Anthropic's Honest Confession. AI Builds the Next AI
Claude now writes 80%+ of Anthropic's code, engineers are 8x more productive, and Claude's research speedup hit 52x in April 2026, beating skilled humans.
Three futures are possible: plateau, continued human-directed compounding, or full recursive self-improvement with minimal human involvement.
Anthropic says it would back a verified global pause on frontier AI — but only if all major labs agree, since a unilateral pause just hands the lead to someone else.
The real danger isn't capability itself; it's humans slowly losing the ability to oversee how AI is built, not just what it does.
Source: Anthropic