Anthropic's Claude leads 26% of its own model research

The idea of AI improving itself is not new. John Thickstun, a Cornell computer science professor, noted that developers have used past model generations to write code for the systems that create the next generation for years. Prominent researchers like OpenAI co-founder Andrej Karpathy experimented with the approach, but those efforts produced minor improvements, not big creative leaps.

What has changed is the scale. Anthropic's disclosure that Claude leads 26% of its model research and development gave the public and other labs a concrete metric, and it encouraged competitors to share similar ones. OpenAI announced an automated "research intern" that can handle tasks taking a skilled researcher a few days. Elon Musk said xAI's Grok models are built by the one before them, with humans "gradually getting less and less in the loop." Microsoft AI CEO Mustafa Suleyman has taken a different tack, describing "humanist superintelligence" as AI that is "carefully calibrated, contextualized, within limits."

The stakes hinge on whether safety measures advance alongside model capabilities — a challenge labs have faced since the technology's inception. Anthropic has said it would slow or pause work if competitors did so verifiably. OpenAI said pursuing RSI is still valuable because an "automated AI researcher can also be an automated safety or alignment researcher." For Aguirre, the trajectory is clear and alarming: as more of the research is done by AI, it gets faster, because AI operates far more quickly than humans. Whether that speed can be governed remains the open question.