Fireship
May 22, 2026
Google I/O 2026: Gemini's AI Agent Era & The Death of Traditional Search
Key Takeaways
- 1Google is pivoting from organizing information via hyperlinks to becoming the interface to reality itself through AI agents integrated into every product
- 2Gemini Omni is a multimodal model that takes any input (text, video, sound) and produces any output, with understanding of language, physics, and motion
- 3Google's token serving capacity has exploded from 9.7 trillion to 3.2 quadrillion tokens per month in 2 years, enabled by new TPU-T and TPU-I chip specialization
- 4Gemini Flash 3.5 offers near-Opus 4.7/GPT-5.5 performance at significantly faster speeds, though pricing has increased 3-30x compared to previous versions
- 5Anti-gravity IDE demo showed AI agents building a complete OS in 12 hours and generating drivers on-the-fly, shifting focus from code writing to agent management
- 6HTML on Canvas API enables developers to render native HTML elements directly in canvas for highly interactive UIs combining WebGL/WebGPU with HTML
Chapters
1. Google's AI Scale & Infrastructure
Google's token serving capacity has grown from 9.7 trillion to 3.2 quadrillion tokens per month. The company is splitting TPU chips into TPU-T (training) and TPU-I (inference) to optimize for different workloads, with massive capex investments driving infrastructure growth.
2. Gemini Omni & The Agentic Era
Google announced Gemini Omni, a multimodal model accepting any input format and producing any output. The company is positioning AI agents across all products (search, Gmail, Android, glasses), marking a shift from hyperlink-based information organization to becoming the interface to reality itself.
3. Neural Expressive Design System
The new Gemini app design system optimizes for on-demand UI generation, creating diagrams, timelines, and mini apps dynamically based on user prompts. The visual refresh includes new icons and gradients alongside functional improvements.
4. Gemini Flash 3.5 & Model Updates
Gemini Flash 3.5 delivers near-flagship performance (comparable to Opus 4.7 and GPT-5.5) with significantly faster inference speeds. However, pricing increased 3-30x relative to previous versions. Gemini 3.5 Pro remains unreleased until summer 2026.
5. Anti-gravity IDE & AI Coding
Anti-gravity (formerly Windserve) demonstrated building a complete OS in 12 hours with AI agents handling design, testing, and deployment. The tool pivoted from code writing to agent management, with a live demo showing real-time driver generation for Doom gameplay.
6. Web Developer Tools: HTML on Canvas API
Chrome introduced the HTML on Canvas API, allowing native HTML elements to be rendered directly in canvas. This enables developers to build interactive UIs with pixel-level control via WebGL/WebGPU while leveraging HTML for basic UI elements.
Glossary
- TPU (Tensor Processing Unit)
- Google's custom silicon chip optimized for AI workloads; now split into TPU-T for training and TPU-I for inference
- Tokens
- Units of text processed by language models; Google now serves 3.2 quadrillion tokens per month
- Gemini Omni
- Google's latest multimodal AI model that accepts any input (text, video, sound) and produces any output while understanding language, physics, and motion
- Agentic AI
- AI systems that can autonomously take actions and make decisions within applications rather than just generating responses
- World Model
- AI models that simulate and understand reality, capable of reasoning about physics, motion, and spatial relationships
- Neural Expressive
- Google's new design system for Gemini optimized for generating UI elements dynamically based on user prompts
- WebGL/WebGPU
- Web APIs that enable developers to harness GPU power for graphics rendering and high-performance computing in browsers
- Anti-gravity IDE
- Google's AI-powered coding tool that uses specialized agents to autonomously build full-stack applications from prompts