Anthopic did a thing...

TL;DR

• 1

  Anthropic says Claude wrote 80%+ of its merged code by May 2026: Berman treats that as a major industry signal, especially because Anthropic claims this jumped from low single digits before Claude Code launched in February 2025.

• 2

  Task horizon is accelerating fast, from minutes to hours to potentially days: He highlights Anthropic's claim that models went from handling 4-minute software tasks with Opus 3, to 90-minute tasks with Sonnet 3.7, to 12-hour tasks with Opus 4.6.

• 3

  The missing ingredient for full recursive self-improvement is research taste, not coding ability: Models can reproduce papers and execute experiments, but Berman says they still struggle with deciding what truly novel problem to pursue next.

• 4

  Anthropic's own numbers suggest AI code is prolific but not yet clearly better: Employees estimated roughly 4x productivity gains with Mythos preview, even while code output rose about 8x, which Berman reads as a sign that AI-written code was still lower quality or harder to absorb operationally.

• 5

  Research automation is improving at a startling rate: He calls out Anthropic's benchmark where Opus 4 found 3x speedups in May 2025, while Mythos preview reached 52x by April 2026, plus CoreBench moving from about 20% success to near benchmark saturation in 15 months.

• 6

  Berman sees Anthropic's safety language as strategic as much as sincere: He repeatedly frames the paper as 'Anthropic coded', arguing the company is asking the world to slow down while using internal models like Mythos to extend its own lead.