Google DeepMind KeynoteBuilding in the Gemini Era
How Gemini 3 Pro and AI Studio are democratizing software development through agentic capabilities and vibe coding
"We get to be the first generation of engineers who are building tools for a world where anyone can build software."
— Kat Kampf, Google DeepMind
Gemini 3
Latest Model
State-of-the-art with agentic tool calling and design sensibilities
14
People in Images
Nano Banana Pro handles up to 14 people consistently
23+
Live Players
Multiplayer racing game demo with real-time participation
Free
AI Studio
No API key required for most models
The Democratization of Development
Google DeepMind is betting that AI-assisted development will transform who can build software. Kat Kampf and Ammaar Reshi demonstrated Gemini 3 Pro—their most intelligent model with dramatic improvements in design understanding and agentic tool calling—alongside Nano Banana Pro (Imagen Pro), an image generation model powered by Google Search for real-time world knowledge.
The live demos showcased capabilities that would have seemed impossible a year ago: a multiplayer racing game that handled 23+ simultaneous players, comic book generation with perfect text rendering, one-shot website cloning from screenshots, and AI-generated stickers based on real-time web search results. Each demo highlighted different aspects of Gemini 3's capabilities—creative storytelling, design sensibility, real-time knowledge integration, and agentic problem-solving.
Underlying these demos is AI Studio, Google's free vibe coding platform that requires no API keys and supports one-shot creation of full applications. The platform abstracts away technical details—package installation, infrastructure setup, environment configuration—allowing users to focus on what they want to build rather than how to build it. This represents Google's vision for the future: not replacing developers, but enabling a new generation of creators who can build software without traditional programming expertise.
Major Announcements
What Google DeepMind unveiled
1. Gemini 3 Pro: Design Understanding Meets Agentic Capabilities
"Very very strong at design understanding and generating websites and good UIs in one shot. Vastly above in performance in agentic scenarios and leaps above our previous models."
Gemini 3 Pro represents a significant leap forward in two critical areas. First, design sensibilities: the model understands UI/UX principles, avoids common aesthetic pitfalls (like the dreaded "purple gradients"), and generates visually appealing interfaces in a single attempt. Second, agentic tool calling: the model can autonomously decide when and how to use tools to complete complex tasks. This combination enables the one-shot website generation demos that would have required multiple iterations with previous models.
2. Nano Banana Pro: Search-Powered Image Generation
"Powered by Google Search for world knowledge. Improved text rendering that can wrap text perfectly. Multi-language support with localization."
Imagen Pro (referred to as "Nano Banana Pro" in the talk) is Google's latest image generation model with several breakthrough capabilities. Google Search integration means the model can access current information—you can generate "weekend builder" stickers based on what's trending this weekend, not what was in the training data. Text rendering has been dramatically improved, with perfect text wrapping and multi-language support. The model can consistently handle up to 14 people in a single image, addressing a common failure mode in previous image generators. These capabilities enable creative applications like personalized comic books and contextual sticker generation.
3. AI Studio: Free Vibe Coding Platform
"Free to use. No API key required for most models. We don't want you to think about those details. You should just be able to ask, I want to make a multiplayer app."
AI Studio is Google's answer to the question: "What if AI-assisted development was accessible to everyone?" The platform is completely free for most models, requires no API key setup, and abstracts away infrastructure complexity. Users can generate full applications with a single prompt—"I want to make a multiplayer racing game"—and AI Studio handles package installation, server setup, and deployment. The demos showed migration from web-based AI Studio to "anti-gravity" (Google's new Agentic IDE), suggesting a workflow where users can prototype in the browser and then move to professional tools when needed. This positions Google as offering a on-ramp to AI-assisted development with minimal friction.
4. Real-Time Capabilities with Live API
"We've also released a new Live API that you can use to build real-time applications with things like webcam input and audio."
The multiplayer racing game demo wasn't just impressive because it handled 23+ simultaneous players—it was built in real-time during the presentation. The Live API enables applications that can process webcam input, audio, and other streaming data sources. This opens possibilities for interactive AI experiences: gesture-controlled interfaces, real-time video analysis, voice-responsive applications, and collaborative tools where multiple users interact with the same AI system. The demo showed that AI Studio can handle WebSocket connections and state management automatically, abstracting away one of the most challenging aspects of real-time application development.
Live Demo Analysis
What each demo revealed about Gemini 3's capabilities
Comic Book Generator: Creative Storytelling & Text Rendering
The demo showed users uploading photos and selecting a genre, with Gemini 3 generating personalized comic book stories. What made this impressive:
- Perfect text wrapping in speech bubbles, a historically difficult challenge for image models
- Humor and creativity—the presenters noted this was the first time a model made them laugh with genuine humor
- Multi-language support with proper localization for different regions
- Consistent character rendering across multiple panels maintaining narrative coherence
"It's actually the first time that the models have made me laugh. And it's not because I'm biased. I was actually surprised at how funny these were."
— Ammaar Reshi • 08:00
Laptop Sticker Generator: Real-Time Knowledge Integration
Ammaar asked for "weekend builder" stickers, and the system used Google Search to understand what's relevant this weekend and generate contextual designs. This demo showcased:
- Google Search grounding—accessing current information beyond the training data
- Contextual creativity—generating designs based on real-time trends and events
- Personalization—incorporating user-specific details into generated content
"Generate laptop stickers for me. I'm a weekend builder."
Result: The AI searched for current "weekend builder" context and generated personalized sticker designs
Animated Website & UI Cloning: One-Shot Design Generation
The presenters generated a slick animated website with shader effects and beautiful typography, then cloned AI Studio itself in a single shot. Key takeaways:
- Shader animations—generating complex visual effects that typically require specialized graphics programming
- Typography selection—choosing appropriate fonts and layouts automatically
- Avoiding design pitfalls—the presenters specifically noted the lack of "purple gradients," a common AI-generated aesthetic flaw
- Screenshot-to-code—cloning AI Studio's UI from an image and adding export functionality
"No purple gradients. Just beautiful, professional design."
— Kat Kampf, emphasizing Gemini 3's improved aesthetic judgment
Multiplayer Racing Game: Real-Time Collaboration at Scale
The final demo built a 3D racing game in Three.js, then converted it to multiplayer with live audience participation. 23+ players joined simultaneously, and the system handled the load seamlessly. This demonstrated:
- Three.js integration—generating complex 3D graphics code with proper scene management
- WebSocket management—handling real-time connections for two dozen concurrent players
- State synchronization—keeping game state consistent across all clients without lag or corruption
- Live risk tolerance—building and deploying a multiplayer game with audience participation during a keynote
Live Demo Stats:
23+ concurrent players • Real-time WebSocket connections • Zero crashes • Generated and deployed in minutes
Top 12 Quotes from the Talk
Direct insights from the presenters
"We get to be the first generation of engineers who are building tools for a world where anyone can build software."
— Kat Kampf
"Very very strong at design understanding and generating websites and good UIs in one shot."
— Kat Kampf, on Gemini 3 Pro
"Vastly above in performance in agentic scenarios and leaps above our previous models."
— Kat Kampf, on Gemini 3's benchmark results
"Purple gradients and things that just, you know, they kill me as a designer."
— Kat Kampf, on avoiding AI design cliches
"So many folks who were struggling with design who might have still tried to grok their way around Figma don't have to do that anymore."
— Kat Kampf, on AI democratizing design
"It's actually the first time that the models have made me laugh. And it's not because I'm biased."
— Ammaar Reshi, on the comic book generator
"We don't want you to think about those details. You should just be able to ask, I want to make a multiplayer app."
— Kat Kampf, on AI Studio's philosophy
"No cyberpunk. Just beautiful, professional design."
— Kat Kampf, during the animated website demo
"Backend support coming very very soon. Full stack runtime planned."
— Kat Kampf, on AI Studio's roadmap
"Powered by Google Search for world knowledge."
— Kat Kampf, on Nano Banana Pro
"Can wrap text perfectly. Multi-language support with localization."
— Kat Kampf, on Nano Banana Pro's text rendering
"Up to 14 people in one image consistently."
— Kat Kampf, on Nano Banana Pro's person handling
AI Studio: Google's Vibe Coding Platform
Features and capabilities
What is AI Studio?
AI Studio is Google's free platform for AI-assisted development. It embodies the "vibe coding" philosophy—focus on what you want to build, not how to build it.
Key Features
- • Completely free (no API key for most models)
- • One-shot website and app generation
- • Automatic package installation (ShadCN, etc.)
- • Google Search grounding integration
- • Google Maps integration
- • Live API for real-time applications
- • Anti-gravity IDE export workflow
Coming Soon
- • Backend support ("very very soon")
- • Full stack runtime
- • Enhanced multi-model support
AI Chips: Unique Integrations
AI Studio offers "chips"—pre-built integrations that add powerful capabilities to your projects without manual configuration:
Google Search Grounding
Connect your AI to current web information. Perfect for apps that need up-to-date data.
Google Maps Grounding
Integrate location data, maps, and geographical context into your applications.
Live API
Webcam input, audio processing, and real-time streaming for interactive experiences.
Gemini 3 vs Previous Generations
What changed and why it matters
Performance Improvements
Can generate professional UI/UX in one shot, avoiding common aesthetic failures
Dramatically improved SWE benchmark performance and tool usage
Genuine humor and creative storytelling in generated content
Real-World Use Cases
How developers can leverage these capabilities
Content Creation
Comic book generators, personalized sticker creation, marketing materials, and social media content with perfect text rendering and multilingual support.
Rapid Prototyping
One-shot website generation, UI cloning from screenshots, animated landing pages, and full-stack application scaffolding.
Interactive Experiences
Multiplayer games, real-time collaboration tools, gesture-controlled interfaces, and voice-responsive applications using the Live API.
Knowledge-Integrated Apps
Context-aware applications using Google Search grounding, location-based features with Maps integration, and real-time data updates.
Try AI Studio Today
Google AI Studio is free to use and requires no API key. Start building with Gemini 3 Pro's state-of-the-art capabilities today.
Related Talks
More AI engineering insights
Future-Proof Coding Agents
Bill Chen and Brian Fioca from OpenAI on building coding agents that survive model evolution
The AI Developer Experience Doesn't Have to Suck
Eric Bernhardson from Modal on serverless infrastructure and sub-second cold starts
Building Cursor Composer
Lee Robinson from Cursor on AI-native development environments
Research Methodology: This analysis is based on the full VTT transcript from the AI Engineer Summit keynote.
All quotes are verbatim from the talk with timestamps. For more details, watch the full video on YouTube.