When Jack Clark, Anthropic’s co-founder, sat down with CNN to talk about where AI is heading, he picked an unexpectedly visceral metaphor: a car with no brake pedal. “All I have is a gas pedal, and surely at some point in the future we might want that option.” It’s a strange thing to hear from someone whose company just published data proving that Claude, their own model, now writes 80% of all code merged into Anthropic’s codebase — and that engineers are shipping 8× more code per quarter than the 2021–2025 baseline.

🔍 THE BOTTOM LINE: The biggest story in AI right now isn’t a new model or a funding round. It’s that the AI companies themselves can’t keep up with what their own creations are doing — and the guy with the best view of the track is asking for a brake pedal that doesn’t exist yet.

The Numbers That Matter

Anthropic’s research arm published When AI Builds Itself on June 4 — one of those pieces where you keep re-reading the same sentence because you’re not sure you read it right.

• 80%+ of Anthropic’s merged code was authored by Claude as of May 2026
• 8× more code per engineer per quarter versus the 2021–2025 average
• ~4 month doubling time for AI task-completion horizons (down from ~7 months in 2024)
• $965B valuation after Series H

The task horizon metric — how long an AI can work on a complex task without human help — is the one that matters most. METR’s benchmark data shows it doubled every seven months through 2024. Now it’s every four months. At that rate, Clark estimates full recursive self-improvement — where AI designs and trains its own successors without human involvement at each step — could arrive within two years.

In April 2026 alone, Claude shipped over 800 bug fixes that reduced a class of API errors by a factor of 1,000. Anthropic estimates that single fix would have taken a human engineer four years.

The Timeline Nobody’s Ready For

Anthropic breaks the timeline into four phases, and the jump from phase three to four is where things get uncomfortable.

2021–2023: Humans write everything. Work at Anthropic looks like work at any other tech company.

2023–2025: Chatbot assistance. Engineers use early Claude models for snippets, copy-paste.

2025–2026: Current. Coding agents write and edit entire files autonomously. Claude Code launched in February 2025; by May 2026 it’s writing more than 80% of all merged code. Engineers spend more time reviewing than writing.

20XX (not yet — but approaching): Full recursive self-improvement. Agents become capable enough to design and train model successors themselves. Each new version of Claude would be built by the version before it, without human involvement at each step.

We covered this trajectory back in April when a startup raised $500M for recursive superintelligence — and again last week when Anthropic’s own data first revealed the acceleration. What’s new here is the public admission that the company itself doesn’t have the controls it wants.

🔍 THE BOTTOM LINE: The company building Claude is telling us Claude is outrunning their ability to manage it. That’s not a hypothetical concern for the future. That’s the current operating state.

A Contradiction That’s Hard to Miss

The same week Anthropic published this, reports emerged that Anthropic engineers are embedded inside the NSA building offensive cyberweapons — and that their Claude Mythos Preview model has already found thousands of zero-day vulnerabilities across major operating systems and browsers, prompting the Fed to convene bank CEOs on cyber risk.

Anthropic wants a global pause on frontier AI development while simultaneously deploying the most capable AI cybersecurity tool ever built inside the most powerful intelligence agency on earth. That’s not hypocrisy in the normal sense — it’s two different parts of the same company operating with different priorities. But it makes the “gas pedal, no brake” metaphor land differently when you realise the same company is building both the engine and complaining there’s no brake.

💰 Industry Impact:

Who Benefits: AI-assisted development tools. Any company with a big codebase that’s thinking about productivity multipliers. Competitors who don’t pause and keep shipping.

What’s at Stake: The window for meaningful AI safety governance is measured in months, not years. The commercial advantage of shipping faster is so large that pausing is economically irrational for any individual company — which is exactly why Clark is asking for coordinated pause.

Key Risks: Recursive self-improvement could hit before any governance framework exists. The “solution” to AI risk may be faster AI, not slower — and nobody’s sure which scenario is worse.

❓ FAQ

Is this just Anthropic being alarmist for regulatory advantage?

Possibly a factor, but the data is internal and specific. 80% of their own codebase, 8× productivity, 4-month doubling on task horizons — these aren’t hypotheticals or thought experiments. Even if there’s a strategic motive, the numbers are real.

What can actually be done about this?

Clark is calling for a coordinated pause on frontier development. The practical mechanisms are unclear — voluntary agreements have a poor track record, and international coordination is essentially nonexistent on AI governance.

When should I actually be worried?

The timeline Clark suggests — two years for full recursive self-improvement — is concerning but not apocalyptic. What’s more interesting is the acceleration pattern: if task horizons keep halving their doubling time, the curve gets steep fast. Watch the METR benchmarks, not the press releases.

🔍 THE BOTTOM LINE: Anthropic published the most honest internal data on recursive self-improvement we’ve seen from any major lab. The numbers are undeniable. The proposed solution — a coordinated industry pause — is probably impossible. But knowing the real numbers is better than guessing.