This Week's Term: Recursive self-improvement (RSI) - the point at which AI systems can design and build their own successors with little or no human involvement, creating a loop where each generation of AI speeds up the development of the next.
In the past few weeks, RSI moved from a long-term thought experiment to the centre of the AI industry's agenda.
On June 4, Anthropic published a paper called "When AI Builds Itself", arguing that RSI "could come sooner than most institutions are prepared for". The evidence comes from their own operations. Claude now writes more than 80% of the code merged into Anthropic's systems, up from low single digits before Claude Code launched in early 2025, and their engineers ship roughly eight times more code per quarter than they did a few years ago.
OpenAI has gone further and put dates on it. The company has set internal goals of an automated AI research intern by September 2026 and a fully automated AI researcher by March 2028.
We are not there yet. Researchers like Helen Toner of Georgetown's CSET point out that engineers leaning heavily on AI is a different thing from classic RSI, where no human is needed in the loop at all. Today's models still struggle with exactly what a closed loop depends on. They lack the self-direction and judgment to run their own research, and they cannot yet manage the kind of ambiguous, week-long task that has no clean specification. But the gap is closing.
Why it matters for business
You do not run an AI lab. Every AI plan you make, though, rests on an assumption about how fast capability will improve. Even partial automation of AI research squeezes the time between model generations. That means any strategy built on "AI cannot do this yet" has a shorter expiry date than you assumed when you wrote it. This is closely related to the question of how much autonomy you hand a system, which I unpack in the five levels of AI autonomy, and it is the backdrop to this issue's piece on executives building software themselves.
The useful part is that you do not need to believe in an intelligence explosion to act on this. The pace of change becomes the one assumption in your plan you should revisit most often, and the one you should be least surprised to find moving faster than your roadmap assumed.
Recommended resource
Watch Matt Baxter's 15-minute non-technical explainer on how RSI works as a feedback loop and why it could lead to an intelligence explosion. It is a clear, jargon-light way to get the concept into your team's shared vocabulary.
Your action step
Take one strategic plan that currently rests on a line like "AI cannot do this reliably yet". Write down when you last tested that assumption, and put a date in the next quarter to test it again. Treat the pace of capability as a variable you check on a schedule, not a fact you set once.
If your leadership team is trying to plan around AI capability that keeps moving, that is exactly the conversation I have in AI strategy advisory engagements and on stage as an AI keynote speaker across Europe.