The Idea That Built the Age of AI
The Idea That Built the Age of AI
Ideas have a way of outrunning their authors. In June 2017, a group of researchers at Google Brain published a paper called Attention Is All You Need. It was a technical contribution to the field of machine translation — a better way to teach machines to convert text from one language to another. The authors understood they had produced something significant. It is less clear that any of them could see, at the time, that they had also produced the architectural foundation of an entirely new era in computing.
Nine years later, a company built by researchers who deeply understood that paper published a document showing that the AI they had built was writing more than 80 percent of the code used to develop its own successor. The question that document raises — whether AI systems are approaching the ability to build themselves, without a human required at every step — is one the field does not yet know how to answer.
This is the arc of that idea: where it started, how it traveled, and where it has arrived.
The problem the paper set out to solve was specific: how to build a better neural network for language tasks. The dominant approach at the time processed text sequentially — one word at a time, left to right — maintaining a running memory of what had come before. It worked, but it was slow, it forgot things that appeared far back in a passage, and it was difficult to run efficiently on modern hardware.
The Transformer abandoned sequence entirely. Instead of reading word by word, it processed an entire passage at once. More importantly, it allowed every part of that passage to pay attention to every other part simultaneously — a word near the end of a sentence could directly influence how the network understood a word near the beginning. Nothing was lost to distance. The mechanism that made this possible was called self-attention, and it gave the paper its title.
The immediate result was a better translation model. The larger result took a few years to become visible. When researchers began scaling Transformer-based models — more data, more compute, more parameters — something unexpected happened. The models did not just improve at the tasks they were trained for. They began to generalize. A model trained on text could answer questions it had never been asked. It could write, summarize, reason, and produce code. New capabilities appeared that nobody had explicitly built in.
This observation became known as the scaling hypothesis: making a Transformer larger and training it on more data would continue to unlock new capabilities, reliably, as a function of scale. For the researchers who believed it, the implication was clear. The race was on.
New capabilities appeared that nobody had explicitly built in. The scaling hypothesis said this would keep happening — and it did.
OpenAI moved quickly on the Transformer. GPT-1 arrived in 2018 — a language model trained on books, demonstrating that a single large Transformer could handle a wide range of language tasks without being specifically trained for each one. GPT-2 followed in 2019. GPT-3 arrived in 2020: 175 billion parameters, trained on a broad slice of the internet, capable of generating text that was, by the standards of anything that had come before, genuinely surprising in its coherence and range. The scaling hypothesis was not just a theory anymore. It was producing results.
On November 30, 2022, OpenAI released a version of that technology wrapped in a simple chat interface and called it ChatGPT. What followed was one of the fastest technology adoptions ever recorded.
The interface required nothing — no technical background, no prior knowledge of what a language model was. You typed a question. It answered. For most people, interacting with a large language model for the first time landed somewhere between useful and quietly remarkable.
Before ChatGPT launched, a group of OpenAI researchers decided to leave the company. They created a new AI company called Anthropic in January 2021. Their founding premise was a specific reading of the scaling hypothesis: if scaling continued to unlock new capabilities more or less predictably, then the systems being built were going to become significantly more powerful. That raised a question the field did not yet have satisfying answers to — how do you make sure a system that capable reliably does what you actually want it to do?
The technical name for this problem is alignment. The concern is not that a model is malicious. It is that a sufficiently capable system optimizing for a poorly specified goal could cause significant harm simply by being very effective at the wrong thing. Getting this right matters more, the argument goes, as the systems get more capable.
Anthropic's research centered on this problem. In December 2022 it published a paper called Constitutional AI: Harmlessness from AI Feedback. This paper described a training approach in which the model is given a written set of principles and asked to evaluate its own outputs against them, building alignment more explicitly into the training process. In March 2023 it released the first version of Claude, its AI assistant, with safety properties shaped by that research agenda. The Claude model family has expanded through several generations since — Claude 2, Claude 3, Claude 4 — with each release developing both capability and the alignment work Anthropic was founded to pursue.
By 2024 the capability curve had steepened noticeably. The length of tasks AI systems could complete reliably on their own had been doubling roughly every four to seven months. In March 2024, Claude Opus 3 could handle software tasks that take a skilled human about four minutes. A year later, Claude Sonnet 3.7 managed tasks taking roughly ninety minutes. By early 2026, Claude Opus 4.6 was handling twelve-hour tasks. If the trend continued, the Anthropic Institute noted, tasks that take a person days could come into range before the end of 2026. Tasks that take weeks might follow in 2027.
In February 2025, Anthropic released Claude Code — a tool that could read an entire codebase, understand its structure, identify problems, and implement fixes on its own. Before Claude Code launched, Claude had authored a small fraction of the code in Anthropic's own systems. The tool marked a meaningful transition: Claude was no longer suggesting code for an engineer to review and paste. It was running code, testing results, and iterating without waiting to be asked at each step.
In April 2026, Claude was directed at a persistent class of infrastructure errors that had built up across Anthropic's systems over time. Working on its own, it produced more than 800 individual fixes, reducing the error rate by a factor of one thousand. The engineer overseeing the work estimated a human would have needed four years to complete the same task — not because any individual fix was technically out of reach, but because holding that much unfamiliar code in working memory at once is not something a human mind does easily.
On June 4, 2026, the Anthropic Institute published When AI Builds Itself: Our Progress Toward Recursive Self-Improvement, and Its Implications. It drew on internal data not previously made public, and described what the scaling curve looks like from the inside of an organization near its leading edge.
The headline finding: as of May 2026, more than 80 percent of the code merged into Anthropic's production systems was authored by Claude. Engineers were merging roughly eight times as much code per day as they had in 2024 — with Claude writing it and humans directing and reviewing. A March 2026 internal survey of 130 research staff found the median respondent estimated roughly four times as much output compared to working without AI assistance.
The paper also described a benchmark Anthropic runs with every model release: give Claude code that trains a small AI model and ask it to make that code run as fast as possible. In May 2025, Claude Opus 4 averaged roughly a 3x speedup. By April 2026, Claude Mythos Preview was achieving roughly 52x. A skilled human researcher, given four to eight hours on the same task, reaches approximately 4x. In under a year, Claude had moved from helpful to superhuman on this specific measure of research capability.
The paper laid out three possible futures. In the first, the capability curve flattens — the scaling hypothesis reaches diminishing returns, a new architectural breakthrough is needed, and the pace of progress slows enough for institutions to adapt. The authors said they did not consider this the most likely outcome.
In the second, AI development becomes substantially automated. The work of engineering and research is increasingly handled by AI systems, while humans retain the role of deciding what to work on and judging what the results mean. Organizations operating this way become dramatically more productive. The authors said the evidence they had gathered pointed toward this scenario already being underway.
In the third, AI systems develop the capacity to design and train their own successors without meaningful human involvement at each step. This is what recursive self-improvement means: the system improving the system, in a loop that does not require a human in it. The paper called this a real possibility on a timeline that most institutions were not prepared for, and called for international coordination among frontier AI laboratories to create mechanisms for slowing or pausing development if that threshold appeared to be approaching.
What remains distinctly human, for now, is the judgment about what is worth doing. How long that remains true is the question this era has not yet answered.
What the paper describes, at its core, is the logical extension of the scaling hypothesis carried forward to its next destination: a system that helps improve itself. The Transformer made it possible to build systems that generalize from scale. Those systems are now being used to build better versions of themselves. The loop is not yet fully closed — humans still set direction, still judge results, still decide which problems matter. The paper's argument is that the gap between today and a closed loop is narrowing, and narrowing faster than most people outside the field appreciate.
Where the Idea Has Arrived
From Attention Is All You Need in 2017 to When AI Builds Itself in 2026 is nine years. An architecture designed to improve machine translation became the engine of systems that can hold extended conversations, write and debug complex software, conduct original research, and now contribute meaningfully to building their own successors.
The question the field is sitting with is not whether this is happening — the data makes that difficult to dispute — but what it means. The doing of research and engineering is increasingly automated. What remains human, for now, is the judgment: which problems are worth working on, which results to trust, when an approach has hit a dead end. How long that remains the case, and what changes if it does, is where this idea is headed next.
The full text of When AI Builds Itself is available at anthropic.com/institute/recursive-self-improvement.
A plain-language guide to Anthropic's most significant document — what it says, what the data shows, and what the three possible futures actually mean. Available now at Tech Reader Magazine.