A new study found that AI is now helping to build faster AI — nearly twice as fast as before. Meanwhile, the company behind those models warns it may be close to machines that improve themselves. This is the moment the future arrived early.
There is an old saying in science: the best way to make a discovery is to build a machine that makes discoveries. For decades, that meant better microscopes, faster telescopes, or more sensitive detectors. But in 2026, the machine that makes the discovery is beginning to make itself — and that changes everything.
This month, an independent research group called METR published a study with a striking finding: the artificial intelligence models being developed by the company Anthropic are improving nearly twice as fast when AI is used to help build them. In other words, AI is now accelerating its own development. Around the same time, Anthropic itself warned that it may be very close to what researchers call "RSI" — the ability for a machine to improve its own intelligence, also known as self-improving AI or "self-building AI."
Together, these two developments are not just news headlines. They are, for many experts, the first clear signs that we have entered a new era in the history of intelligence — one where the tool and the maker become the same thing. Let's break down exactly what happened, why it matters, and what it could mean for the future.
Why this matters: For the first time, artificial intelligence is not just a tool we use — it is becoming a participant in its own creation. That shifts the pace of technological change in ways we are only beginning to understand.
The story began with a study by METR, an independent AI safety research organization. According to coverage by Reuters, ABC News, Dealroom, and Nautilus, METR found that Anthropic's models were improving "nearly twice as fast" with help from AI. The implication is profound: when AI tools are used to write code, test systems, and design new models, the development process speeds up dramatically.
The findings were reported by multiple major outlets within days. Reuters led with the headline that Anthropic's models were "improving nearly twice as fast with help from AI," while ABC News framed it as Anthropic "rapidly accelerating AI development amid slowdown calls." Dealroom and other industry analysts noted that AI is "now building AI" — a phrase that has become the defining summary of the moment.
But the more provocative development came from Anthropic itself. As reported by TechCrunch in early November, the company warned that it is "very close to RSI capability, aka self-building AI." This is not a claim that Anthropic has already built a self-improving model. It is a public acknowledgment that the company believes such a capability may be imminent — and a caution that the road ahead carries real risks.
The combination of these two stories — the METR study and Anthropic's warning — has ignited a fierce debate about the pace and safety of AI development. Some see it as proof that we are on the verge of an explosion of progress. Others see it as a warning bell that we may be moving faster than we can safely control.
To understand why this is significant, we need to unpack what "AI helping to build AI" really involves. It sounds abstract, but it is already happening in concrete ways.
Modern AI models are enormously complex. They are made of millions of lines of code, trained on vast amounts of data, and fine-tuned through countless experiments. Building and improving them requires enormous effort from teams of engineers and researchers. Now, AI tools can help with much of that work — writing code, debugging software, testing new ideas, and even proposing architectural changes to the model itself.
The METR study measured the effect of this help. When AI is used in the development process, models improve roughly twice as fast as they did without it. That may sound like a modest improvement, but in the world of technology, doubling the pace is enormous. It means discoveries that used to take a year now take six months. It means problems that once seemed intractable may be solved in weeks.
The acceleration loop: Imagine a scientist using an AI assistant to write code. That assistant gets better at writing code. Better code makes a better AI assistant. That better assistant writes even better code. The loop spins faster and faster. That is the promise — and the risk — of AI building AI.
The implications of this acceleration reach far beyond laboratories. Let's look at what it means for different people.
Imagine you are a scientist trying to solve a difficult problem — perhaps a new material for batteries, or a way to treat a disease. For years, this work might take a team a decade. Now, with AI helping to run experiments, analyze data, and propose hypotheses, that same work might be completed in a fraction of the time. This is already happening. Earlier this year, an unreleased Anthropic model reportedly helped settle an 18-year-old quantum computing conjecture in just three days — work that had stumped researchers for nearly two decades.
Imagine you work at a technology company racing to release a new product. With AI doing much of the coding, testing, and design work, you can move twice as fast as your competitors. This creates intense pressure to adopt AI — not because it is better, but because the alternative is being left behind. That pressure is already reshaping entire industries, from software to pharmaceuticals to finance.
Imagine you are a programmer, a designer, or a writer whose job involves creating something new. If AI can now do much of that work faster and cheaper, your role is changing. This does not necessarily mean your job is disappearing — but it does mean the nature of the work is shifting. The most valuable skill may not be doing the task, but knowing how to direct, verify, and improve the AI that does it.
Imagine the medicines you take, the energy you use, and the products you buy. All of these will be shaped by how fast AI can help us invent them. If AI truly accelerates innovation, we could see breakthroughs in health, climate, and technology at a pace unlike any in human history. Or we could move so fast that we lose control. The difference between those two futures may come down to how wisely we manage this new power.
The "AI building AI" trend touches many different sectors, and the implications differ for each.
The "AI building AI" trend did not appear overnight. It is the culmination of decades of progress in machine learning. The breakthrough that made it possible came in 2017, with the invention of the "Transformer" architecture — the same technology behind the AI models we use today. Transformers allowed AI to understand language and patterns in ways that were previously impossible.
Since then, models have grown enormously more capable. Today, they can write code, create images, and even help solve scientific problems. The latest development — AI helping to build better AI — represents a new phase in this evolution. As Anthropic's warning about RSI suggests, we may be approaching a point where AI can improve its own intelligence, a milestone that could change the trajectory of human civilization.
The key question: Will AI's self-improvement be a tool we carefully guide, or a force we can no longer control? The answer depends largely on the safety research that is happening right now.
Like any powerful technology, AI that builds AI carries both enormous benefits and serious risks. A balanced view requires acknowledging both.
The situation is still developing rapidly. METR's study has added rigor to a debate that was previously dominated by speculation. Anthropic's willingness to publicly discuss RSI is unusual — most companies have been more cautious. But the trajectory is clear: AI development is accelerating, and the question of how to manage that acceleration is one of the most important challenges of our time.
Experts are divided on what to expect. Some believe the acceleration will lead to an "intelligence explosion" — a rapid, self-sustaining improvement in AI that could solve problems we cannot currently imagine. Others worry that we are moving faster than we can safely control, and that the risks outweigh the rewards.
What is certain is that this is not a distant future. It is happening now, in laboratories and companies around the world, and it will shape the lives of everyone on the planet.
The discovery that AI can build AI twice as fast is a turning point in human history. It marks the moment when the maker and the tool became the same thing — when intelligence began to improve itself. The potential is staggering: cures for diseases, solutions to climate change, and breakthroughs we cannot yet imagine. But so is the risk: a loss of control that could have consequences we cannot foresee.
As Anthropic's warning reminds us, we are close to uncharted territory. The choices we make now — about safety, governance, and how fast to go — will determine whether AI becomes humanity's greatest tool or its greatest danger. The future is no longer something that happens to us. It is something we are building, one AI at a time.
What do you think? Is accelerating AI development an opportunity we should seize, or a risk we should slow down? Share your thoughts below — this is a conversation that concerns everyone.
Join the Discussion