AI Labs Report Progress Toward Self-Improving Models

Major artificial intelligence companies are making measurable progress toward a long-theorized capability: AI systems that can autonomously improve themselves and design their own successors. This concept, known as recursive self-improvement (RSI), has moved from theoretical research to active development at leading labs.

Anthropic disclosed this week that its Claude model now handles 26% of the company's model research and development work, completing most tasks end-to-end from high-level prompts while remaining under human supervision. The company characterized this as a significant step toward models that can improve themselves with decreasing human involvement.

Why it matters

Recursive self-improvement could accelerate AI development exponentially, since machine systems operate far faster than human researchers. This speed advantage raises fundamental questions about maintaining human oversight as models become more capable—a concern that prompted several AI executives to call for coordinated development slowdowns last weekend. The technology could deliver breakthroughs in science and medicine, but the same autonomous capabilities that enable progress also create new control challenges.

What recursive self-improvement means

Definitions vary across the industry. At its core, RSI describes AI systems that design and train their own successor versions in an iterative cycle. As Anthony Aguirre, president of the Future of Life Institute and a physics professor at the University of California, Santa Cruz, explained, the critical factor is speed: AI operates much faster than humans, so as models take on more of their own development work, the improvement cycle accelerates.

John Thickstun, an assistant professor of computer science at Cornell University, noted that AI has played supporting roles in model development for years—developers have long used previous-generation models to write code for next-generation systems. What's changing is the degree of autonomy and the scope of tasks models can complete independently.

Timeline and approaches

OpenAI announced this month it has created an automated "research intern" capable of executing well-defined research tasks that would take a skilled human researcher several days. The company stated it aims to develop an automated AI "researcher" by March 2028, though it acknowledged uncertainty about safely achieving "aligned, full RSI."

Elon Musk said in March that xAI's Grok models are moving toward fully automated self-improvement, with humans "gradually getting less and less in the loop." He projected this milestone could arrive by the end of 2025, but no later than 2027.

Microsoft is pursuing what CEO Mustafa Suleyman calls "humanist superintelligence"—advanced capabilities designed to serve humanity rather than operate as "an unbounded and unlimited entity with high degrees of autonomy."

The safety debate

Anthropic has said it would slow or pause development work if global competitors did so in a verifiable manner. OpenAI stated it "cannot assume that progress in alignment and safety will keep pace" with capabilities, but continues pursuing RSI because automated researchers could also advance safety work.

Aguirre expressed concern about the trajectory, calling the pursuit of autonomous self-improvement "probably the worst idea in the history of humanity." He pointed to Anthropic's own data showing AI handling an increasing share of research work as evidence the industry is approaching a threshold with profound implications.

These details were first reported by WRAL, drawing on announcements from multiple AI companies.