In May 2026, Google I/O introduced more than a hundred AI features over the course of two days. A good number of the reporters covering the event couldn't keep up with all of it, and among tech analysts a phrase started making the rounds: "I/O spaghetti." The features sprawled off in so many different directions that it was hard to tell what the center even was.
This isn't a Google problem alone. That same week Microsoft, OpenAI, and Meta each rolled out new AI features of their own, and as the competition accelerated, one line became the most common thing said by people trying to adopt these tools: "I don't know what I'm supposed to use." The gap had widened between how fast new features ship and how fast anyone can actually fold them into real work. The more features there are, the heavier the burden of choosing becomes. Google I/O was simply the largest stage on which that played out.
What Google Actually Did This Time
At I/O 2026, Google chose not so much to launch new products as to thread AI through the services it already runs.
AI Overviews now sit by default at the top of Search. Type a question and you get a summarized answer first, instead of a list of links. Gmail gained suggested draft replies and automatic sorting of incoming mail. Google Photos analyzes the background and context of an image and writes a caption for it automatically. Chrome bundles the contents of your currently open tabs into a single summary, and Google Maps turned its directions into a conversational question-and-answer flow. On YouTube, AI was wired into video summaries and comment replies.
There were features for developers, too. Jules, a coding agent, connects to a GitHub repository, hunts down bugs on its own, and proposes fixes. Veo 3, a video-generation model, produces high-resolution clips from nothing more than a short description. NotebookLM expanded its ability to analyze several documents at once and quickly surface a specific point.
Google DeepMind also laid out a research direction it calls "World Models"—technology in which an AI understands the cause-and-effect and spatial structure of the physical world as if simulating it. It's a long-term bet that connects to fields like robotics, medical diagnosis, and climate prediction, but which product it lands in, and when, was left unclear at the announcement stage.
There's a logic to this strategy. Build a separate app and users have to learn something new. Bring AI into the tools they already open every day, and adoption climbs without that learning curve. It's the same move Microsoft made by tucking Copilot inside Word and Excel. The more AI gets used, the more usage data piles up, and that data feeds the training of the next model. For a platform company, the longer this loop holds, the more it becomes a competitive advantage.
But the Strategy Comes With Its Critics
There's no shortage of skepticism about all this.
When you insert AI into every feature, the polish of each one becomes uneven. Users can't know which ones actually work well until they've tried them firsthand. With something like a hundred features, finding the handful worth using becomes its own separate time sink. And when one feature behaves differently than expected, the disappointment doesn't stay contained to that single feature—it can shake a user's trust in the surrounding services as a whole.
Tensions inside DeepMind get reported now and then, too. DeepMind began as an independent organization focused on AI safety research and the long-term alignment problem. Since being folded into Google, it has taken on a structure where it must also contribute to commercial product development, and from the outside it's hard to see clearly where the real priority sits between safety research and revenue contribution. Some researchers point out that commercial pressure could crowd out long-term safety work. The choice of how fast to move—between rushing World Models technology into products and validating its safety carefully—is shaping up to be a defining question going forward.
It's also worth adding: putting AI everywhere is a platform's competitive strategy. Trying out every AI is how a working professional allocates their time. A strategy being rational for the platform doesn't make the same strategy right for the practitioner. Lose sight of that difference, and you end up in a loop where your own work cycle gets pinned to the platform's release calendar.
What This Announcement Leaves for a Korean Solo Operator
Some of the features unveiled at Google I/O can go to work right now.
Gmail's draft suggestions can cut the time spent on repetitive email. If you regularly trade similar messages with the same clients or partners, editing an AI-generated draft is faster than writing from scratch. NotebookLM's expanded features are useful for getting a fast read on long contracts, proposals, and market research reports—finding a specific clause or a key figure without reading the whole document. And note-tidying that links into Google Docs cuts the time it takes to structure what came out of a meeting.
On the other hand, World Models research, Google Beam's holographic video calls, and complex AI agent frameworks are not, in 2026, at a stage a solo operator can actually deploy. Roughly half of this announcement was really about previewing where things are headed one to two years out.
The trouble is the pressure that sets in right after the keynote—the sense that you have to try everything immediately. Your social feeds fill up with new-feature explainers, and dozens of "I/O recap" videos go up on YouTube. React to that noise by pouring energy into chasing new tools, and the work that was already running smoothly slows down. The pattern of AI launch news disrupting a work routine now repeats every quarter.
There was a similar wave when smartphones went mainstream. Once hundreds of thousands of apps came flooding in, people figured they had to try them all. Yet look at the ones whose productivity actually rose, and they tended to use fewer apps, not more. The person using two apps deeply got work done faster than the person using five. What the practitioners who survived an era of fast-churning smart technology had in common wasn't being first to adopt anything new. It was knowing where the bottleneck in their own workflow was, and picking only the tool that fit that exact spot.
A Practitioner's Read After the I/O Keynote
There are a few things you can try in your own work right after Google I/O.
Write down the list of AI tools you're currently using. Then split that list into the ones you've actually opened in the past two weeks and the ones you haven't. The ones you haven't opened are probably not ones you need right now. The fact that the work still got done means you can do without them.
Pick the one repetitive task that eats the most of your time each week. Email triage, research, drafting a report—anything. Spend just three days finding a way to handle that one task with AI. That's less time than it would take to read every Google I/O article. If it works, that tool stays in your workflow. If it doesn't, you've gained a piece of information: that among the hundred-odd features announced, this one doesn't fit how you work.
How Google will weave DeepMind's World Models research into which products, and how DeepMind will balance safety research against commercial goals, will all get clearer in announcements still to come. For now, that question matters less than figuring out where AI actually fits in your own work.
On the day Google put out a hundred features, what a practitioner should pick is two or three of them—and the person who gets one more bit of mileage out of a tool they already use is more likely, six months from now, to be getting more done than the person who tried to keep up with all of I/O.
