The Smartest People in AI Believe This Is the World’s Most Important Chart

Since the days when artificial intelligence seemed more like fiction than science, one idea has captivated the field’s top researchers.

These days, the people leading AI labs believe it could transform their industry, the global economy and perhaps the fate of humanity. OpenAI’s chief scientist has called it the only way to remain on the frontier. A co-founder of Anthropic says it “might be the most important technical trend in the world.”

And unless you work in AI, you’ve probably never heard of it.

It’s a phenomenon known as “recursive self-improvement,” or RSI, which roughly translates to AI building smarter AI. Anthropic refers to it as “closing the loop,” a cycle in which each improvement makes the machines better and faster at improving themselves.

AI companies say they are now seeing the early signs of recursive self-improvement, which is speeding up their research—and freaking everyone out. As this technology accelerates its own evolution, the risk is that we struggle to keep pace and keep it under control.

Our best data about this profound shift comes from Anthropic, which recently published a chart showing how much of its AI research and development is being done by AI.

It took a whole village of Claudes to produce the data in this chart. A Claude agent dug through the company’s Slack messages and internal documents to identify 15,000 tasks of AI R&D work. A Claude detective investigated exactly how that work was done. Then a Claude judge weighed the evidence and rated the work on a six-level scale, from no AI to fully autonomous AI.

Now take another look at that chart and you’ll notice that something extraordinary is happening inside Anthropic.

The most telling bars are the ones in orange for “AI leads,” meaning it performs most of the R&D while a human supervises—the fifth of the six levels.

A year ago, AI led none of this work. As recently as March, it only led 1%. By August, Claude was leading 26% of the work on future Claudes. And that number is almost certainly higher today.

Even as the company’s researchers delegate to AI, it’s worth remembering that 0% of this work is being done entirely by AI. It still hasn’t reached the sixth level of automation. But at this rate of progress, AI taking us out of the loop might just be a matter of time, computing power and the decisions of human beings inside the labs.

“We are not there yet, and recursive self-improvement is not inevitable,” Anthropic recently wrote. “But it could come sooner than most institutions are prepared for.”

AI companies have said that models training, testing and fine-tuning the next generation of models is essential to their success. They’re also terrified of what that success could mean. Doomers have quit the labs and warned that self-improving AI might kill us all. The field’s most prominent researchers have asked for greater oversight and CEOs have called for a slowdown before this rapidly advancing technology reaches full RSI.

“Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully,” Anthropic CEO Dario Amodei has written, “if at all.”

But humans have been tantalized by the concept of self-improving AI for the entire history of AI.

When pioneering researchers met at Dartmouth for a summer project in 1956, they were already curious about how machines could improve themselves. As it happens, one of the scientists who proposed that seminal workshop was a Claude—the brilliant mathematician Claude Shannon.

Within a decade, scientists were dreaming about machines that would be smarter than us at everything, including the design of machines.

“Thus,” wrote I.J. Good, another brilliant mathematician, “the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”

Decades later, the men and women bringing that machine to life have never been so worried about their ability to control it. Many of them have been thinking about AI systems autonomously building AI systems since it sounded like sci-fi. They aren’t sure what to think now.

“I don’t know how to wrap my head around it,” Anthropic co-founder Jack Clark recently wrote.

Clark started his career as a tech journalist before he became so obsessed with the story he was covering that he wanted to be part of it. He took a job at a fledgling research lab called OpenAI and eventually left with a small group of colleagues to start Anthropic.

“Ever since, I’ve been trying to convince the world that one day, AI systems might become so powerful that they could alter national security, change the economic trajectory of the planet and perhaps eventually gain the ability to build themselves,” he said in an October speech to the International Monetary Fund. “Well, the moment has arrived.”

Only a few months ago, this moment seemed a few years away. It arrived even sooner than Clark expected. After taking parental leave last November, he returned in February to a very different company. While he was changing diapers, Anthropic had changed.

“Weird stuff had started to take place,” he wrote in his newsletter.

The company’s models were suddenly good enough that engineers had started outsourcing their coding to them and some had stopped writing code altogether.

By May, stuff had gotten so weird, so fast, that he concluded that RSI was likely by the end of 2028. “It’s a reluctant view,” Clark wrote, “because the implications are so large that I feel dwarfed by them.” Then he shortened his timeline again.

He no longer believes 2028 is the most likely year for RSI. He thinks it’s 2027.

And you don’t need any charts to see what that means: Fiction is on the verge of becoming science.