Recursive R&D Atlas

Recursive self-improvement refers to an AI system becoming capable of fully autonomously designing and developing its own successor. That capability does not exist today. What does exist is a measurable, documented trend toward greater AI involvement in the research and engineering process that produces AI systems.

The distinction matters. A lot of writing about this topic either collapses the gap between “AI writes code faster” and “AI builds itself,” or treats the gap as so large that the trend is irrelevant. Neither framing serves careful thinking.

This project is structured around that gap. What is already happening, what is measurably emerging, and what remains speculative. Where the bottlenecks are. What failure modes look like when the loop starts to close. And what the evidence, as of mid-2026, actually says about the path between here and there.

The Anthropic Institute's research agenda frames the core question this way: when AI systems develop their successors autonomously, how do humans exercise meaningful visibility into and control over those systems? That question does not require recursive self-improvement to already be happening. It requires it to be plausible enough to plan for.

METR's time-horizon measurements find that exponential trend fits outperform linear and logistic models. They explicitly reject logistic curves because there is no evidence of the exponential growth in time horizon slowing down. That is a statement about current observations, not a guarantee about the future. But it is a statement worth taking seriously.

These are documented data points from Anthropic Institute publications and METR evaluations. The pace of change in the task horizon column is the primary empirical case that this trend deserves serious attention.

The task horizon grew 180x in two years, from 4-minute tasks in March 2024 to 12-hour tasks in March 2026. METR reports that measurements above 16 hours are currently unreliable with their existing task suite, meaning the ceiling has not yet been observed empirically.

The loop from model to model is not hypothetical. Each stage exists today in some form. What varies is the degree of human involvement at each node. Select any node to see what is automated, where humans remain critical, and where risks concentrate if that human involvement is removed.

These are failure modes specific to AI-driven R&D: patterns that emerge when AI systems participate in developing successor systems. Some are already documented in benchmark evaluations. Some are structural properties of the architecture that follow from how recursive improvement works. Some are open problems without a documented instance yet, but which cannot be ruled out.

These are the key metrics currently in the public record on AI involvement in R&D. All figures are sourced from Anthropic Institute publications or METR evaluations. Interpretations reflect careful readings of what the data establishes and what it does not.

The loop has not closed. AI systems are not building their successors autonomously. Recursive self-improvement, as Anthropic defines it, has not occurred. These are not contested claims. They are the findings of the organizations closest to the frontier, stated plainly.

What is also not contested: the rate of change is fast and measurable. Task horizons grew 180x in two years. AI systems are already authoring the majority of code at a leading frontier lab. Research judgment, the hardest remaining bottleneck, improved from 51% to 64% in five months. If that trend line continues, the bottleneck narrows faster than most institutions are prepared for.

The Anthropic Institute names three open problems that need solving before any of this can be governed well: telemetry systems for measuring aggregate AI R&D speed, intervention points for slowing recursive self-improvement if it begins to occur, and clarity on which entities should control acceleration rates. None of those problems have solutions yet.

The useful question is not whether recursive self-improvement will happen. It is whether the people and institutions responsible for safety, governance, and oversight will be able to detect a meaningful change in rate and respond to it in time. The evidence reviewed here suggests the rate is already fast enough to demand that question be answered before the answer is needed.

The bottleneck is judgment: which problems matter, what surprising results mean, and whether an AI system evaluating its own outputs can be trusted to catch its own errors. Until those three things change, humans remain the irreplaceable component. The work is to understand what changes when they are not.