pull\nup Liia 3. Um,\nuh, which is,\num, which is, uh, an app that was\ncreated by my colleague Guom that allows\nyou to create music clips, um, either 30\nseconds or longer just using, um, just\nusing the Gemini models. If you click\ncode, you can also see all of the code\nthat was used to generate the app um\nincluding the Genai service. So you can\nsee how um so you can see how Gemini is\ninvoked. You can see all of the prompts\nassociated for it. Um the it sounded\nlike uh it sounded like there was um a\nlittle ding so I should be able to\ncreate my sound. Um, and then maybe I\nsay something like electronic dance\nmusic,\num, in German,\num, or maybe in Hindi, um, about, and\nyou're going to have to fact check me,\num, in about the trials and tribulations\nof linear algebra,\num, and fine-tuning open-source models.\nUm, we'll see how this goes. And, uh,\nyou can generate a song.\nYou can also see that you can add custom\nlyrics or instrumentals. It looks like\nthe song is being uh song is being\ngerminated.\nI'm very worried.\nSo, we'll see. Um and then I'm going to\nplay\nmodel opens\nfunction [music]\nmachine learning.\n[music]\nSo, okay. So, I'm I'm like I'm very m\nokay like\nI'm hoping that means it was really good\nand not very bad. Uh but uh but this is\nkind of the things that you can do with\nthe LIA API which is one of our music\ngeneration um music generation models\nalso available via AI studio. So you can\ncreate a background audio track for\nlyrics that you define. You can create\num 30-cond long clips or even longer. Um\nand you can also specify different\ninstrumentals that you want to pull in\nto kind of modify or enhance an existing\naudio clip. So, if you've got a really\ngreat baseline, but you want to add some\nother things to it, um LIA is a is a\ngreat option for you. It also looks like\num the Gemini 3.1 Pro is still cooking.\nSo, it's adding the bento grid design\nthat I had selected for borders and\nrefining the visual style. While that's\nworking, I'm going to show you Genie 3,\nwhich is kind of our our platform for\nworld models. If you've never seen Genie\n3 before, you can kind of just describe\na scene that you'd like to experience.\nSo, this is someone navigating a\ndifficult terrain in the middle of a\nvolcanic landscape on kind of like a\nfour-wheeler thing. Um, you can also\nselect different things like a jet ski\nrunning around during the festival of\nlights. Um, there's no physics engine\nbehind the scenes. So, if you hit a\nlight and it bounces back, that's just\nbased on the training data that was used\nas part of this kind of world model\nharness. Um, and if you actually knock a\nlight into the water and then circle\nback, it will remember for the extent of\nyour one minute long exploration, which\nis pretty wild. Um, but it's everything\nfrom kind of experimenting with\nrealworld scenarios. So, like what it\nwould it be like to experience a\nhurricane firsthand in Florida? Um to uh\nyou know, what would it be like to\nexplore underwater vents um and kind of\nthese volcanic uh style underwater um\nexperiences, these hydrothermal vents\nwith jellyfish and a whole bunch of\nother whole bunch of other scenes. Um,\nso it sounds like we've got our first\nimplementation of the brand identity as\nwell. Um, so, so we'll take a look at\nthat in a second, but I'm going to do\nProject Genie first.\nI'm going to describe a scenario. So,\none that I've never tried before. Uh, so\nmaybe um the Golden State Warriors\nbasketball court.\nUm, but it's on top of a skyscraper\nin the middle of a futuristic\ncity with flying cars\nall around and clouds.\nAnd then maybe the character is a cat\nwith\na pink mohawk\nand uh\na jetpack.\nUh so I'm going to create the sketch.\nWhat's happening behind the scenes is a\nwhole bunch of different models are\nbeing invoked. So things like Nano\nBanana, our VO models, um our Gemini\nmodels are being used to help refine my\nclumsy prompt into something a little\nbit more um a little bit more robust so\nit can be used to better inform the\noutput of the game environment that gets\ncreated. It should create this initial\nimage. So it looks like I've got my cat\nwith a mohawk and the jetpack and the\nGolden State Warriors. Um and then I'm\ngoing to create this world. Um\nAnd then uh I should be able to use the\narrows to kind of look around um the\nspace key to jump around. I can use the\num some of the other keys on my laptop\nto change the direction and position of\nthe video. And then once this world gets\ncreated um I've got uh I've got 60\nseconds to explore it. So it looked like\nit gave me a very large cat. Um but\nWhoa. Okay, cool. So, it looks like it\ndoes have a jetpack.\nUm, and it's using the jetpack to kind\nof like escape and move into the city.\nUm, which is wild. Um, I'm also moving\nthe arrow keys. So, you can see the cat\nkind of defying gravity. Um, which is a\ngreat song from a great musical. Um,\nbut, uh, but all of this is kind of\ndynamically generated on the spot. Um,\nif I want to go like try to make it to\nthat skyscraper far in advance.\nUm, I wonder if the jetpack will like\nallow me to go there a little bit\nfaster. Um, but it looks like it's\nrunning through uh running through the\nsky. You can also see gravity starting\nto kick in because the the cat gradually\ngets lower and lower. Um, and then\nbefore I get to before I get to the\nskyscraper, I think my 60-second timer\nis going to going to cut out. Um, but\nyou can see the flying cars also off to\nthe right, like those little those\nlittle golden specks. You can see the\nskyscrapers. Um, and then all of this\njust gets converted into a playable\nexplorable world. Um, that behind the\nscenes is just pixel by pixel. So, not a\nUnity environment, not something that\nwas created in Unreal Engine, just like\na dynamically created video game\nenvironment that is completely up to you\nto create and define. Um, so wild to\nsee. Um, the uh and then let's go back\nto this brand identity forge that we had\ncreated before and we'll see how well it\ngoes. Um, so I'm going to describe my\ncompany mission and the emotion I want\nto evolve or evoke a sustainable coffee\nroaster in Portland. I'm going to say a\nvegan basketball food truck. Um,\nuh, upload mood boards. I don't think I\nhave that many good mood boards. Um, but\nI do really really love uh, um, I was\ngoing to say like Trans Metropolitan the\ncomic book. Um, but maybe uh to to win\npoints with my um to win points with one\nof my little nieces, I'm going to\ndownload um I'm going to download\nsomething that looks like a fluffy\nunicorn.\nUm so I'm just going to save this image,\nupload it,\nand hope that it supports ABF.\nand then uh generate this brand bible.\nI could have also taken just like a\nscreenshot of all of the of all of the\nunicorns that we saw on the screen um as\nwell. But it's quite cool that for the\nthe apps that you create, you can also\nclick share. So you can share the apps\num with people that you specify or you\ncan just open them up into the entire\nworld to share the URL. Um, you can see\nthat it incorporated the the kind of\naesthetics of the unicorn into the\ndesign of the app. So, you can see the\nstyle selector of minimal, bold, or\nelegant. Um, as well as kind of the\nability to export the image. Um, it's\ngenerating the primary logo. Um, and\nthen it also has some secondary marks.\nUm, as well as a brand voice of\nenergetic, playful, fiercely magical.\nUm, we combine the competitive hype of\nstreet basketball with plant-based\nstreet food. Fun and approachable. I\nlove it. Um, clearly this is my post AGI\nplan is to have a vegan food truck. Um,\nbut uh, but this is pretty cool in order\nto create these apps um, relatively\nsimply. You can incorporate AI into\nthem. So the chatbot features, the\nGemini Live features, Imagine for um,\nfor kind of these logo creations. Um and\nagain you can kind of publish directly\nto Cloudr Run or share um the URL for\nfolks to try as well. Um and with that\nuh I'm going to go back to the slide\ndeck with the caveat that I absolutely\ndetest slides just to tell you about\nanother couple of the models um that we\nhave available. So all of the Gemini\nAPIs are able to be used for robotics\npurposes. You can use Gemini Live to\ninvoke other models that are either\nrunning on device or triggering other\nrobotic actions. We have a whole bunch\nof them on Mountain View Campus. So, if\nyou want to stop by and take a look, um,\nping your favorite DeepMind roboticist\nand they'll see if you uh if you can get\nearly access. Um, augmented reality also\nvery very helpful with Gemini Live. So\nto be able to dynamically explain what\nyou see, um give you directions in real\ntime, kind of explain the environment,\num even do things like dynamic\ntranslation.\nWe already talked about the world\nmodels. And then for video generation,\num we have our VO3.1 series of models,\nparticularly VO3.1 Flash or V VO3.1\nLight. Um, and I don't think any\npresentation is complete without a quote\nfrom Andre Carpathy. So, here's mine.\nUm, but video really, really is a\npowerful medium, especially today for\nbeing able to communicate complex topics\nin a way that's more easily digestible,\nespecially for audiences that are that\nare really really more comfortable\nconsuming information via video as\nopposed to text. Um, though, as an avid\nreader, I hope that text still persists.\nUm, so VIO 3.1 is everything from\nreference powered video. So giving you\nthe ability to have a reference person,\na reference place and stitching together\nthe two, being able to show the same\nreference in multiple contexts.\nBeing able to show um kind of animated\nimages, including animated images with\nguidance, being able to do things like\ncontrol cameras just with natural\nlanguage. So moving back, moving right,\nrotating up, zooming in, being able to\ndo outpainting, so extending a scene or\ninpainting,\nbeing able to do um object additions or\neven removing objects from a scene,\nand then also doing things like\ncharacter control. So having these\nreference space movements and an initial\navatar. Um all of this is possible with\nVO3.1.\num first and last frame also a great\nexample.\nAnd then these are just a couple of the\nvideos that you can create. Um so as an\nexample,\num this is kind of a Yeti style figure\nwalking around in an icy landscape. You\ncan barely hear the snow on the\nbackground dropping. Um this is\num kind of two characters at a clifftop.\nYou can hear the waves crashing.\nYou can have scenes from Mordor\nor what feels like Mordor.\nAnd all of this is just using the VO\nseries of models. So, it's video and\nalso audio um tightly controlled. But I\nwanted to to kind of test it out to see\nif I could use different models to\nreplicate a commercial. This commercial\nin particular. Okay, my name's Paige and\nwhat makes the Chick-fil-A chicken\nsandwich original to me is the\ncrispiness of the breading and the\ntenderness of the fillet. It's tasty,\nit's warm, it's total satisfaction. So,\nso the page in this video is not me\npage. It's like some other page. Um, but\nto to recreate it with the um\nHey, my name's Paige and\nto recreate,\nlet me try let me try otherwise. So to\nrecreate it with V2, which was one of\nour models from about a year and a half\nago, um the process ended up taking\nquite a long time. Um you would have to\nuse Gemini to do the prompt generation,\nGemini text to speech for the background\naudio, music effects or something\nsimilar like LIA for the um for the\nmusic generation. And then you would\nhave to stitch it all together using\nsomething like Camtasia or movie pie to\nget the final output. Um, when you did\nall of that, you ended up with something\nlike this.\nHey, my name's Paige, and what [music]\nmakes a Chick-fil-A chicken sandwich to\nme is the crispiness of the breading and\na tenderness of the fillet. It's tasty,\nit's warm, it's total satisfaction.\nAnd so, this uh end to end took me about\n20 minutes to make. again using kind of\ndated tools. Um, with VO3 um, which was\nour best model about a year ago, the\nprocess got a lot simpler. So, you would\nhave the original video, you would put\nit through Gemini 2.5, which would\ngenerate a detailed description, and\nthen you would put that into VO3. Um,\nand it ended up looking like this.\nMy name is Paige and what makes the\nChick-fil-A chicken sandwich to me is\nthe crispiness of the breading and the\ntenderness of the fillet.\nSo, a little bit better. I still like my\nfirst version though. And then the\nprocess with VO3.1\num took that same prompt which was\ngenerated with Gemini 2.5\num put it into the put it into the model\nand I got this as the first example.\nHi, my name's Paige and what makes the\nChick-fil-A chicken sandwich to me is\nthe crispiness of the breading and the\ntenderness of the fillet,\nwhich is pretty wild. I think it took\nless than I want to say like a dollar in\norder to generate that. Um and it also\ntook less than a couple of minutes. Um\nso the process of creating these these\nkind of ads that are hyperpersonalized\nor that kind of speak in a certain way\nthat are um you know sort of translated\nacross different languages um is really\nreally near and uh it's great to see how\npeople are starting to experiment with\nthese um as part of their projects. So,\nwe already talked about our model\nfamilies. We talked about our Jimma\nseries of models. Um, and I just want to\nclose by saying before we open up for\nquestions, there's never really been a\nbetter time to be a founder. Hopefully,\nI've commenced you of that. Um, it's\nreally compelling what people are able\nto do with small numbers of teams now.\nUm, especially one or two, three people\nwho really feel convicted and uh and\nvery passionate about a given topic. you\ncan do much more than you could have\never expected um using a smaller kind of\ndollar footprint, a smaller GPU\nfootprint. And there's uh also never\nbeen an easier time to do marketing at\nscale that's hyperpersonalized to many\ndifferent customers um and to to really\nscale out GTM. So, so just want to throw\nthat out there. If you've ever wanted to\nbuild a company, now's your time. Like,\nyou know, build multiple companies\nsimultaneously. A lot of people have\nbuilt companies using AI Studio or tools\nlike Lovable. Um, and they're revenue\ngenerating sources, even passive revenue\ngenerating sources, similar to how\npeople would previously have like\nvending machines. Um, so with that, I\njust want to say thank you. Go build\nthings. Uh, show us what you build and,\nuh, let me know how I can help. [music]",
  "transcript_chars": 34506,
  "ingested_at": "2026-05-21T19:17:11.562712+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 396,
    "like_count": 11,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}