Veo 3 Took First Place by Beating Veo 2. Google Labs Says Chapter Two Just Started

Days after Google I/O, Sonya Huang opens Training Data with the mood shift in tech: public opinion had flipped almost overnight, and the new line was that Google was out in front in AI. She sits down with three Google Labs leads: Thomas Iljic on creative tools around Whisk, Flow, and the Veo video models; Jaclyn Konzelmann on the computer-use agent Project Mariner; and Simon Tokumine on NotebookLM. Asked why the court of public opinion turned so fast, Iljic's first answer is the models.

Konzelmann adds the flood of products—breakthroughs landing in existing surfaces plus net-new experiences, not only on I/O week but in the days leading up to it. Tokumine, who had his own moment the day before, says the validation is real and still feels, from the inside, like the result of a lot of work. They have been improving in generative AI for at least three years. Leaderboards and state-of-the-art demos help, including things he thinks only Google's models can do. Internally, though, it feels like the end of chapter one and the start of chapter two.

A joke from the week makes the point. Veo 3 climbed to first on a video-generation ranking, and the previous occupant of that slot was Veo 2. The team was happy to report that Veo had beaten Veo. Huang's personal eval is simpler: the Will Smith spaghetti-eating test that used to shred AI video. She thinks they have passed it. Iljic's verdict is that there is still room, but it is pretty cool.

Iljic dates the real start of Labs' creative-product work to two and a half or three years ago. Google already had image models such as Imagen, and DALL-E existed outside, but two papers changed what a product could be. ControlNet made generation steerable: you could specify a pose or a scene instead of pushing a button and waiting. LoRAs let you show the model a range of examples and pull the possibility space for a particular look. That sense of iteration—that you could start controlling outputs—felt like the right moment to explore the creative process.

What followed was a lot of stumbling. They trained people internally on ComfyUI-style workflows, then worked with artists on half an episode of 2D animation, *The Not-So-Supervillain*, which they published on YouTube. Out of that exercise came three convictions. Creation has to be iterative, so controls have to live next to the models. Media comes with a blueprint: if I generate something, you should be able to pick up where I left off. And it should be show-and-tell—instead of a very long prompt, show the model images and say it should look kind of like this.

Those convictions became two products on a spectrum. Whisk is the consumer play: everyone now has a visual language at their fingertips, even if they do not have advanced storytelling, and they can quickly remix each other's work. Lessons from that exploration, Iljic says, will probably graduate into how Google handles user inputs across surfaces. Flow is the other pole. You already have a vision, and the job is to give you the tools to make the best version of it in video.

Flow's ideal user, for now, is the AI filmmaker—an ill-defined term he prefers to "filmmaker" because the extreme end of the craft needs bespoke tools and entire existing workflows. One tier down is previs: get a version out fast, then do the full process. The tier that moves him most is people who do not have $100,000 to put an idea into the world and can now at least take a shot. The thing the team wants to build has a name: the generative AI camera. Build the DSLR of generative video first, then distill the Android, Pixel-camera version out of it. Whisk, meanwhile, is exploring everything from funny chat remixes with friends to visuals for slides.

On Veo 3, Iljic credits Google DeepMind with a huge step up: prompt adherence is rising, the six-finger problem is gone, and physics are getting pretty good. What is still coming is choreographing multiple characters, full consistency across scenes, refining an output, and propagating a change across clips. The biggest reveal this time was audio generated with the video. An image is one thing, a video is more than an image, and a video with sound is way more than a regular video. That complete audiovisual illusion, he says, is what opened up the virality.

Huang asks how much of the remaining R&D for the ideal craft tool sits in the product and UI versus the model layer. Iljic thinks both, but the control ideas—character, scene, and location consistency, reshooting—already have a visible path, and new pieces show up week after week. "Not solved," he says, "but we know how to do it, and it will happen." The hard part is still the abstraction of all of it. Where do you define the voice? How do you attach it to the character? How do mannerisms propagate? What inputs do you even want from users? Audio is a new frontier of those questions: talk to drive the audio, type the text, do diarization. How do you let people mold clay with all these models?

Huang relays a founder's analogy: the user should direct like a director on a set—change the lighting here, say that line with a little more of an accent. Iljic still thinks it is show-and-tell everywhere. Forcing everything through text is counterintuitive; there should be showing, acting, mimicking, and reference images as inspiration in addition to language. What is getting clearer to him is that video generation, simulation, and games are becoming the same thing: world-building. You define the stage, the assets, how things are supposed to look, then you shoot in it—reshoot, refine, pause, correct, go back in time and regenerate. The UI, he says, is going to be fairly novel.

DeepMind's Genie model is already on that edge: give it an image, move a character, and the world builds in front of your eyes. Games are fun because the constraints are set; movies are good because small details matter—expression, moment, timing. The work is constraining capabilities toward what you need. He also doubts the old static formats. If I share an image and you can instantly walk into it as a scene, am I sharing an image or an experience? Story, setting, and rules may be the common core of a good game and a good movie: you define the rules, then let other people enjoy themselves in it. On cost, he is optimistic. Image and video are seeing the same speed of cost reduction as language models, from better hardware and from distillation that takes less processing to get to whatever you asked for.

Konzelmann named the project Mariner for the spirit of open-ended exploration—letting users and agents go out and wander. Under the hood it is an action-tuned Gemini model that plans a task, then takes screenshots in a loop, using Gemini's multimodal abilities to understand the screen and keep walking the trajectory. The work started after last year's I/O, when AI Studio and the Gemini API graduated out of Labs onto the developer team and freed them to explore agents that could take action, not just answer questions or generate content. They shipped in December as a Chrome extension that took over your browser, then iterated with a large group of trusted testers into the version announced at this I/O.

The first magic moment was unanimous: watching Mariner control the mouse, click, scroll, and type into boxes felt different once you knew an agent was doing it. The novelty faded into one request: this is super cool, can I please use my browser again? That is why the new version runs tasks in virtual machines in the background. A companion extension can still see your open tabs, so you can be looking at a chicken recipe, tell Mariner to add the ingredients to Instacart, select that tab, and it carries the context into the VM. Huang calls it almost superhuman: people are bad at switching among dozens of tabs, and the agent can see them all at once. The bigger unlock is parallelism. Mariner can run ten tasks at once. Konzelmann's own ritual, back from an errand with a pile of to-dos, is to open Mariner, send three tasks, and return to the document—progress underway, and the list off her mind.

People can still go full-screen, watch the clicks, and pause to take over; oversight still matters on an open-ended experiment. Watching at first builds comfort with how the agent thinks. Testers then said the saved conversation at the end was too long. They wanted a summary of what Mariner did, so they could check it. Konzelmann's read: users want to hand the task off and verify at the end, not sit there the entire time. On the technical bet, some people said screenshots would be too slow and they should read the DOM. Google went with screenshots so the skill could apply to things that are not websites. DOM and accessibility trees are another lever, and she says everything is evolving; they will pivot when it makes sense. Speed has two dimensions: hill-climbing model calls and tool use, and agents doing things in a way that is inherently faster than a person. Planning a four-year-old's birthday loot bags, Mariner searched for ideas, remembered five items, found a good link for temporary tattoos, stored the huge URL instead of pasting it into a doc, and handed back all five links at the end.

Ahead, Mariner capabilities are landing in the Gemini app and in AI mode in Search. Konzelmann buckets the rest as agent, environment, and ecosystem: smarter models, tools, memory, and context; a path from local foreground Chrome to VMs, with a middle layer that still runs on your device but in the background, and eventually an agent that is omnipresent across devices; then agent-to-agent interaction with the outside world. Booking a flight or ordering a pizza is the demo because it is easy to understand, starts from scratch, and needs no prior context. The real spectrum is "do it with me" versus "do it for me," and more complex use cases wait on capability and trust. Agents that skip ads and find the relevant content will force business-model change. Tokumine admits he often abandons checkout because the page is confusing or he cannot be bothered—he likes the outcome of shopping, not the act. Konzelmann imagines a universal cart: half-built baskets across sites, remembered by an agent that represents her, then one click to check out everywhere. Huang's line is that once human laziness at checkout is no longer the filter, the best product wins.

Tokumine has tried to drop the "LM" from NotebookLM and failed; after multiple hockey-stick moments it is part of the team's identity. The product was one of Google's biggest viral hits last year, on the back of Audio Overviews—two AI hosts talking through whatever you uploaded. His own way into the project was those overviews. The team's hypothesis was personal content for an audience of one, maybe two or three, not content for everybody, and the notebook looked like a good place to test product-market fit. They were totally unprepared for the success. The first couple of months, he says, was hanging on for dear life, starting with making sure the TPUs did not melt.

Late last year they added join mode, so you can talk with the hosts and ask questions. This year they took stock of what users actually do once Audio Overviews gets them in the door. A lot of people use a notebook for longer-running work: hobbies, ongoing projects at work, preparing for a presentation. That is now the strategy. They also shipped mobile apps and international Audio Overviews, the end of a long road off a research-grade launch model onto native Gemini audio. Huang frames the split as the RAG workspace versus the podcast killer. Tokumine hopes Training Data does not get replaced, and then describes three superpowers. Notebooks help you accumulate information over time on first-party Google database technology. They bundle intelligence, from Gemini 1.5 Pro at launch to thinking models now. And they make content adaptive to your situation: the same knowledge as an audio overview, a comic, a short movie, or a mind map, fitted to the moment.

He still uses Audio Overviews and still drops out of the illusion—why did they say that, they lost the plot, the narrative is off. It looks like a last-mile fix; it is actually a ton of work to make people suspend disbelief. They have also had one show type for a bit too long. New formats are coming, many inspired by users. People have been dropping their LinkedIn into a notebook since the beginning, partly because it is fun to hear hosts talk about you, and partly to get feedback they cannot get from a real person: how would someone else talk about my strengths and where to improve. Real feedback is hard to find, and AI hosts became a cheap on-ramp. One working example: a 150-page PhD dissertation on invasive wolves in some part of Europe. You could look at a mind map, or spend ten or fifteen minutes on an audio overview, but the comic-book rendition is what helped him grasp the overall narrative. Comics sit next to storyboards, which is exactly the line Iljic is on. Huang jokes about a hero's-journey comic of someone's LinkedIn career. For an audience of one, Tokumine says, that would probably be the most awesome movie you would ever see.

From here the focus is being more useful on those long-running projects, for knowledge workers and for students. The project is where those users need the most help and where value accumulates—the unit of work, or as the team calls it, the unit of knowledge. Passing a test and lifelong learning both count. He is also excited about mobile as more than a shrunk desktop. Google's last mobile-first era put big products onto small screens, and they became hard to change. He wants a companion that uses the form factor, the sensors, and the fact that the phone is always with you. Native recording is the example: in a discussion with smart people, drop Notebook, hit record, and later transform and accumulate the conversation. Doing that on a laptop is weird. On a phone it is the point of the device.

Lightning round. Favorite Labs project they did not already discuss: Iljic picks virtual try-on, which hits a real user need, plays to Google's inventory, and is fun because you can see things on yourself. Konzelmann and Tokumine both pick Stitch, where you describe a front-end design and it comes out. Hottest application area for the next twelve months: Konzelmann says video; Iljic says remixable content, the thing where you generate something and someone else riffs off it, the direction Whisk is already heading and that Veo can power.

Where they have been really wrong, on three, two, one: timing. They have been onto something more than once and simply too early, with projects paused, stopped, and in some cases coming back around. Konzelmann treats that as a good problem. Being early means you have a jump start when the models and capabilities are ready. What they have been right to stick with, for Iljic, is show-and-tell: do not ask users to write two pages of text to describe an image; show and tell the way you would with a friend or an artist. That conviction is moving people off prompting and toward instructing, relying on the intelligence behind the model, and he is sticking to his guns.

Tokumine's version has more history in it. When they joined Labs, Google did not have a functional instruction-tuned language model, let alone an LLM API, and the consensus was that these things were too expensive to build a business around. The team stuck with the technology anyway. That bit of timing they got right: capabilities up, costs down. Labs lead Josh Woodward's tradition, when they joined, was to write docs around what happens in two years. That curve, Tokumine says, inspired a lot of them.