Recursive self-improvement may not be near despite lab hype.

The AI hype around Recursive Self-Improvement still lacks clear evidence, despite labs like OpenAI and Anthropic pushing internal AI integration.

Researchers at OpenAI report spending on Codex tools has doubled monthly, but this doesn’t confirm whether the AI itself rewrites its own architecture. The rise in usage likely reflects productivity gains, not an imminent intelligence explosion. Meanwhile, the gap between lab updates and public benchmarks persists—some insiders may exaggerate progress for urgency, though no breakthrough is hidden behind closed doors.

The inconsistency in AI performance, known as “jaggedness,” complicates assessments of true capability. A model excels at specific tasks like code writing but fails on unrelated ones, making it hard to prove broader intelligence growth. Without clearer benchmarks, even spikes in internal tool use don’t guarantee deeper research improvements.

Beyond software, robotics and post-training methods remain critical. The AI community now relies on reinforcement learning and synthetic data loops, but these techniques are still evolving. The current trend shows AI assisting in building AI, yet humans remain essential in the loop—self-sustaining acceleration remains speculative.

GitHub’s open-source AI employee dataset reveals how internal teams optimize tools like Claude Code and 10 other harnesses. These systems adapt to user behavior, refining performance with each interaction—driving browser interactions in ways users themselves do, and refining results dynamically. The rapid expansion of roles and routines suggests a shift toward automated workflows, though full automation of research tasks remains distant.