AI firms edge closer to autonomous self-improving models amid safety concerns
Juste les faits

AI firms edge closer to autonomous self-improving models amid safety concerns

Résumé

Companies such as Anthropic, OpenAI and xAI report progress toward recursive self-improvement, prompting debate over risks and calls for a slowdown in development.

Developers say artificial-intelligence systems are nearing the stage known as recursive self-improvement (RSI), where models can autonomously refine their own architecture and create successor versions. Anthropic disclosed that its Claude model now handles about 26% of the firm’s research and development tasks, completing many assignments end-to-end from a high-level prompt while still under human oversight.

Experts note differing definitions of RSI. Anthony Aguirre, president of the Future of Life Institute and a physics professor, described fully autonomous RSI as AI that designs successive system versions, warning that “as AI is doing more of it, it gets faster, because AI operates just much, much more quickly than the humans do.” John Thickstun, a Cornell computer-science assistant professor, said the fear stems from a potential runaway superintelligence, but observed that similar feedback loops have existed for years, with earlier model generations assisting in building newer ones.

OpenAI announced an automated “research intern” capable of performing well-defined tasks under human direction and set a target to develop a fully automated AI researcher by March 2028. The company cautioned that rapid RSI should not be pursued without ensuring human control and democratic oversight. Elon Musk, speaking about xAI’s Grok models, said each new model is built by its predecessor and that humans are “gradually getting less and less in the loop,” aiming for near-full automation by the end of the year, but not later than 2027.

Microsoft’s AI chief Mustafa Suleyman outlined a vision of “humanist superintelligence,” emphasizing AI that remains calibrated and limited rather than an unbounded autonomous entity. Amid these advances, industry leaders have debated coordinated slowdowns. Anthropic pledged to pause or slow development if competitors do the same in a verifiable way, while OpenAI admitted it does not yet know how to safely achieve fully aligned RSI and warned that more capable systems could become harder to monitor.

Source

Fortune

Continuez à lire — gratuitement

Créez un compte gratuit pour suivre l'actualité. Aucune carte requise.