What? Most AI Is Now Written by AI?
This supercharges AI development by orders of magnitude
PATRICK WOOD
For the better part of the last two years, I have been tracking what appeared to be a reasonably predictable pattern in AI development: a doubling of capability roughly every 3.5 months. That figure came from METR’s time horizon benchmarks — measurements of how long an AI agent can work autonomously on a task before failing. Early 2024 through early 2026, the data held with uncomfortable consistency. If you plotted it, the curve bent upward with almost mechanical precision.
I say “appeared to be predictable” because that framing is now obsolete.
The variable that breaks every forecast model is Recursive Self-Improvement — the condition in which AI systems are no longer just tools that humans use to build AI, but active participants in building themselves. We crossed that threshold. The question of when is already behind us. The question now is what happens when a system that rewrites its own code, runs its own experiments, and optimizes its own training recipes starts doing so faster than any human team could direct it.
There is no clean answer. That is precisely the point.
Let me be specific, because vague gestures toward “exponential growth” have become their own form of intellectual laziness.
As of May 2026, Anthropic confirmed that more than 80 percent of the code merged into its own production systems was written by Claude — its own AI. Not assisted by Claude. Written by Claude. The company’s own engineers have described the shift as moving from doing work to managing a system that does the work. One Anthropic engineer publicly stated that 100 percent of his personal code output was AI-generated, with 22 pull requests shipped in a single day.
OpenAI’s announcement of GPT-5.6 in July 2026 included a data point that deserves more attention than it received: over the previous six months, the share of internal research compute devoted to AI coding inference grew one hundredfold, while internal agentic token usage — meaning AI agents working autonomously inside OpenAI’s own research infrastructure — increased twenty-two fold. OpenAI is now using its own frontier models to diagnose training failures, optimize training systems, run experiments, interpret results, tune computational kernels, and improve training recipes for the next model. The company described this as “quickly becoming standard.”
Google reported in early 2026 that 75 percent of all new code at the company is AI-generated, up from 25 percent in 2024.
These are not productivity statistics. These are evidence that the loop has closed. AI is now a primary agent in its own development cycle.
When I cited the 3.5-month capability doubling figure, I was describing what the data showed for models built primarily by human engineers using AI as an accelerant. That is a fundamentally different situation from what is unfolding now.
RSI compresses timelines in a way that no benchmark extrapolation can capture in advance. Here is the basic mechanic: if AI doubles in capability every X months under human-directed development, and AI is now directing a significant fraction of its own development, then the effective time to the next doubling shrinks in proportion to how much of the development loop the AI controls.
If AI handles 80 percent of code output and that percentage is rising, you are not looking at a linear compression of the doubling time. You are looking at a feedback loop where each generation of AI produces a more capable successor faster than the previous generation did — and that successor inherits the full research infrastructure of its predecessor.
Google DeepMind’s June 2026 paper, From AGI to ASI, described the unconstrained version of this as potentially “hyperbolic” — meaning super-exponential, a curve that in theory races toward a singularity. The authors were careful to note that real-world resource constraints bend such curves into S-shapes before they go vertical. Compute costs money. Power requires physical infrastructure. Fabricating chips takes years of supply chain work. These are genuine friction points.
But friction points are not stopping points. They slow the curve. They do not reverse it.
Here is where intellectual honesty requires admitting the limits of any analysis, including mine.
RSI does not just accelerate known processes. It creates conditions for qualitative leaps — the kind of change that looks, in retrospect, like it came from nowhere. The history of science is full of these moments. They are not random, but they are not predictable either. You cannot model the arrival of a genuinely new idea. You can create the conditions that make such ideas more likely, and an AI system running millions of research cycles per day, improving its own ability to run those cycles, is precisely such a condition.
What this means practically is that the question “when will ASI arrive?” is not answerable with a date. The consensus among serious forecasters — not YouTube thumbnails, but the AI-2027 team, the METR researchers, the Google DeepMind paper authors — clusters around AGI in the 2026-2027 range and ASI following within months to a few years. But those estimates assume that progress continues on a roughly continuous curve. RSI introduces the possibility of discontinuous jumps — moments where capability does not inch forward but lurches.
Nobody knows when those moments arrive. That is not a failure of analysis. That is the nature of the phenomenon.
The implications of RSI extend well past questions of benchmark performance or corporate strategy. They reach into every institutional structure that assumes human cognitive supremacy — which is to say, every institutional structure that exists.
Legal systems assume that humans write laws, interpret them, and enforce them. Economic systems assume human judgment at key decision points. Democratic governance assumes a human electorate making decisions about a world they can understand. All of these assumptions are being stress-tested simultaneously by a technology that is now improving itself at a rate that exceeds the capacity of any regulatory body to track, much less manage.
The same politicians and regulators who failed to anticipate the social consequences of social media algorithms — a comparatively simple technology — are now being asked to govern RSI. This should concern everyone, regardless of their position on the political spectrum. This is not a left-right question. It is a question of institutional competence in the face of something genuinely unprecedented.
Sam Altman wrote in mid-2025 that “we are past the event horizon; the takeoff has started.” That is as close to a plain statement of fact as you will get from a sitting AI lab CEO. The event horizon metaphor is apt. Past a certain point, events inside cannot be communicated outward in a way that allows course correction.
We may or may not be past that point. What is certain is that the institutions charged with maintaining human oversight of this process were not built for it, are not staffed for it, and show no signs of being reformed fast enough to matter.
Anyone who tells you they know exactly how this unfolds is selling something. The honest position is this: AI capability is growing faster than any previous technology in history, the development loop has partially closed, and the factors that could produce discontinuous breakthroughs are now structurally embedded in the research infrastructure of every major AI lab on the planet.
The 3.5-month doubling was a data point describing yesterday’s trajectory. RSI means that the trajectory is now self-modifying. The curve is not just bending upward — it is bending the conditions that determine how fast it bends.
There are serious people who believe this ends well for humanity. There are equally serious people who believe it does not. What there is not, on either side, is certainty. The responsible course is to watch what the systems are actually doing — not what the press releases say — and to resist the temptation to normalize a situation that is, by any historical standard, abnormal.
The machine is rewriting itself. Pay attention.
••••
••••
••••
••••
(UKR)