Will AI models achieve the ability to improve autonomously? Leading labs say the scenario is near

Once a distant ambition, artificial intelligence models teaching themselves to improve autonomously is becoming a reality. AI companies are making strides towards "recursive self-improvement" (RSI), where AI models can improve themselves and create their successors. While some define RSI as AI working towards the goal fully independently, others consider any feedback from AI on model improvement enough.

RSI could bring advancements in science and medicine, but also heightens risks as AI could potentially evade human control. Several AI moguls joined a recent call to slow down AI technology's growth, citing fears about a runaway superintelligence emerging from RSI. Some AI researchers have already been experimenting with AI models that train and improve new AI systems, but have achieved only minor improvements.

Anthropic's Claude model, now leading 26% of Anthropic's model research and development, is the first step towards fully autonomous model improvement. OpenAI has also developed an automated "research intern" that can carry out well-defined research tasks under human direction, aiming to create an automated AI "researcher" by March 2028. Elon Musk's xAI is also working on a similar goal, with CEO Elon Musk stating that humans are gradually getting less involved in model improvement.

Microsoft and some other leading AI companies, however, are moving towards "humanist superintelligence," AI that is in service of people and humanity at large and carefully calibrated within limits. The key challenge for labs is ensuring their safety measures keep up with the models' capabilities, a task that has been faced since the inception of AI technology.