The intro before the intro. Like every year I watched the keynote & developer keynote with excitement, taking my usual notes and reactive thoughts as it streams. This year though instead of myself turning those thoughts into a blog post, I asked Gemini to do it!
This post is the outcome of using Gemini’s new DeepResearch feature. Coming soon will be a blog post of the process. For now, have a read of what caught my eye from the Google I/O keynotes through the medium of Gemini! It’s honestly a great read, expanding on the notes I took for myself, I am suitably impressed with the content or I wouldn’t post it. 🙂
Introduction
Google I/O 2025 was a deep dive into the future of Artificial Intelligence. Sundar Pichai’s opening statement set the tone: Gemini, Google’s AI model, would integrate into tools used by two billion people, impacting billions of lives. The sheer volume of “AI” mentions (between 92 and 121 times) underscored its central role.
The message was clear: AI is no longer just a cloud technology. It’s moving to every device and into every developer workflow. For Android developers, this means on-device AI is becoming a standard expectation. Google aims to make AI ubiquitous, context-aware, and optimized for everything from cloud infrastructure to edge devices.
The post will explore Gemini’s pervasive reach, the expanding Gemma family for on-device intelligence, AI agents in Android Studio, and the broader implications of agentic AI, from smart glasses to wildfire detection.
I. Gemini’s Pervasive Intelligence: Beyond the Cloud
The Scale of AI Adoption: The Staggering 50x Increase in Processed Tokens
A notable statistic from Google I/O 2025 was the surge in Google’s monthly processed tokens: from 9.7 trillion in April 2024 to over 480 trillion in April 2025. This represents a 50-fold increase in just one year. This isn’t just a big number; it signals a rapid acceleration in AI adoption and usage depth across Google’s ecosystem. The volume suggests not only more users engaging with AI but also deeper, more complex interactions, as “thinking models” consume more tokens.
For developers, this rapid AI adoption means the “AI-first” development paradigm is here. If users are interacting with AI at this scale, applications must integrate AI capabilities to stay competitive. This creates a clear need for Android developers (like us) to sharpen their AI and machine learning skills, as demand for AI-powered features will only grow.
Gemini Everywhere: Integration into Wear OS, Android Auto, and Google TV
Gemini is extending its reach beyond a standalone model, integrating into watches, car dashboards, and televisions. This widespread integration reflects Google’s vision for ubiquitous contextual computing.
On Wear OS, Pixel Watch users will soon verbally interact with Gemini, setting reminders or retrieving event details via Google Workspace apps like Calendar and Gmail. This aims for a more hands-free, context-aware assistant experience on the wrist.
For Android Auto and Google Built-in systems, Gemini will take over voice control from Google Assistant. Drivers will be able to summarize text messages, translate communications, or request a book synopsis, all while driving. This pushes developers to consider how their apps can intelligently interact across diverse form factors, emphasizing robust voice and contextual AI integration.
On Google TV, Gemini will offer custom movie recommendations or pull up educational YouTube videos using natural language queries. This aims to transform passive content consumption into an interactive, intelligent experience.
Gemini’s Problem-Solving Prowess: The “Pokemon Blue” Feat
A curious, if not entirely practical, demonstration of Gemini’s capabilities was its successful completion of the classic game Pokemon Blue, defeating all gym leaders and the Elite Four.
While some might dismiss this as trivial, it does highlight Gemini’s ability to reason, plan, and execute complex, multi-step tasks in a dynamic environment. The fact that it took Gemini roughly 800 hours – significantly longer than a human and that it did so with a notoriously “woof” (poorly balanced) team, relying heavily on an overpowered Blastoise underscores its persistent, adaptive problem-solving nature, but also whilst powerful, these AI agents may still lack human intuition in certain domains, suggesting a hybrid human-AI workflow will likely remain optimal for now.
II. The Expanding Gemma Family: On-Device AI for All
Gemma3n: Lightweight, Multimodal, On-Device Power
Google’s Gemma 3n is a step forward for on-device AI. This lightweight, multimodal model is designed for efficient, offline operation on devices with as little as 2GB of RAM. This is significant for mobile developers, enabling on-device processing of audio, text, images, and video, which enhances user privacy by keeping data local. Gemma 3n shares its architecture with Gemini Nano, indicating a consistent approach to on-device AI. It’s currently only available for developer preview.
Gemma 3n points to a future of privacy-first, edge AI applications. For Android developers, this means building sophisticated AI features—like local summarization and real-time speech transcription—without sending sensitive user data to the cloud. This addresses privacy concerns and reduces latency and cloud costs. It also makes powerful AI more accessible on lower-cost hardware. This development responds to the need for AI that works “anywhere, anytime, even in poor network conditions”, opening new use cases for Android apps.
Specialized Gemma Models
Beyond Gemma 3n, Google introduced specialized variants, showing a focus on tailoring AI models for specific domains.
MedGemma: Healthcare AI
MedGemma, a Gemma 3 variant, is designed for analyzing medical text and images. MedGemma is a starting point for healthcare AI apps, supporting tasks like medical image classification and clinical reasoning. MedGemma highlights a trend: the shift from general-purpose LLMs to highly specialized, domain-specific AI models. This targeted approach improves accuracy in niche fields like healthcare.
SignGemma: Bridging Communication Gaps
Google also previewed SignGemma, an upcoming open model for sign language recognition, specifically translating American Sign Language (ASL) into English text. Google claims SignGemma is “currently the most powerful sign language understanding model”. Its goal is to help developers create more inclusive communication tools for the deaf and hard-of-hearing community.
This opens a new area for assistive technologies, potentially integrated into Android XR glasses or other mobile devices, making real-time communication more accessible.
DolphinGemma: A Frontier in Interspecies Communication
Perhaps one of the more unusual announcements was DolphinGemma, described as the world’s first large language model for dolphins. This model is trained to discern dolphin vocalization structures and generate novel, dolphin-like sounds. While it doesn’t interpret “meaning,” it helps scientists identify structural features in vocalizations for further study. Google plans to open-source DolphinGemma later in 2025.
DolphinGemma, despite its niche, has implications. First, it shows LLM architectures’ versatility beyond human language. Second, its ability to run on a Pixel phone reinforces the trend of powerful on-device AI, even for specialized research.
Colab’s AI-First Transformation: Agentic Colab
Google Colab is transforming into a new, fully agentic experience. Powered by Gemini 2.5 Flash, this enhanced Colab will allow developers to state their desired outcome, and the environment will proactively act within the notebook—autonomously fixing errors, transforming code, and even building user interfaces. This aims to reduce coding time and streamline complex tasks like model fine-tuning.
This is a step towards AI-driven development environments acting as productivity multipliers. By letting AI handle boilerplate and error correction, developers can focus on higher-level problem-solving.
III. Supercharging Android Development with AI Agents
Journeys for Android Studio: Natural Language-Driven End-to-End Testing
Android Studio introduces “Journeys,” an agentic AI tool using Gemini to simplify end-to-end testing. Developers can describe complex user actions and assertions in natural language, and Gemini will autonomously perform the tests.
“Journeys” marks a significant change in Android testing. By enabling natural language tests, it makes testing more accessible. This is an example of AI reducing cognitive load, allowing developers to focus on higher-value tasks, potentially leading to faster delivery of higher-quality applications.
Gemini AI Agent for Dependency Updates: Automated Dependency Management
Android Studio is gaining a “Version Upgrade Agent” that automatically manages dependency updates. Apparently it will be under: “Refactor Menu > Update Dependencies”.
This Version Upgrade Agent directly addresses a long-standing challenge: keeping project dependencies up-to-date. Manual management is error-prone and consumes valuable time. An AI agent that automates this—presumably by analyzing compatibility, resolving conflicts, and suggesting optimal versions—could substantially boost productivity. This allows developers to focus on feature development, potentially leading to faster release cycles and more stable applications.
Gemini Code Assist: Boosting Coding Performance and Code Reviews
Gemini Code Assist, Google’s AI-powered coding assistant, is now generally available for individuals and GitHub integration, powered by Gemini 2.5. It offers coding performance assistance for tasks like app creation, code transformation, and editing. For Standard and Enterprise developers, a 2 million token context window is coming. This expanded context will help with complex tasks like bug tracing and generating onboarding guides for large codebases. Gemini should be able to understand the entire project, leading to more precise suggestions.
Firebase Studio IDE: Building Full-Stack AI Apps
Firebase Studio (I had not heard of!) is a new agentic, cloud-based development environment designed to streamline the creation and deployment of production-quality full-stack AI applications. It integrates Projects with specialized AI agents and Gemini within the IDE, providing a collaborative workspace.
Firebase Studio represents Google’s push to make full-stack AI app development more accessible and faster. The ability to import Figma designs and have Gemini “add features and functionality” or “detect, set up, and provision your app’s backend” is a significant step towards low-code/no-code AI development.
For Android developers, this could mean accelerated prototyping, reduced boilerplate, and the ability to spin up complex backend infrastructure with minimal manual effort. This aims to streamline the development lifecycle, allowing developers to focus on unique AI logic and user experience rather than infrastructure. This is particularly valuable for rapid iteration.
Jules: Your Asynchronous Coding Agent
Jules, Google’s asynchronous coding agent, is now in public beta and widely available. Its purpose is to free up developers’ time by handling tasks they might prefer to delegate, such as bug fixing, version bumps, writing tests, and initial feature building. Jules works directly with GitHub repositories, cloning them to a Cloud VM, formulating a development plan using Gemini 2.5 Pro, and creating a pull request for human review. It even generates an audio summary of changes.
Jules embodies autonomous development. By taking on “random tasks that you’d rather not”, like clearing bug backlogs or generating tests, Jules aims for productivity gains, allowing human developers to focus on more complex problem-solving or “more desirable coding tasks”. This is a practical AI application that directly impacts developer workflow, moving towards a future where AI functions as a “co-worker.”
IV. Agentic AI and the Future of User Experiences
Agentic Checkout: The Convenience and the Caution
Google Search is introducing “Agentic Checkout,” a new feature in its AI Mode that allows users to track product prices and authorize an AI agent to autonomously purchase an item when its price drops to a desired threshold, using Google Pay.
While this offers convenience, my first thought was: “I can totally see fake websites taking credit card details off agents,” highlights a critical aspect of agentic automation: security and trust challenges. As AI agents gain autonomy and access to sensitive information, the attack surface for sophisticated fraud expands. For developers building such systems, security, transparency, and robust user control are paramount. This feature, while convenient, necessitates a deeper consideration of the ethical and security implications of highly autonomous AI, emphasizing the urgent need for stringent security measures and comprehensive user education.
Android XR: Immersive AI Experiences
Android XR is Google’s new operating system for extended reality, deeply integrating Gemini AI into headsets and smart glasses. This platform aims to merge immersive and AI technologies, supporting existing mobile and tablet apps, as well as XR-tailored content.
For Android developers, this is the next major platform frontier, requiring a new mindset for app design. This shift demands focus on contextual awareness, multimodal input (voice, gaze, gesture), and blending digital information with the physical world, exemplified by live translation. This is a long-term play that will redefine Android app experiences, requiring innovation in UI/UX and robust data privacy considerations.
FireSat: AI for Global Impact
FireSat is an initiative supported by Google ($13 million from Google.org) in collaboration with the Earth Fire Alliance. Its primary objective is to enhance wildfire detection using high-resolution multispectral satellite imagery and advanced AI. When fully operational, a constellation of 50 satellites will provide global imagery updated every 20 minutes. The first prototype satellite launched in March 2025. The AI component compares current images with prior ones and factors in local weather to reliably identify fires.
(I wrote a note that Google could be using this an in-road into satelittes and making google maps cheaper / replacing Google Maps cars .. and the AI wrote:) Regarding whether FireSat data will replace images on Google Maps: it’s highly improbable. Google Maps uses static, high-resolution imagery for general navigation, updated less frequently. FireSat’s imagery, while high-resolution for detection, is multispectral and optimized for identifying heat signatures and fire-indicative changes. It’s a specialized dataset. (Think outside the box AI! I’m not saying its Aliens but…)
Instead of replacement, a more plausible application for Google Maps would be as a dynamic overlay. Imagine real-time wildfire boundaries, smoke plumes, or active fire hotspots overlaid onto Google Maps.
Conclusion
Google I/O 2025 solidified AI as the core of Google’s strategy, placing Android developers at the forefront of this transformation. The 50-fold increase in processed tokens, from 9.7 trillion in April 2024 to over 480 trillion in April 2025, signals an unprecedented acceleration in AI adoption. This is mirrored by Gemini’s pervasive integration across Wear OS, Android Auto, and Google TV, embedding AI into nearly every aspect of the user’s digital life.
The expanding Gemma family, particularly Gemma 3n, ushers in an era of powerful on-device, privacy-preserving AI, empowering developers to build smarter, more accessible, and privacy-conscious applications directly on user devices.
This isn’t just about new APIs; it’s a fundamental shift in how applications are conceived, developed, and experienced. The future of Android is undeniably AI-first, agentic, and deeply integrated into the user’s entire ecosystem, presenting immense opportunities for innovation for Android developers.
Thank you for reading, and thanks to Gemini for the main draft 🙂 It didn’t manage to add my usual sign off, so here it is typed by a human!
Happy coding!
I’d love to hear your thoughts. Connect with me on:
BlueSky @Blundell_apps (or Threads, or X..?)