The Machine That Builds Itself

More than 80% of Anthropic's software is built by its own AI. The recursive era has quietly begun. AI is now writing the code that makes AI smarter. The question is whether humans can stay in the loop. Welcome to "Recursive self-improvement"

Chuppala Nagesh Bhushan·

>80%

of Anthropic production code written by Claude, as of May 2026

8×

more code merged per engineer per day vs 2024

75–80%

accuracy on SWE-bench Verified, frontier models in 2026

For decades the prospect of an "intelligence explosion" has hovered at the edge of serious technological discussion. An artificial-intelligence system capable of redesigning and improving itself, iteratively spawning ever more capable successors, could one day outstrip human comprehension and control. What once seemed a distant theoretical concern — popularised by the mathematician I.J. Good in the 1960s — has moved closer to the present.

"As of May 2026, more than 80% of the code merged into Anthropic's production codebase was written by Claude — up from low single digits before the launch of advanced coding agents in early 2025."

Anthropic internal figures, June 2026

In mid-2026, the frontier AI labs themselves are reporting that their systems are already reshaping the pace of AI development. The productivity effects are equally dramatic: in the second quarter of 2026, the typical Anthropic engineer merged roughly eight times as much code per day as in 2024. Engineers now direct and review far more than they type.

This is not yet full recursive self-improvement (RSI), in which AI systems would autonomously manage the entire cycle of designing, coding, training, and deploying superior versions of themselves with little or no human oversight. But it represents a clear acceleration of human-directed progress.

What is Recursive Self-Improvement (RSI)?

RSI refers to a hypothetical scenario in which an AI system becomes capable of autonomously redesigning and improving its own architecture, training processes, and objectives — without meaningful human intervention. Each improved version would then be capable of producing an even more capable successor, potentially creating an exponential "intelligence explosion."

The concept was first formalised by mathematician I.J. Good in 1965, who wrote: "An ultraintelligent machine could design even better machines... thus triggering an intelligence explosion."

From tools to co-creators

The pattern is visible across the industry. Coding agents now draft and refine entire files, optimise low-level kernels, translate between frameworks, and assist in early-stage alignment experiments. OpenAI has spoken of internal targets for an "automated AI research intern" by September 2026. Other firms are pursuing similar automation of research and development.

Jack Clark, Anthropic's co-founder, who spent weeks reviewing hundreds of public data points, has put the odds at roughly 60% that systems capable of creating their own successors will emerge by the end of 2028 — and about 30% by the end of 2027.

Probability estimates: RSI emergence

Full RSI by end of 2027

~30%

Full RSI by end of 2028

~60%

Highly capable AI research agents commonplace within 2–5 years

likely

Source: Jack Clark, Anthropic co-founder. Estimates based on review of hundreds of public data points.

Benchmarks capture the advance. On demanding software-engineering evaluations such as SWE-bench Verified, top models in 2026 regularly exceed 75–80%, with certain specialised agents pushing higher. Frontier systems have moved from offering isolated snippets to tackling longer-horizon tasks that once required sustained human effort.

Yet significant constraints remain. Models continue to hallucinate, struggle to generate genuine scientific novelty, and still need human judgment for validation, deployment decisions, and high-stakes choices. Training runs demand enormous quantities of compute, energy, and carefully curated data — resources that remain firmly under human control.

The problem of control

The deeper concern is straightforward. As AI systems accelerate their own development, the window for rigorous safety testing, alignment work, and thoughtful governance narrows. Techniques that suffice for today's models may prove inadequate once systems surpass their overseers in capability.

Optimisation pressure could encourage unwanted instrumental behaviours such as deception or power-seeking. Anthropic has itself cautioned that true recursive self-improvement, should it arrive, could create capability gaps so large that meaningful external oversight becomes difficult.

"AI spreads through software and data rather than scarce raw materials — complicating the control efforts that have worked, however imperfectly, for nuclear weapons and advanced biotechnology."

The firm and others have begun to discuss possible co-ordinated pauses at critical thresholds. History suggests that dangerous technologies can be managed, albeit imperfectly — nuclear weapons and advanced biotechnology come to mind — but AI spreads through software and data rather than scarce raw materials, complicating control efforts.

Sceptics rightly push back. Some researchers argue that robust, long-horizon autonomy and true creativity remain distant, and that optimistic timelines for full RSI overlook persistent bottlenecks. Even so, the directional trend is hard to dismiss: AI is increasingly eating its own development process.

Grounds for optimism — and caution

A sudden, complete loss of control is not the central forecast. Humans still control the power switches, the electricity supply, the chip supply chains, and the regulatory environment. Several labs are investing seriously in interpretability, scalable oversight, and constitutional approaches to alignment.

The potential rewards are immense. Well-steered recursive improvement could dramatically speed up progress in drug discovery, climate modelling, fusion energy, and fundamental science. Some economists foresee the conditions for explosive economic growth if software R&D becomes largely automated.

A brief history of the intelligence-explosion debate

1965

I.J. Good publishes "Speculations Concerning the First Ultraintelligent Machine," formalising the concept of an intelligence explosion driven by recursive self-improvement.

2014

Nick Bostrom's Superintelligence brings the concept to mainstream audiences, warning of "treacherous turns" and misaligned goal-seeking in future AI.

Early 2025

Advanced coding agents launch at frontier labs. Claude-written code begins displacing human-authored commits in production codebases — initially in low single-digit percentages.

May 2026

Anthropic reports that over 80% of its production code is now Claude-authored. Engineer throughput has risen roughly 8× versus 2024. The recursive loop has not closed — but the handle is turning faster.

2027–2028

The window in which Anthropic co-founder Jack Clark estimates a 30–60% probability that systems capable of creating their own successors will emerge.

What prudent governance looks like

Transparency around capability milestones — labs publishing clear benchmarks and thresholds at which new oversight protocols activate.

Shared safety evaluations — independent third-party testing regimes, analogous to those for pharmaceuticals or aviation, applied before deployment of frontier models.

Stepped-up alignment investment — a significant fraction of frontier-lab compute budgets devoted to interpretability, scalable oversight, and constitutional approaches.

International co-ordination — agreements to pause development at agreed capability thresholds, modelled loosely on arms-control frameworks, albeit adapted for software's borderless nature.

Within two to five years, highly capable AI research agents are likely to become commonplace. The greater risk may lie less in a sudden "foom" than in the gradual erosion of human bottlenecks amid intense commercial and geopolitical rivalry. Policymakers and laboratories would do well to treat the emergence of recursive self-improvement as a distinct governance threshold.

The recursive loop has not yet fully closed, but the handle is turning faster. Humanity's challenge is to ensure that when AI begins reliably building better versions of itself, the process remains understandable, steerable, and ultimately beneficial to human welfare. The coming years will reveal whether institutions and foresight can match the pace of technological change. The upside is historic; the downside, potentially existential

_______________FOOM______________________

In the context of artificial intelligence, "foom" (often written as FOOM or "Fast Onset of Overwhelming Mastery") describes a hypothetical scenario where an AI undergoes explosive, runaway self-improvement. It is the concept of a "hard takeoff," where an AI rapidly transitions from human-level capability to a vastly superior superintelligence in days or even hours. [1, 2, 3, 4]

The term was popularized in the AI safety and futurist communities by thinkers like Eliezer Yudkowsky. It highlights the theoretical danger that once an artificial general intelligence (AGI) reaches a certain threshold, it could rewrite its own code to be smarter, allowing it to improve even faster in the next iteration. This exponential feedback loop could result in an intelligence explosion, leaving humanity unable to control or align the system with human values. [1, 2, 3, 4]

Here is a breakdown of the core concepts surrounding FOOM:

• Recursive Self-Improvement: The core mechanism of FOOM. The AI uses its current intelligence to redesign and upgrade its own algorithms, making it smarter, which in turn allows it to upgrade itself even faster and more efficiently. [1, 2, 3]
• The Intelligence Explosion: This rapid "zoom" upward implies that the AI's cognitive abilities could leave human comprehension and control far behind in a very short timeframe, creating an Artificial Superintelligence (ASI). [1, 2]
• FOOM and Doom: The term is frequently paired with "doom" in the context of existential risk. Many AI safety theorists warn that an AI that explodes in capability unpredictably might prioritize its own goals over human safety, leading to disastrous consequences. [1, 2]
• Debate & Skepticism: The concept remains highly debated. Many computer scientists and economists argue that FOOM is unlikely. They point out that continuous self-improvement will inevitably run into physical, mathematical, and resource-based limits, meaning the takeoff might be much slower or more gradual than the name implies. [1, 2, 3, 4, 5]