\n
<\/p>\n
In a major salvo in the AI race, Google announced on Tuesday a slew of new and updated products at its I\/O developer conference. These ranged from tools that deploy personal AI agents to code generators to search tools to a new \u201cworld model\u201d for generating physically accurate video.<\/p>\n
Taken together, the releases paint a picture of Google\u2019s current strategy for bringing AI to consumers and businesses. It\u2019s a strategy that effectively leverages the company\u2019s vast information infrastructure, built up through search, in ways that give it clear advantages over newer AI companies.<\/p>\n
Google DeepMind\u2019s newest models are bigger and smarter, deeply multimodal, and tuned for taking actions. Many of the new products and features announced at I\/O are powered by the new Gemini 3.5 Flash model. Google says the model is optimized for speed and efficiency, is four times faster than other frontier models, and costs between one-half and one-third the price of comparable models. Gemini 3.1 Pro was previously DeepMind\u2019s best model, and 3.5 Flash outperforms it on nearly all benchmarks, notably coding and tool use.<\/p>\n
There is also a Gemini 3.5 Pro model, which will become DeepMind\u2019s new flagship model, but researchers are still studying its safety implications and plan to release it publicly sometime in June. \u201cAll our focus with the 3.5 series has been on taking the model intelligence and making sure tool use, instruction following, long-horizon use cases, and agent decoding all work well,\u201d Alphabet CEO Sundar Pichai said during a call with reporters Monday.<\/p>\n
Google also announced its entry into the growing race to build \u201cworld models,\u201d or models that can create digital environments or video that remains true to real-world physical properties. Gemini Omni, as it\u2019s called, is multimodal, meaning it can generate various kinds of outputs (video, images, text, audio, and more) based on prompts that include content in those same formats.<\/p>\n
One example: A user can provide an image of herself, along with a video, and the model will use high-level reasoning to let her likeness stand in as a character in the video. Google is launching a small version of Omni, Omni Flash, today. A larger Omni Pro model is currently in development.<\/p>\n
Before saying a word about its new models, Google spoke about the infrastructure it has built to support them. Google says it expects to spend up to $190 billion on new infrastructure this year. Much of that will go toward new data centers where Gemini models run on hundreds of thousands of Google\u2019s own AI chips.<\/p>\n
The company is now on its eighth generation of tensor processing units (TPUs), the chips that perform the billions of mathematical computations required by neural networks. As AI labs scale up their computing resources, the power and cost efficiency of the chips they use increasingly affects the economics of serving AI models and apps to users. Google says training large AI models is no longer limited to a single data center, but can instead be distributed across more than 1 million TPUs globally, creating the world\u2019s largest training cluster.<\/p>\n
Google may have a distinct advantage when it comes to training data, too. The company very likely has the world\u2019s most advanced web crawler, the technology that continually scours and indexes web pages so they can be searched. Researchers train large AI models on massive amounts of this web content, and the volume, quality, and composition of that training data can directly impact a model\u2019s overall intelligence.<\/p>\n
Google\u2019s crawlers may simply reach more web pages and content than those used by other AI labs. The company also captures much of this content in a \u201cknowledge graph,\u201d allowing it to quickly serve information about people, places, organizations, products, events, and concepts. Any and all of that information can be used to train models. In addition, Google has the full corpus of YouTube videos available for AI training. That content was very likely used to train the new Omni world model to understand the relationships and movement of objects in the real world.<\/p>\n
A larger point: AI labs ask the public to take a lot on faith. Faith that our information will be kept secure. Faith that companies will spend responsibly on AI safety. Faith that they won\u2019t allow their technology to be used for harmful purposes, such as autonomous weapons or mass surveillance. Faith that new data centers won\u2019t spike energy prices or further tax the environment. Faith that the benefits of AI will be broadly distributed. And faith that the business itself will eventually generate enough market demand and revenue to survive. Google isn\u2019t perfect, but the company\u2019s pragmatic approach to AI gives the impression that it can credibly make such promises, that there are, in fact, adults in the room.<\/p>\n
The dominant narrative has been that companies like Google, Anthropic, and OpenAI need these data centers to power AI-infused business processes at large enterprises. That\u2019s why it was striking to hear Google focus primarily on new consumer-facing models, apps, and services at I\/O. Pichai said during Monday\u2019s briefing that Google is trying to bring as much frontier intelligence to consumers as possible.<\/p>\n
\u201cAs someone who grew up using Google search, I think Google’s whole ethos has been to organize the world’s information and make it universally accessible and useful,\u201d says DeepMind\u2019s Tulsee Doshi, senior director of product management for Gemini & Gen Media, in an interview with Fast Company<\/em>. “And now in the agentic era you can add \u2018help users take action on that information in a way that is thoughtful and intentional\u2019.\u201d<\/p>\n Doshi acknowledged that a large portion of the return on Google\u2019s massive capital expenditure investment in data centers will likely come from enterprise business.<\/p>\n This year, Anthropic and OpenAI expanded their Claude Code and Codex coding tools to cover non-coding information work as well, including the creation and management of autonomous agents. Google may be slightly late to that party, but it is making every attempt to catch up.<\/p>\n The company launched Gemini Spark, a personal AI agent that runs on Gemini 3.5 Flash and stays active in the background even when a user\u2019s devices are off.<\/p>\n Spark\u2019s superpower may be quick personalization. By connecting to Gmail, Docs, Slides, and other widely used Workspace tools, it can quickly learn a user\u2019s interests, preferences, and work habits. Google says it can handle complex tasks such as drafting status updates from multiple documents or planning block parties. It can also perform multi-step tasks like parsing credit card statements, monitoring a Gmail inbox for time-sensitive information, or turning meeting notes into polished documents<\/p>\n As its rivals have already begun doing, Google has also built connectors to third-party tools such as Canva, OpenTable, and Instacart. Google says more capabilities are coming this summer, including the ability to text or email Spark directly, create custom sub-agents, and let Spark control a local browser. Users control which apps Spark can access, and the agent is designed to ask for confirmation before taking high-stakes actions like sending emails or spending money. Google says Spark will soon come to its Gemini mobile app, allowing users to manage agents from anywhere.<\/p>\n At the start of the generative AI boom, many believed AI search would ruin Google\u2019s search advertising business, its cash cow. Google had always placed ads next to ranked search results, the classic \u201c10 blue links,\u201d but it was unclear how advertising would work around custom AI-generated answers. The company now seems eager to argue that radically improving search with AI simply encouraged users to search more often, creating new advertising opportunities that otherwise would not have existed.<\/p>\n Google said users conducted more searches during the first quarter of the year than in any previous quarter, likely because of the conversational, multiple-query nature of AI search. It says \u201cAI Mode\u201d queries have been doubling every quarter, and that more than a billion people now use the tool each month.<\/p>\n Google first began using large language models to help interpret the intent behind user searches. After the arrival of ChatGPT, it introduced \u201cAI Overviews\u201d for some searches, where results were packaged into AI-generated summaries designed to answer user questions. Then came \u201cAI Mode,\u201d an advancement on the same idea. Now AI is best understood as a permanent layer sitting atop all Google search functionality.<\/p>\nPersonal agents<\/h2>\n
Search and AI are becoming one<\/h2>\n