On June 4, two Anthropic researchers published a paper asking the entire AI industry to consider hitting pause. Marina Favaro and Jack Clark called it "When AI Builds Itself." The argument: frontier models are getting close to improving themselves faster than humans can supervise, and the world should keep the option to slow down before that happens.

Three days earlier, Anthropic had quietly filed to go public.

That's the whole story, really. A company that wants the planet to agree on a brake pedal had, seventy-two hours before, started the paperwork to race onto the public markets at a $965 billion valuation. Both things are true. Both are happening at the same desk.

The numbers underneath the pause proposal are the part nobody can wave away. As of May 2026, more than 80% of the code merged into Anthropic's own production codebase was written by Claude — not by people. Before Claude Code shipped in early 2025, that figure sat in the low single digits. The typical Anthropic engineer now ships roughly eight times as much code per day as in 2024. On the hardest internal coding tasks the company tracks, Claude's success rate climbed from about 26% to 76% in six months.

So the recursive self-improvement Favaro and Clark warn about isn't a future hypothetical. It's their Tuesday. The machine is already writing most of the machine.

Here's the context that makes the timing land funny. Anthropic's annualized revenue hit $47 billion in May, up from around $9 billion at the end of 2025 — a fivefold jump in five months. The $65 billion Series H that set the $965 billion price was reportedly oversubscribed. President Daniela Amodei went on stage at Bloomberg Tech and explained, plainly, that the IPO exists because training and serving these models costs a fortune and public markets are where the money is. Goldman Sachs pegs cumulative AI capex from 2026 to 2031 at $7.6 trillion — about a quarter of annual US GDP.

You do not raise that kind of money to slow down. You raise it to go faster.

And the pause Anthropic describes is engineered to be safe to propose. It only kicks in if every major lab in every major country agrees, with verifiable proof everyone actually stopped, plus triggers, adjudicators, and exit conditions. Translation: it requires the simultaneous cooperation of China, the US government, OpenAI, Google, and a dozen others, on a verification system that does not exist. Anthropic gets the halo of raising the alarm while facing roughly zero chance of ever having to ease off the gas.

My Opinion

I'll be blunt. This reads less like a conscience and more like a press release with footnotes. Bentley University's Noah Giansiracusa put it cleaner than I can: "I do not think it is a genuine call to slow down." When the only version of a pause you'll endorse is the one that can never actually trigger, you've written a hedge, not a warning.

What bugs me is that the underlying fear might be real. Eighty percent of your code written by your own model is a genuinely strange place to be standing. If Anthropic believes that, the honest move is to slow its own training runs and eat the competitive cost. Instead it filed an S-1 and asked everyone else to go first. That's not leadership. That's wanting credit for the worry without paying for the fix.

The tell is the IPO itself. A pause freezes the board at the exact moment Anthropic holds one of the top two or three seats. New entrants get locked out, incumbents consolidate, and the company calling for the freeze happens to be winning when the whistle blows. I don't think that's an accident — and I don't think the next lab to publish a "we're worried" manifesto will be acting alone either. Watch who files to go public the same week.

Author: Yahor Kamarou (Mark) / www.humai.blog / 16 Jun 2026