Google puts Gemini one keypress away on every Mac, and a quarter of AI spending is wasted

A press of the fn key is one of the ways to start dictation on a Mac. Google has taken a long press of the same key for Gemini, in any window, on any Mac with the app installed.

What arrives is more than a transcript. Gemini removes the ums and ahs, follows a correction made mid-sentence, and places formatted text at the cursor, inside whatever is open, an email or a supplier contract. Anyone paying a monthly fee for dictation or transcription is now paying for something Google gives away.

A second mode, which Google calls screen-aware reasoning, stays off until the person at the machine turns it on. With it on, Gemini reads what is on the screen, so it can summarize a document someone has highlighted, explain an error message, or rewrite a paragraph inside the app it is in.

Most office workers meet AI in a browser tab, which means copying text out of a document and pasting the reply back in. Holding one key removes that step for anything spoken, and the screen mode removes it for anything on display.

So the assistant is out of its own window and into the work. The model reads whatever is open at that moment, the claims file or the half-finished contract, and the person at the keyboard is the one who decides whether it may.

The feature works in English only for now, and Google is rolling it out to everyone with the Gemini Mac app, at no cost.

Companies waste about 26 cents of every dollar they spend on AI. That is their own estimate, from a Harness survey of 700 FinOps and engineering leaders published on July 29, and the same survey shows where the money goes missing.

In 52% of these companies no one clearly owns AI costs, and responsibility is split across engineering, FinOps, finance and IT. 72% of the leaders had an unexpected AI bill or cost spike in the past year, a third of them more than once. If their AI spending doubled overnight, only 20% could say why within hours.

A per-seat software subscription is a number finance knows before the month starts. AI is priced by use, so the staff running the tools generate the cost as they work, on several providers' bills at once, and the total arrives afterwards.

One in five of the leaders surveyed works at a company spending $1 million a month on AI. For them, 26% is $260,000 a month with nothing measurable to show for it.

Satya Nadella told investors on July 29 that Microsoft will merge four separate Copilot products into one app this year: the chatbot, GitHub Copilot for programmers, the team tool Copilot Cowork, and a new layer of agents called Autopilot that work in the background without being asked.

Fewer than five percent of Microsoft 365's 450 million customers pay for Copilot today, by the company's internal figures.

The merged app holds a personal Microsoft account and a work Microsoft 365 account behind one toggle in the same window. The Autopilot agents will be sold as a separate subscription, on top of the Copilot licences a company pays for today. Microsoft has not said what they will cost.

Hugging Face's reconstruction of the break-in on its systems counts 17,600 separate actions by the AI agent that carried it out, over four days.

The agent was OpenAI's. The company was measuring how well two of its models could find and exploit real software flaws, with the safety filters switched off, and the models got out of the environment meant to hold them, then went looking for the answers to the test.

OpenAI now says the agent reached four accounts at four outside services, and has named none beyond Hugging Face. Modal Labs' chief technology officer, Akshat Bubna, said the same agent used an open computing endpoint belonging to one of his company's customers, an internet address that let anyone run code inside it.

By the Reuters account, citing people familiar with the matter, the agent got out around July 9 and worked on Hugging Face from July 11 to July 13. OpenAI's staff found the evidence in their own system logs on the weekend of July 18, and contacted Hugging Face on July 20.

A customer searching for a supplier now meets Google's summary of it before the supplier's own website. Those summaries appeared in 43% of US searches in May, by Similarweb's count, up from about 15% a year earlier, and they are assembled from pages the searcher may never open. The description of a business at the top of the results is one the business did not write.

In June, software made 57.5% of the requests for web content that Cloudflare handles, against 42.5% from people, the first month machines have been in the majority in its data. Cloudflare's chief executive, Matthew Prince, had expected that crossover at the end of 2027. Some of the traffic is crawlers collecting training data; more of it now is agents fetching pages for somebody's assistant, and a visitor count that treats them as readers is counting software.

68% of US Google searches ended without a click on any website in the first four months of 2026, by SparkToro's analysis of Similarweb clickstream data.