AI now writes most of Anthropic's code, and the bottleneck just moved

Information Technology 4 min 12 sources

Anthropic says more than 80% of the code it shipped in May was written by its own AI, Claude — eight times the old output per engineer — moving the constraint from writing code to reviewing it. Meanwhile the AI money keeps flowing even as doubt about whether the spending pays off drags tech stocks down, and Washington calls Nvidia's CEO in over chip sales to China.

Key takeaways

Anthropic says AI now writes more than 80% of its code, shipping eight times more per engineer — but the bottleneck just moved from writing code to reviewing it, and reviewing is now the scarce skill.
The AI money keeps flowing — a robotics startup at $2 billion, the fintech Ramp at $44 billion — even as doubt about whether the spending pays off drags tech stocks down.
Washington called Nvidia's CEO to testify on selling AI chips to China, while the old giants Microsoft and Apple reshape their oldest products, Windows and Siri, around AI.

The code that writes itself

Anthropic, the company behind the Claude AI models, made a striking claim this week [7]. More than 80% of the code added to its main software in May was written not by its engineers but by Claude itself [7]. The company frames it as a step toward “recursive self-improvement” — AI that helps build the next, more capable AI [7]. The output is real: Anthropic says its engineers now ship roughly eight times as much code per person each quarter as they did a few years ago [7].

The path there was gradual. Anthropic describes three stages. Engineers wrote code by hand through 2023. Then they pasted in snippets from early chatbots. Now they hand whole tasks to “coding agents” that write and wire up the work themselves [7]. Each stage moved more of the typing from the human to the machine.

But the bottleneck didn’t vanish. It moved. Eight times more code means eight times more code that someone, or something, has to check [7]. Writing used to be the slow step. Now it’s reviewing. For anyone who builds software, that’s the shift worth noting. The valuable skill is sliding from producing code to judging it — catching the bug, the security hole, the wrong assumption an agent wrote confidently and fast. The work didn’t disappear. It changed jobs.

The money keeps flowing — and the doubt with it

Away from the workbench, the cash kept moving. Generalist AI, a robotics startup backed by the chipmaker Nvidia, raised $400 million at a $2 billion valuation to build AI that controls robots [1]. Ramp, whose software helps businesses track and trim their spending, raised $750 million at a $44 billion valuation [2]. Investors, as one outlet put it, are hungry for “fintechs with an AI story” [3].

The valuations climb, and a question climbs with them: does the AI spending actually pay off? Ramp’s own pitch is that it helps companies rein in their AI bills — a tell about where the worry sits [3]. Ahead of Anthropic’s planned stock-market debut, its president Daniela Amodei waved off doubts about AI’s returns [8]. And the doubt showed up in prices. After the chipmaker Broadcom gave a softer forecast than investors wanted, tech stocks slid across Asia and dragged US names down with them [11]. A valuation is a bet on a story. This story is still being tested.

Washington eyes Nvidia’s China sales

In Washington, the fight over who gets the chips sharpened. Senator Elizabeth Warren, a Democrat known for grilling executives, invited Nvidia’s chief executive, Jensen Huang, to testify before the Senate about the company’s sales of AI chips to China [4]. Nvidia makes the processors that train almost every large AI model, which turns its export choices into a national-security question as much as a business one. Whether Huang appears — and what he says about the line between selling to a huge market and arming a rival — is the thread to watch [4].

The old giants replant their flags

The incumbents spent the week reshaping their oldest products around AI. At Build, its conference for developers, Microsoft put Windows back at the center of the keynote [6]. That’s a notable turn for a company that has spent years pushing its cloud and AI services ahead of the operating system [6]. Next week Apple holds its own developer conference [9]. It is expected to show a long-delayed overhaul of Siri built on Apple Intelligence — the company’s name for its on-device AI features [9]. Same story, two flags: the giants bending their flagship products toward the new technology.

A country bets on AI for jobs

Finally, a government placed its bet out loud. Canada said its national AI strategy would create 250,000 jobs and lift the country’s economy by 3% [10]. Those are projections, not results — the round, confident numbers governments attach to a plan they want to sell. But the direction is the point. Much of this year’s tech conversation has been about AI taking jobs; Ottawa is staking its pitch on AI making them. Which forecast proves right is a question every worker has a stake in.

02 · Lesson · why it matters

You don't remove a bottleneck. You move it.

Make one step of any process faster and the work piles up at the next step — the constraint relocates, it does not disappear.

The eight-times-faster trick that wasn’t

Anthropic said this week that its AI now writes more than 80% of its code, and that its engineers ship eight times as much code as they used to. Read quickly, that sounds like software got eight times easier to make.

Look again at the company’s own words and the catch is right there: eight times more code means eight times more code that someone has to review. The slow step didn’t go away. It moved — from writing the code to checking it. The typing got automated. The judging didn’t, and now there’s a mountain of it.

That move — from one bottleneck to the next — is the whole lesson. Once you see it, you see it everywhere.

Every process is a chain, and one link is always slowest

Think of any job as a line of steps, one feeding the next. Writing code, then reviewing it, then shipping it. Or: take the order, cook the food, wash the dishes. The speed of the whole line is set by its slowest step — the bottleneck. Everything upstream of it backs up; everything downstream of it sits idle, waiting.

Here’s the part people miss. When you speed up the slowest step, the line doesn’t become uniformly fast. It just has a new slowest step. The constraint doesn’t leave the building. It walks to the next room. You haven’t finished the problem — you’ve relocated it.

So Anthropic didn’t make software eight times easier. It made the writing eight times faster, which means writing is no longer what holds things up. Reviewing is. The bottleneck moved one step down the line.

The new bottleneck is usually the part you couldn’t automate

There’s a pattern to where it moves. The step that speeds up is almost always the mechanical one — the typing, the copying, the routine. The step it moves to is almost always the one that needs judgment — deciding if the code is right, if it’s safe, if it does what was actually wanted.

That’s not bad luck. Machines clear the parts that are easy to mechanise, and pile the load onto the parts that aren’t. So “we automated the writing” quietly becomes “we made the reviewing bigger.” The harder, more human step doesn’t just remain — it grows, because now it has eight times as much to chew through. Automating the easy half of a job tends to enlarge the hard half.

The same shape, far from any computer

This isn’t a software quirk. It’s the shape of almost every improvement.

Widen a congested highway and the traffic jam doesn’t vanish — it reappears at the next interchange that didn’t get widened. Hire a faster cook and the orders fly out of the kitchen until they pile up at a single overwhelmed dishwasher. Add an AI tool that floods you with job applicants, and your bottleneck is no longer finding people — it’s the human interviewers who now can’t keep up. Automate your inbox so messages sort themselves, and you discover the real pile was never the messages. It was the decisions inside them.

Each time, the same thing happens. Relieve one constraint, reveal the next. The work moves; it doesn’t end.

Why it surprises us every single time

If this is so common, why does it keep catching people off guard?

Because a bottleneck is invisible until the step in front of it clears. While the cooking is slow, nobody notices the dishwasher — there’s never a backlog there to see. The new constraint is hidden behind the old one. And the two arrive together: the win (this step is faster now) and the new problem (the next step is drowning) land in the same moment. We see the win, celebrate it, and get blindsided by the pile we didn’t know was coming.

Where the constraint went

Something in your work suddenly gets much faster — a new tool, an automation, an AI that does the tedious part. The useful question isn’t “are we done now.” It’s “where did the bottleneck go.”

It went somewhere. It almost always went to the step that needs a human to think, because that’s the step the machine couldn’t take. The job isn’t finished; it has a new slowest point, and that point is where the next real work lives. Anthropic’s engineers didn’t stop working when the AI started writing. They became the reviewers. The bottleneck moved, and so did they.

03 · Lab · your turn

Where Does the Jam Go

Speed up one step of a pipeline and watch the bottleneck jump to the next, landing on the human step that won't automate.