It’s starting…

TL;DR

• 1

  Anthropic says recursive self-improvement is visible, but not complete: Berman highlights the paper's core claim that AI is taking over more of AI development, yet still lacks the key missing ingredient of truly novel research taste and goal-setting.

• 2

  Claude now writes most of Anthropic's code: As of May 2026, Anthropic says over 80% of merged code was authored by Claude, up from low single digits before Claude Code's February 2025 research preview.

• 3

  Task horizon is expanding fast: Berman calls out Anthropic's chart showing autonomous task length doubling every four months, from Claude Opus 3 handling roughly 4-minute software tasks in March 2024 to Opus 4.6 handling 12-hour tasks a little over two years later.

• 4

  Research reproduction is nearly saturated, original insight is not: On benchmarks like CORE-Bench, AI systems went from reproducing research results about 20% of the time in 2024 to nearly 100% 15 months later, but Berman stresses that coming up with the original idea still belongs to humans.

• 5

  Anthropic's own data suggests AI code is prolific, not clearly better: Employees estimated Mythos preview made them about 4x more productive while code output per engineer rose about 8x, which Berman reads as evidence that AI-written code was at best around parity with human code, not obviously superior.

• 6

  The biggest critique is Anthropic's politics, not its charts: Berman repeatedly says Anthropic's suggestion that the world should slow frontier AI development sounds like fear-based marketing from a company that benefits most if everyone else eases off while it keeps compounding gains internally.