Google I/O is Google A/I as search biz goes all-in on AI

Google I/O is Google A/I as search biz goes all-in on AI

Time Stamp: May 14, 2024 3:18 PM
Source Node: 2578292

Google’s annual developer conference is still called Google I/O, though this year Google A/I would be equally apt.

Almost every developer-oriented announcement coming from the event, which began on Tuesday, includes a reference to AI, a one-size-fits-all term that’s been enlisted to describe everything from text summarization to image recognition and language translation.

If the fawning over blockchains got on your nerves by the time it peaked around 2018, the incessant rumination about robo-thought probably is probably a serious threat to your sanity. Hang in there, it’s just software and the hype will subside eventually.

Google, facing rivalry from OpenAI and others, continues to enhance its marquee AI models, with Gemini 1.5 Pro entering public preview in over 200 countries alongside a new speed-tuned variant called Gemini 1.5 Flash. Both models sport one million token context windows – meaning they can accept large amounts of data as input – and both of them are being offered via Google AI Studio with a two million token context window to developers participating in a private preview.

The Gemini API now can handle parallel function calling and native video frame extraction. And soon it will support Context Caching, which can help manage costs by storing frequently used context data.

Google’s open family of models, Gemma, is expanding with a new sibling. Sitting next to CodeGemma, for code completion and generation, and RecurrentGemma, for better memory usage, there’s now PaliGemma, for multimodal vision-language tasks.

Google also plans to show off its forthcoming Gemma 2 series of models, due in June. The first member of that set will be a 27 billion parameter model.

“It’s optimized to run on Nvidia’s next-gen GPUs or a single TPU host in Vertex AI,” explained Josh Woodward, VP of Google Labs, during a media briefing. “So that’s what makes it easy to use. And we’re already seeing some great quality. It’s outperforming models two times bigger than it already.”

Gemini in Android Studio, which began under the name StudioBot, is getting retrofitted to handle multimodal input using Gemini 1.5 Pro later this year.

“Android is an OS with AI at the very core,” said Matthew McCullough, VP of Android Developers. “And we’re helping developers build incredible experiences that only AI can unlock and then get those experiences into users’ hands.”

McCullough said Google is offering developers multiple ways to integrate AI into their Android apps, including the Google AI Client SDK and the Vertex AI SDK, both of which are in public preview.

“If a developer is new to building with generative AI, then they can experiment and prototype an AI studio, and then they can seamlessly integrate Gemini into their Android app with the Google AI Client SDK,” said McCullough. “Similarly, if an enterprise user is already using Vertex AI, then they can use the Vertex AI SDK to access the full capabilities of Gemini.”

Google’s big gambit this year is adding a machine learning model to Chrome, specifically its Gemini Nano model, thanks to improvements in WebGPU and WebAssembly that are making AI work better on a wider set of hardware.

“Starting in Chrome 126, Gemini will be built into the Chrome desktop client,” said Jon Dahlke, director of product management for Google’s Web Platform group. “We will use Gemini to power our own AI features, including Help me write, which uses on-device AI to help users write short-form content like product reviews, social media posts, and customer feedback forms.”

Google will be using the integrated Gemini Nano model for its own browser-based AI features and making it available to developers through a preview program.

“Our vision is that developers will be able to deliver powerful AI solutions to Chrome’s billions of users without having to worry about prompt engineering, fine-tuning, capacity, or cost,” explained Dahlke. “All they’ll have to do is call a few high-level APIs like translate, caption, or transcribe.”

Other browser makers may take similar steps. According to Dahlke, Google has “started to engage with other browsers,” meaning web standards for AI models in browsers may emerge.

The Register asked whether Gemini Nano will come with any new reporting capabilities to inform Google about model errors or abuse. Following the media event, a Google spokesperson replied, “Safety is an integral part of our model development as well as deployment. Gemini Nano integrations are built according to Google’s work with SAIF and Responsible AI practices.”

As that statement doesn’t make it clear whether Chrome will emit novel telemetry data to watch for model abuse, we’ve asked for further clarification.

Chrome’s AI infusion has also seeped down to the Chrome DevTools Console, which will see explanations of errors and debugging solutions, thanks to Gemini Nano. This feature, dubbed Console insights, will be initially offered as an experimental feature in the US next week, and should roll out to other countries somewhat after that.

Even the Speculation Rules API, which facilitates faster page loads by pre-fetching and pre-rendering of pages in the background, has been seasoned with AI. And according to Google, AI can be used for further optimization by predicting navigation patterns to improve efficiency of resource preloading.

Google’s cloud-based integrated development environment, Project IDX, has now left public preview and entered its beta phase. Yes, it can be used for AI-powered apps, in case anyone is asking.

The Chocolate Factory’s Dart programming language and Flutter cross-platform development framework get bumps to 3.4 and 3.22 respectively, which follow layoffs that hit Google’s Flutter and Python teams.

The most significant changes include native support for WebAssembly in Flutter web apps, which is said to boost frame rendering time by a factor of two in most cases. In Dart, the update adds Macros, which may prove useful for automating Flutter development workflows.

Google’s Firebase service is making GenKit, an AI toolkit, available as a beta release. It’s an open source framework suited for JavaScript/TypeScript, and soon Go, for writing Node.js backends that bring AI features to apps. It provides a way to integrate data sources, models, cloud services, agent software, and the like.

And finally, after infusing everything with AI, Google is making its Checks service available to Android and iOS developers. Checks uses AI to verify that apps, AI-enabled or otherwise, comply with privacy and data collection requirements. ®

Time Stamp:

More from The Register