Anthopic did a thing...

2026-06-04 · Source: Matthew Berman · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, extended

Summary

Anthropic's recent paper, "When AI Builds Itself," details the company's progression towards recursive self-improvement (RSI) in AI development. By May 2026, Claude authored over 80% of code merged into Anthropic's codebase, a significant increase from low single digits in February 2025. The paper notes an accelerating trend where AI agents' task completion length doubles roughly every four months. In research, the internal Mythos preview model achieved a 52x speedup in code optimization experiments by April 2026, compared to Opus 4's 3x speedup a year prior. While AI systems are increasingly capable of execution and reproducing research, human judgment remains crucial for generating novel ideas and setting research directions. Anthropic also controversially suggests a global slowdown in AI development, a position seen as self-serving given their current technological lead.

Key takeaway

For Directors of AI/ML evaluating team structures, recognize that AI is rapidly automating code generation and research execution. You should strategically reallocate human talent from coding and routine experimentation to defining novel problems, setting research directions, and verifying AI outputs. This shift maximizes productivity by leveraging AI for "perspiration" while preserving human "inspiration" as the critical bottleneck. Prepare for increased demand in marketing and sales as development velocity accelerates.

Key insights

AI systems are rapidly automating their own development, shifting human roles towards high-level judgment and novel problem definition.

Principles

• AI automation shifts human roles to direction-setting.
• Execution is increasingly automated, making novel ideas paramount.
• Accelerating one process exposes new bottlenecks elsewhere.

Method

Anthropic's AI development evolved from direct human coding to AI agents writing and reviewing code, accelerating research by optimizing existing code and proposing experiments.

In practice

• Delegate code generation and optimization to AI agents.
• Employ AI judges for reviewing AI-authored code.
• Prioritize human effort on novel research and problem definition.

Topics

• Recursive Self-Improvement
• AI Development Automation
• AI Agent Productivity
• AI Research Acceleration
• Human-AI Teaming
• AI Alignment Risks

Best for: Research Scientist, Investor, CTO, AI Scientist, AI Engineer, Director of AI/ML

Related on AIssential

• It’s starting… — AI systems are increasingly automating their own development, accelerating progress while shifting human roles to high-level judgment.
• Why AI hasn’t replaced software engineers, and won’t — AI primarily automates the "execute" layer of software development, leaving human "decide" and "deliver" functions intact.
• Coding Agents Meet Data Science — AI coding agents boost developer productivity but introduce new bottlenecks and require specialized adaptation for data science workflows.
• AI Coding Assistants Haven’t Sped up Delivery Because Coding Was Never the Bottleneck — AI coding tools shift bottlenecks from coding to specification and verification, demanding new team structures and workflows.
• AI Generalists Are Quietly Changing Who Wins in Software — AI shifts software value from coding speed to strategic direction and cross-functional judgment.
• How Autonomous AI Agents Are Reshaping the AI Game of Productivity — Autonomous AI agents enhance productivity by independently managing tasks, freeing human focus for creative and strategic work.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Matthew Berman.