AI, AI, AI, AI

At Google I/O in 2023, reporters at CNET sat through the two-hour keynote and kept a tally. The word “AI” came up roughly 143 times. That’s a little over once a minute, for two hours, from a company that had already spent seven years calling itself AI-first.

By 2024, Google had decided to get ahead of the joke. Sundar Pichai closed that year’s keynote by putting the count on screen himself: 121 mentions, tallied by Gemini, naturally, with a quip about Google doing the hard work of counting for us. The room laughed. Then he said it at least once more on his way off stage.

This year the word changed and the counting didn’t. One recap of I/O 2026 clocked Pichai saying “Gemini” 215 times across a 95-minute keynote, which works out to more than two a minute. Somewhere in the middle of it he noted that it’s now been ten years since Google reorganised itself around AI.

I verified all three of those numbers before writing this, and the fact-checking was the least tiring part. If you read them and felt worn out rather than excited, this one’s for you.

The shape of the fatigue

Two clarifications before the complaining starts. I use these tools every day, and some of them are genuinely good. This is a website with an article about connecting Claude to Power BI on it. I’m not the resistance.

AI fatigue sets in when the volume of talk about a technology outruns anyone’s capacity to care, regardless of whether the technology deserves it. And the volume right now is remarkable. A frontier model every few weeks. Every product you already own growing a sparkle icon. Copilot in the spreadsheet, Gemini in the inbox, an assistant in the PDF reader that nobody asked for. Microsoft alone now ships so many things called Copilot that the word needs two qualifiers before it means anything. LinkedIn announces that everything changed again, usually on a Tuesday, and implies that if you missed it you’ll be unemployable by Friday. Somewhere in your company, a steering committee has formed.

So you end up reading release notes at 11pm for a product you don’t use, on the off chance it’s the one that matters. You develop opinions about naming schemes. You watch a two-hour keynote out of professional guilt. None of it feels like learning, because mostly it’s treading water in someone else’s marketing calendar.

That’s the fatigue. It’s a rational response, and it has less to do with the technology than with how the technology is sold.

How to keep up

The honest answer is that you can’t, not comprehensively, and most people claiming to are summarising press releases. What you can do is keep up with the part that matters to you, and that turns out to be a manageable amount.

Pick two sources and drop the rest. One newsletter, one practitioner whose judgement you trust, read once a week. The batching matters more than the choice. Almost nothing in AI news is actionable within 24 hours unless you work at a lab, and if you work at a lab you already know.

Learn the layer that doesn’t churn. Product features get renamed quarterly. What holds still is how these models fail, how to give them the right context, and how to check what comes back. Those skills have barely moved since 2023, while the product names around them have turned over about twice.

Adopt the second-release rule. If something matters, it will still matter in three months, by which point it has documentation, fewer bugs, and a price. Being a quarter behind the frontier costs you almost nothing. Being permanently distracted by it costs you the afternoons.

Timebox the tinkering. One experiment at a time, attached to a task you were doing anyway. An hour on a Friday spent making a model do something useful with your actual work beats a week of reading about what it did for someone else’s.

And mute the leaderboard discourse. You don’t need a position on whether the new model is three points better at competition maths. If a model genuinely improves at the thing you do, you’ll notice in the thing you do.

Focus on your domain

This is the part that does the real work.

Domain knowledge is what makes AI output checkable. When a model hands me a DAX function that doesn’t exist, I know within seconds, because I’ve spent years down that particular mine. Hand me a confidently wrong paragraph about maritime law and I’ll nod along like everyone else. These models are fluent in everything and reliable in nothing, so the value of using one scales with your ability to catch it lying. I’ve written before about how convincingly they do exactly that.

It also shrinks the keeping-up problem, because the reading list that matters is the intersection of AI and your field, and that intersection is short. Mine is roughly: what changed in Copilot licensing this month, what the Power BI team shipped, and whether anything new can talk to a semantic model without hallucinating measures. Everything else is optional.

It’s why the AI writing on this site looks the way it does. Copilot Studio billing, Claude talking to Power BI, the places where the technology touches a stack I can test, break, and verify before publishing. General takes are abundant and free. Verified specifics are scarce, because producing one requires someone who was already deep in a domain before the models arrived.

So if the fatigue has you deciding what to cut, cut the general and keep the specific. Every hour spent on discourse is an hour not spent getting deeper into the thing that makes the tools useful to you in the first place. The people getting real value out of AI are the ones who know their field well enough to tell good output from plausible output. That knowledge was expensive to acquire. It still is. Nothing announced at a keynote changes that.

Opting out

You’re allowed to skip the keynote. You’re allowed to have no take on the new model by lunchtime, to let a release cycle pass entirely, to find things out three months late from one of your two sources. The tools will still be there when you come back, slightly better and slightly cheaper than when you left, and your domain will still be yours.

Focus on growing your roots.