AI’s expansion from chat and coding into broader knowledge work includes the work of developing AI itself. Systems already write research code, run experiments and improve training efficiency.

Recursive self-improvement (RSI) is an iterative process in which an AI system designs, implements and evaluates changes to its own code, architecture or learning methods. Each improved version then uses its enhanced capacity for AI development to repeat the cycle.

Sustained, fully autonomous development of successive frontier AI models remains a research goal. Such a loop could significantly accelerate AI research.

For alignment, the challenge is to keep successive generations acting consistently with human intentions and values. An aligned predecessor does not guarantee an aligned successor. Each requires evaluation and effective means to redirect or pause development, especially if capability growth outpaces our ability to assess it.

As the scope of AI execution expands, governance needs to address whose interests are represented, how benefits are realised and risks mitigated, and how people can challenge decisions that affect them.

Human accountability extends to what we authorise AI to build and how those systems affect society. The opportunity is to direct that growing capability towards discoveries and applications that improve people’s lives.