Google I/O '25: Thoughts & Takeaways
It's software spring once again: the couple of months that almost every tech company announces new software and services. We are currently a week out from seeing Apple's WWDC (with a rumored OS redesign, and AI enhancements), and I also just attended my fourth Google I/O a couple of weeks ago.
For a couple of years now: Google has kept an equal footing between their multiple products: Android, AI, Web (Search/Workspace), and Cloud (GCP). There was a time when Google I/O was solely focused on Android and Services, but AI has increasingly been getting a spotlight since about 2020. As an avid user of all these aspects of Google's offering, I'll use this space to note down some of my own thoughts regarding all these aspects.
Android
Our favorite bot turns 16, and while dessert names had been forgone from the mainline a long time ago, internally this version has been christened "Baklava". From an OS and devops position, one of the most notable changes with this version is the introduction of the "Trunk Stable Project" for Android development: which shifts all Android development to internal Google repositories, with the company releasing de-Googled bits in the Open Source Repository.
Android 16 is introducing an "Advanced Protection Mode," which includes things like Theft Detection Lock (which automatically locks your screen if it detects common theft motions), an Offline Device Lock, and stricter enforcement through Google Play Protect, even blocking sideloading of unapproved apps if you opt-in (I have many thoughts on sideloading. In large, I am pro-sideloading, but it makes sense to not enable it for devices that are used in critical sectors like government and finances. To be honest, the barring of sideloading entirely is what makes iOS largely more secure, but I am appreciative of Android as an option in the market that enables sideloading).
Network security is also bolstered by preventing auto-connections to insecure Wi-Fi and offering 2G network protection to guard against interception. On a system level, it can enable intrusion logging for forensic analysis if a compromise is suspected, with logs securely encrypted, and enhances Android Safe Browsing. While Google didn't explicitly name specific ISO certifications for this mode itself, its features align with established security principles like defense-in-depth and least privilege, leveraging Android's hardware-backed security (like a Trusted Execution Environment) and standard cryptography. It's more about an aggressive application of best practices than introducing entirely new open standards for the mode, though it benefits from underlying open standards like Vulkan for graphics or FHIR in the related Health Connect space.
Beyond these deep system-level enhancements, Android 16 is bringing some pretty exciting user-facing changes. Visually, Android 16 is set to be a treat with the full rollout of Material 3 Expressive. Think more vibrant colors, intuitive motion, and adaptive components that should make the whole experience feel more personal and dynamic. This new design language isn't just for phones; it's also making its way to Wear OS 6 + Chrome OS, which is a nice touch for consistency across devices. There are tons of thoughts I have when it comes to UI/UX design across Operating Systems, and I am honestly waiting on seeing what Apple delivers on in about a week to fully form my thoughts regarding this, and make a post about the same.
Android 16 is adding support for an Advanced Professional Video (APV) codec, promising near-lossless video quality with better compression, which is great for those shooting high-res content. We're also seeing Ultra HDR image enhancements and better tools for developers to integrate features like motion photo capture. The OS is also getting smarter about how it handles common tasks. The Quick Settings panel is getting a revamp, with Wi-Fi and Bluetooth toggles reportedly returning to a more straightforward one-click operation (thank goodness!). Multitasking is also being refined with better desktop windowing capabilities and more flexible split-screen options, especially for tablets. This is a special one, and it seems they're partnering with Samsung for their DeX expertise. Essentially, plug in your phone to a monitor, keyboard, and mouse, and you'll get an entire desktop environment with windowing and multitasking capabilities.
Lastly, Google emphasized Android's expansion into new form factors. Android XR got a lot of airtime, with Gemini integration coming to glasses and headsets, and partnerships with companies like Samsung and XREAL. This push into XR, along with continued refinements for foldables, tablets, Wear OS, Android Auto, and Google TV, shows Google's ambition for Android to be a truly ubiquitous platform. My biggest hope is that they deliver on this. Android on tablets only recently has gotten good, and the gap between Chrome OS and Android being separate entities is closing day-by-day.
Takeaways
Wrapping up the Android story from a developer's perspective, the I/O 2025 Developer Keynote hammered home a few core themes.
- First, AI is not just a feature, it's foundational. The deep integration of Gemini across the Android ecosystem, especially within Android Studio (with features like Compose UI generation from Figma designs, code explanations, and even generating unit tests for UI logic), is a clear signal that Google wants AI to be an indispensable part of the development workflow, aiming to boost productivity and help tackle complex problems.
- Second, building for a multi-device world is paramount. There was a significant emphasis on tools and APIs that help developers create adaptive experiences that seamlessly transition across phones, foldables, tablets, wearables (Wear OS 6 getting Material 3 Expressive and improved tooling), TVs, and now, more prominently, Android XR. The push for adaptive layouts, improved emulators for various form factors (including the new embedded XR emulator in Android Studio), and Kotlin Multiplatform (KMP) support are all geared towards making it easier to reach users wherever they are. The app ecosystem, across Apple and Google, are still as vibrant as ever.
- Third, developer productivity and app quality remain key focus areas. Beyond the AI enhancements in Studio, Google highlighted improvements in Jetpack Compose, including better performance and new components. They also showcased advancements in app quality tools, like more insightful crash reporting with Gemini analysis in App Quality Insights, and streamlined testing with better backup/restore support in Studio. The new "Version Upgrade Agent" in Android Studio, designed to help manage dependency updates, also looks like a welcome addition.
- Finally, platform modernization continues. The shift to 16KB page sizes and the establishment of Vulkan as the official graphics API, while requiring some adaptation from developers, are crucial for the long-term health, performance, and security of the Android platform. Google seems committed to providing tools and guidance to navigate these transitions. Overall, the message was clear: Google is investing heavily in making Android development more intelligent, adaptable, and efficient.
Artificial Intelligence
If you thought that Google did a tremendous job with Android and its Operating Systems, the next day was filled with AI to the brim. From consumer chat-bots and image-generators to deep level integrations in Android Studio, context understanding, and devops + cloud: Google has a lot of goals and plans with Gemini, and they're moving fast.
The star of the show was undoubtedly the Gemini family of models, which continues to expand and specialize. We heard about Gemini 2.5 Pro, boasting a massive 2 million token context window, which is frankly mind-boggling and opens up possibilities for understanding and processing incredibly long documents, codebases, or hours of video. Then there's Gemini 2.5 Flash, optimized for speed and efficiency in high-volume, low-latency tasks, and Gemini Nano 3.0 for on-device experiences, which is crucial for privacy and responsiveness. Google even teased Gemini Ultra, their most capable model, now powering a new "Google AI Ultra" subscription tier for those who need the absolute cutting edge (which seems to be positioned to rival ChatGPT Pro).
Overall, here's where I think Gemini shines the most: the platform ecosystem and its "AI-first product" strategy. While other players like OpenAI, Mistral, Perplexity, Anthropic, and DeepSeek have certainly pushed the boundaries of model capabilities, Google's strength lies in its unparalleled ability to integrate Gemini deeply and (mostly) seamlessly into products that billions of people already use daily. Google isn't just building AI models; it's building AI-powered experiences.
Take Search, for example. The new AI Mode in Search, powered by Gemini, isn't just about summarizing web pages; it's about multi-step reasoning, planning, and synthesizing information from multiple sources to answer complex queries. They even showcased "Deep Search" for more thorough responses and "Search Live" (coming from Project Astra) allowing real-time, conversational interaction with Search using your camera. Imagine pointing your phone at a broken appliance and having a conversation with Search about how to fix it! In fact, I actually already use it almost daily for fixing day-to-day issues (most recently to get some sticker residue off my old laptop).
This integration-first approach extends across the board. Workspace apps (Docs, Sheets, Slides, Gmail) are getting more powerful Gemini features, like "Help me write/organize/visualize" and improved summarization. Google Photos is set to get "Ask Photos," letting you search your photo library using natural language in ways that were previously impossible (e.g., "Show me the best photos from all my national park trips").
And the fun part is, I am not just writing this stuff. I actually use it everyday. I recently changed my primary mobile platform from iOS to Android, mainly owing to Google products and ecosystem integrations. I remember been blown away by Gemini integrations in Google Photos just 4 months ago:
From that point on, I have been extensively using Gemini across my email + calendar + notes + code. From parsing my email to tell me the status of my conversations, to finding obscure files in Drive: the integration runs deep. From "did Trish email me back about our photoshoot next month, and what is her preferred payment method, and remind me where we last met 2 weeks ago", to "find me a vegan-friendly sushi and ramen place in Seaport that also allows for service animals, and send directions to my maps for driving there", to even "how do I assemble this table I got from Ikea" during a conversational video chat: Gemini is getting deeper ecosystem strongholds that the competitors currently lack (including Apple, despite them demoing this stuff last WWDC).
Unless OpenAI shifts their focus to creating and curating a device-platform-ecosystem, I believe that Gemini can actually be the better product for the vast majority of people despite OpenAI having the better models (slight detour: OpenAI is actually working on devices w/ Jony Ive's "io" as of writing this).
Beyond these enhancements to existing products, Google unveiled some seriously impressive new generative media tools. Imagen 4, their latest text-to-image model, showed remarkable improvements in detail, realism, and text rendering. Veo 3, a video generation model, can now create more realistic scenes with better coherence and even generate accompanying sound effects and voices. And then there's Flow, an AI filmmaking tool built on Veo 3, designed to help creatives stitch together entire scenes. For music, Lyria 2 and the interactive Lyria RealTime demonstrated incredible potential for music generation and performance.
What I find particularly compelling is Google's vision for Project Astra – their universal AI assistant. The demos showed Astra seamlessly understanding multimodal input (seeing, hearing, speaking), remembering context, and taking action on the user's behalf. This feels like a significant step towards truly helpful, proactive AI agents. They also talked about Agent Mode for Gemini, an experimental feature where you describe an end goal, and Gemini can figure out the steps and get things done on your behalf across different apps and services: akin to a Large Action Model.
And speaking of presence, Project Beam (formerly Project Starline) was another standout for me. This isn't just another video conferencing tool; it's an AI-first 3D video communication platform aiming to make virtual interactions feel remarkably like in-person meetings. Using a combination of advanced 3D imaging, specialized light-field displays, and AI to render life-size, high-fidelity images of participants, Beam creates a genuine sense of presence without requiring headsets or glasses. The goal is to preserve natural eye contact, subtle gestures, and those non-verbal cues that are so often lost in traditional video calls. Google is partnering with HP to bring the first Beam devices to market later this year, with plans to integrate with platforms like Google Meet and Zoom.
Of course, with all this power comes responsibility, and Google made sure to reiterate its commitment to responsible AI development, dedicating significant time to outlining their approach. A big emphasis across all of Google's AI offerings was how they do not intend for these products to replace workers. Rather, a human subject-matter expert will always be required to make the final calls for every matter. This "human augmentation" philosophy is critical. They highlighted practical tools like SynthID, which embeds imperceptible digital watermarks into AI-generated content (images, audio, video, and text) from models like Imagen, Veo, and Gemini. This is crucial for transparency and helping to identify AI-generated media, with a new SynthID Detector portal now available for verification. Google is also expanding SynthID's reach by open-sourcing its text watermarking component and partnering with companies like NVIDIA to encourage wider adoption. The emphasis remains on how human-art is paramount, and how there needs to be critical emphasis put on making it easier for people to distinguish between AI-content and real life content.
Beyond watermarking, Google emphasized its ongoing work in several key areas of responsible AI:
AI Safety and Alignment: Continuous research and development to ensure models behave as intended and align with human values, including red-teaming and building robust safety filters.
Combating Bias: Proactive measures to mitigate bias in training data and model outputs, including using diverse datasets, developing automated bias detection tools, and ongoing monitoring and model updates.
Transparency and Explainability: Efforts to make AI systems more understandable, such as providing detailed model cards that explain a model's capabilities and limitations, and clearly labeling AI-generated content.
User Control and Privacy: Empowering users with more granular controls over their data and how AI processes it, alongside continued investment in privacy-preserving techniques like federated learning and on-device processing with Gemini Nano.
While the "AI will augment, not replace" line is one we hear often, Google's focus on building AI as a tool within existing workflows, coupled with these tangible steps towards safety, transparency, and accountability, lends some credence to this. The labs webpage continues to be the place for early access to some of these experiments, and it's clear they want users to be part of the journey in shaping this technology responsibly.
Takeaways
From a developer standpoint, the AI segment of I/O 2025 was less about a single groundbreaking algorithm and more about democratizing access to powerful, multimodal AI and embedding it deeply into developer workflows and end-user products.
Gemini Everywhere - The API is King: The clear message was that developers should be thinking about how to leverage the Gemini API across its various tiers (Pro, Flash, Nano, and the upcoming Ultra) to bring intelligence to their applications. The expanded context windows, improved reasoning, and multimodal capabilities (processing text, images, audio, and video) open up a vast array of new use cases. Google is pushing hard for developers to build with Gemini, not just marvel at it. For privacy sensitive workflows, Google is emphasizing on Gemini Nano with on-device processing (akin to the goals of Apple Intelligence), which can already be run and tinkered on with in the new Google Pixel 9 devices.
Tooling for an AI-First World: The enhancements to Android Studio with Gemini are a prime example. Features like generating UI from descriptions or Figma, code explanations, and AI-assisted debugging are designed to significantly boost developer productivity. Similarly, Firebase Genkit aims to simplify the integration of generative AI into mobile and web apps, abstracting away some of the backend complexities. The advancements in Imagen 4 (images), Veo 3 (video), and Lyria (music), coupled with tools like Flow, are not just for end-users. These models will increasingly be available via APIs, allowing developers to integrate sophisticated content generation capabilities directly into their products. This could range from dynamic asset creation for games to personalized media experiences.
Rise of the Agents: While still experimental, the vision for Project Astra and Agent Mode in Gemini signals a future where developers can build applications that don't just respond to commands but can proactively assist users and perform multi-step tasks across different services. This will require a shift in thinking about application architecture and user interaction.
Responsible AI by Design: Google is strongly encouraging developers to build responsibly from the outset. Providing tools like SynthID for content authenticity, emphasizing the importance of safety filters, and offering guidance on mitigating bias are all part of this. The expectation is that developers will be mindful of the ethical implications and societal impact of the AI they build.
Openness (with a Google Flavor): While Gemini models are proprietary, Google continues to support the open ecosystem with models like Gemma and tools like Keras. This provides developers with options and pathways, whether they want the full power of Gemini or the flexibility of open models.
Compliment workflows, don't replace: AI should be viewed as a tool: it'll always require a human driver and expert to make proper connections, and make proper use out of. Don't try to do things that humans are innately good at: decision making, logic, emotion, connection. Rather, iterate faster on the tasks that slow humans down: from generating code to making sure the human touch isn't lost during video conferencing. A hammer can only be put to good use in the hands of a great carpenter.
In essence, the AI developer track at I/O 2025 was a call to action: Google is providing increasingly powerful and accessible AI building blocks, and they want developers to start using them to create the next generation of intelligent applications, with a strong undercurrent of doing so responsibly.
Cloud
I'll keep this one quick and short: the crux is that all existing GCP products got updates with the inclusion and integration of the new Gemini models, and Google is delivering on some very ambitious agentic workflows and tooling that enhance how cloud clusters are managed, deployed, and how developer productivity is shifting.
Vertex AI remains the cornerstone, now supercharged with the latest Gemini 2.5 Pro and Flash models. Managing the sheer resources (compute, memory, and high-speed interconnects) required for training and serving these massive models is a monumental task for any underlying operating system. Vertex AI aims to abstract much of this complexity, offering features like "thought summaries" for model auditability and the "Deep Think" mode for highly complex reasoning tasks in Gemini 2.5 Pro. These advanced reasoning capabilities likely involve sophisticated state management and optimized execution paths within the cloud infrastructure.
Developer productivity in the cloud is getting a massive boost with AI-powered coding tools. Gemini Code Assist, now powered by Gemini 2.5 and benefiting from that huge 2 million token context window for enterprise users, is becoming an indispensable coding companion. Jules, the autonomous AI coding agent, takes this further, promising to handle tasks like writing tests and initial feature scaffolding. This hints at AI models performing complex program transformations and understanding developer intent at a high level.
Deploying AI applications is also being streamlined. The ability to deploy applications built in Google AI Studio directly to Cloud Run with a single click is a significant step. This highlights the power of sophisticated containerization and serverless execution environments. Cloud Run instances, likely running on finely tuned OS images, can spin up and down rapidly, efficiently managing resources for potentially bursty AI inference workloads. Managing fleets of AI agents is also getting easier with new Agent Development Kits (ADKs) for Python and Java, and the Agent Engine UI in the Google Cloud console. These tools likely provide frameworks for inter-agent communication, state management, and resource orchestration, abstracting the complexities of distributed AI systems.
Underpinning all this is Google's continued investment in cutting-edge, custom-designed hardware. The announcement of Ironwood, their 7th generation Tensor Processing Unit (TPU), was a major highlight. TPUs are ASICs specifically designed for neural network workloads, and Ironwood is tailored for efficient AI inference at scale, reportedly delivering 10x the performance over the previous generation. The efficiency of such custom hardware relies heavily on a co-designed software stack. This includes specialized compilers that can take high-level model descriptions (from frameworks like TensorFlow or JAX) and generate highly optimized machine code for the TPU's unique architecture, exploiting its massive parallelism, specific instruction sets, and high-bandwidth memory. The runtime system and drivers on GCP must then manage these powerful accelerators, scheduling tasks, managing memory, and ensuring data flows efficiently to keep the compute units fed. The focus on inference performance per watt is also a critical OS-level concern, especially at cloud scale.
And it's not just general-purpose AI; Google is also showcasing specialized AI solutions like MedGemma for healthcare applications and AI-driven tools for things like wildfire monitoring with FireSat, all running on GCP. This demonstrates the platform's capability to support a diverse range of demanding AI workloads, each with its own performance and resource characteristics. The message from I/O 2025 is clear: Google Cloud is not just providing AI models; it's building a deeply integrated hardware and software ecosystem, from the silicon up through the OS, compilers, and managed services, to be the premier platform for building, deploying, and scaling sophisticated, responsible AI solutions.
All in all, great event!!
---
Cover image: UMass Amherst Observatory at Orchard Hill, taken on a Google Pixel 6a in 2023.
