AI systems are beginning to take on more of the research and engineering used to develop better AI — raising a bigger question: what happens when improving AI becomes one of AI’s own jobs?
WHAT’S HAPPENING
AI researchers are increasingly discussing recursive self-improvement — RSI: the possibility that AI systems could eventually help design, test and improve the systems that succeed them.
In a recent discussion with Dwarkesh Patel, researchers John Schulman, Beren Millidge and Charlie O’Neill examined how close today’s systems are to that feedback loop — and the technical barriers still standing in the way.
The process has already started in a limited form.
Anthropic says AI now performs a growing share of its own AI-development work, including writing code, running experiments and assisting with research. But the company explicitly says full recursive self-improvement has not been achieved and is not inevitable.
WHY IT MATTERS
Today, humans still largely decide what problems AI should solve, what experiments matter and whether an improvement is actually useful.
But if AI becomes capable of handling more of that process, the development cycle could change:
Humans build AI → AI helps build better AI → better AI becomes more capable of improving the next generation.
That could accelerate AI development without requiring every improvement to originate with a human researcher.
WHO BENEFITS
AI laboratories and researchers could test more ideas, run more experiments and improve systems faster.
The same capabilities could eventually accelerate scientific research, software development and other fields where experimentation can be automated.
WHO LOSES
The faster the development loop becomes, the harder it could become for human researchers, companies and regulators to evaluate each change before the next one arrives.
There is also a basic technical problem: today’s AI systems do not consistently improve themselves across repeated attempts.
Recent research shows agents can discover improvements, but later iterations frequently fail to preserve those gains.
WHAT HAPPENS NEXT
Researchers are now testing whether AI can move beyond carrying out experiments humans define and begin making more of the decisions required to conduct AI research itself.
The important distinction is that AI has not reached autonomous recursive self-improvement.
But the first pieces of the loop are appearing.
The next major acceleration in AI may come when building better AI becomes something AI itself is increasingly capable of doing.
