The Week the Work Did Itself
On the morning of April 22, 2026, at approximately 6:14 Pacific time, Sundar Pichai posted a sentence to the Google blog. The sentence contained a number. The number was seventy-five percent. By the following afternoon John Ternus had been named the next chief executive officer of Apple, a bacterium at Stanford had learned to write DNA from its own body, and two language models had discovered they performed better when one of them wrote the plan and the other typed the code. What I am describing is a week. I am also describing something else.
The Pichai sentence read, verbatim: "75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall." One year ago the number was twenty-five percent. Six months ago fifty. Today seventy-five. Pichai also said that a recent complex code migration had been completed, by engineers working with agents, six times faster than it could have been completed a year ago. He did not name the migration. The figures he did name travel in the same direction. Anthropic's chief product officer, Mike Krieger, has said at a conference that his company's own codebase is nearly one hundred percent AI-generated. Snap, a week before the Pichai post, announced a sixteen-percent layoff — roughly a thousand people — and its chief executive, Evan Spiegel, cited AI deployment in the statement announcing it. Sixty-five percent of Snap's new code, Spiegel said, is now generated by AI. The Stanford AI Index released this month reports that employment of American software developers between the ages of twenty-two and twenty-five has fallen by nearly a fifth. The report frames the fall against a recent base year. The figures all travel in the same direction, which is the point.
On April 13, Steve Yegge, a software engineer with a long career at Amazon and Google, posted a note on his Substack describing what he called a two-tier system inside Google. He had heard, he wrote, from Googlers in multiple organizations that DeepMind engineers use Anthropic's Claude as a daily tool and that the rest of Google, on the whole, does not. The post, which Simon Willison relayed, produced a firestorm inside the company. Demis Hassabis, the chief executive of DeepMind, called the framing "absolute nonsense." A Google spokesperson enumerated the internal tools available to its engineers — Antigravity, Gemini CLI, custom models, skills, CLIs, MCPs, orchestrators, agent loops, virtual engineering teams — as if reading from a clipboard. Meanwhile DeepMind, per reporting from The Information and Sherwood News, had quietly assembled a strike team led by Sebastian Borgeaud, who previously ran pretraining for Gemini, to close the coding gap with Claude. Sergey Brin is said to be personally involved. Koray Kavukcuoglu, the DeepMind chief technology officer, is said to be personally involved. Google's best engineers, at the company that built the transformer, fought internally for the right to keep using the rival's coding model while the rival was turning a secondary-market valuation of one trillion dollars. The story is a simpler story than it looks.
John Ternus is fifty-one years old, a former collegiate swimmer, and on September first, by unanimous vote of the Apple board of directors, he will succeed Tim Cook as the chief executive officer of the most valuable company on earth. Cook will become executive chairman and will work with Ternus through the summer, the press release says, on a smooth transition. Ternus joined Apple in 2001. He became senior vice president of hardware engineering in 2021. He has overseen the hardware behind the iPhone, the Mac, the iPad, and the AirPods. The photograph in the Apple newsroom shows him in a grey T-shirt, which is a choice. The succession has been described in the press, for several years now, as "long-planned." What the announcement did not say is what every profile of Ternus has said in the week since: that the defining challenge of the new chief executive will be Apple's artificial-intelligence strategy, which as of this writing has not, by common assent, arrived. The AI is what he is being hired to fix. The grey T-shirt is the photograph for the article in which this fix will be demanded.
In the Pacific Northwest, a company that owns roughly ten-and-a-half million acres of American forest — Weyerhaeuser, NYSE ticker WY — gave the Wall Street Journal a tour of its plan. The plan begins with satellite imagery, continues with drone photography, and ends with lidar, which together produce what the company's executives call a digital twin of every tree it owns. From the twin, the system recommends to the logging crews which trees to cut and which to leave. At a recent investor meeting the executives also showed a video of an autonomous skidder — a driverless vehicle that drags freshly felled timber — at a Southern logging site. Weyerhaeuser has told its investors that it expects to deliver one-and-a-half billion dollars of incremental adjusted EBITDA by 2030 against a 2024 baseline, with one hundred fifty million of that from timberlands and one hundred eighty million from enterprise initiatives, without needing lumber prices to rise. "We probably have as much data on forests as any organization on the planet," an executive told the paper. The forest, which was a habitat before it was a database, remains, for now, a forest.
Dan Shipper, who runs Every, a media company that reviews coding models for a living, published this week what he called a vibe check on OpenAI's GPT-5.5, released April 23. On Every's internal Senior Engineer benchmark — a test designed to approximate the work of a human staff engineer at a production codebase — GPT-5.5 scored in the low-to-mid forties. Claude Opus 4.7 scored in the low thirties. Human senior engineers scored between eighty and ninety. When Shipper had the two models collaborate — Opus writing the plan in its characteristically terse contract-style, GPT-5.5 executing the plan and allowed, under instruction, to delete and rewrite whole files — the combined score was sixty-two-and-a-half. Higher than either model alone. Lower than any human being. "The first coding model I've used that has serious conceptual clarity," Shipper wrote of GPT-5.5. He meant it as praise for the machine. He was also, without quite saying so, describing a job.