AI labs shift focus to recursive self-improvement, but definitions vary

In short: More AI labs are using the term recursive self-improvement for AI that can improve itself, but researchers disagree on what counts and say it is not happening yet.

What's going on

Recursive self-improvement, often shortened to RSI, is getting more attention in AI research. The basic idea is simple to describe. It means an AI system that can keep upgrading itself over and over, like a worker who can rewrite their own training manual and then work faster the next day.

TechCrunch reports that new and existing teams are starting to point to RSI in their plans. Researcher Richard Socher recently launched a company called Recursive Superintelligence with RSI as an explicit goal. Other efforts aim at pieces of the same puzzle, like Alex Karpathy’s Auto-Research project, which uses many small AI “agents” (software helpers that follow steps) to make incremental improvements to a model.

There are also examples of AI doing more of the work inside companies. TechCrunch notes comments from Anthropic staff suggesting AI tools are writing a large share of code in some teams. Still, experts caution that using AI tools to help researchers is not the same as an AI lab running end to end without people.

What to watch

A key question is how to measure progress. Researchers at Georgetown’s Center for Security and Emerging Technology and others describe milestones, from AI being able to keep working without humans at all, to matching human researchers, to outperforming human and AI teams combined. For now, even industry leaders like Google CEO Sundar Pichai say the “next level of acceleration” implied by RSI is not here yet.

Source: TechCrunch AI