Recursive Self-Improvement: How Artificial Intelligence Begins to Improve Itself

Contents

• What is Recursive Self-Improvement (RSI)
• Why RSI is considered the basis of the singularity
• How modern LLMs help create new models
• Anthropic Case (April 2026): Agents Closed 97% of AI Safety Research Task
• Why Anthropic states that RSI is "not yet inevitable"
• Where is the line between automation and self-improvement
• Technical limitations of modern models
• Is a fully autonomous AI development cycle possible
• Frequently Asked Questions

What is Recursive Self-Improvement (RSI)

I would explain this term as follows: imagine an engineer who not only writes code but creates the next, better engineer than himself – and that engineer, in turn, creates an even better one. Recursive Self-Improvement is precisely such a cycle, only instead of a human engineer, artificial intelligence acts in it. An AI system participates in creating its own successor: it writes code, optimizes architecture, prepares training data, or formulates research hypotheses – and the result turns out to be smarter and more productive than the original.

The key difference between RSI and ordinary automation is recursion itself. If an AI simply speeds up a specific task (e.g., writes code according to a template), it's useful automation, but not self-improvement. RSI begins where the result of one cycle becomes the input for the next, and each iteration is potentially faster and higher quality than the previous one – and therefore, theoretically, the pace of progress can increase non-linearly, not linearly.

Why RSI is considered the basis of the singularity

If we return to the classic definition of technological singularity – a point after which AI development becomes so rapid that humans lose the ability to predict it – RSI is precisely the mechanism that theoretically creates such a point. For more details on the origin of the singularity concept itself, the history of the term, and researchers' predictions, read our full guide: Technological Singularity: A Complete Guide in Simple Terms.

The logic is simple: if the self-improvement cycle is truly closed and each iteration takes less and less time, the pace of progress ceases to be linear. This is why in the vast majority of scientific and popular science works on the singularity, RSI is featured not as one of the possible scenarios, but as the central technical mechanism around which all argumentation is built – including Vernor Vinge's original 1993 essay.

To understand why a cycle, rather than a one-time leap in quality, is at the core of the concept, it's worth breaking down the mechanics step by step. In a normal, "non-recursive" AI development scenario, each subsequent generation of models is created by humans: researchers pose hypotheses, engineers write code, and the speed of progress is limited by the number of human-hours a team can invest in development. This is a linear process – doubling the research team approximately doubles the progress speed, no more.

RSI changes precisely this proportion. If model N participates in creating model N+1, and model N+1 is smarter, faster, or cheaper to use, it can, in turn, take on an even larger share of the work on model N+2. Each iteration not only repeats the previous one but potentially reduces the time and resources needed for the next. It is this property – when the speed of the acceleration process itself also accelerates – that distinguishes a recursive cycle from simple progress accumulation, and it is this property that mathematically corresponds to the "vertical takeoff" of the curve, which Vernor Vinge described as the limit of predictability, and Ray Kurzweil – as a consequence of the law of accelerating returns.

It is important to emphasize again: the mere existence of such a cycle in principle does not mean that it has already been launched at full capacity or that it cannot be slowed down. How far modern models have progressed along this path and where exactly the line between "the model helps" and "the model independently leads the process" lies – we will discuss in detail below, as well as in the sections on the Anthropic case and technical limitations.

Why RSI is considered the basis of the singularity

If we return to the classic definition of technological singularity – a point after which AI development becomes so rapid that humans lose the ability to predict it – RSI is precisely the mechanism that theoretically creates such a point. For more details on the origin of the singularity concept itself, the history of the term, and researchers' predictions, read our full guide: Technological Singularity: A Complete Guide in Simple Terms.

The logic is simple: if the self-improvement cycle is truly closed and each iteration takes less and less time, the pace of progress ceases to be linear. This is why in the vast majority of scientific and popular science works on the singularity, RSI is featured not as one of the possible scenarios, but as the central technical mechanism around which all argumentation is built – including Vernor Vinge's original 1993 essay.

To understand why a cycle, rather than a one-time leap in quality, is at the core of the concept, it's worth breaking down the mechanics step by step. In a normal, "non-recursive" AI development scenario, each subsequent generation of models is created by humans: researchers pose hypotheses, engineers write code, and the speed of progress is limited by the number of human-hours a team can invest in development. This is a linear process – doubling the research team approximately doubles the progress speed, no more.

RSI changes precisely this proportion. If model N participates in creating model N+1, and model N+1 is smarter, faster, or cheaper to use, it can, in turn, take on an even larger share of the work on model N+2. Each iteration not only repeats the previous one but potentially reduces the time and resources needed for the next. It is this property – when the speed of the acceleration process itself also accelerates – that distinguishes a recursive cycle from simple progress accumulation, and it is this property that mathematically corresponds to the "vertical takeoff" of the curve, which Vernor Vinge described as the limit of predictability, and Ray Kurzweil – as a consequence of the law of accelerating returns.

It is important to emphasize again: the mere existence of such a cycle in principle does not mean that it has already been launched at full capacity or that it cannot be slowed down. How far modern models have progressed along this path and where exactly the line between "the model helps" and "the model independently leads the process" lies – we will discuss in detail below, as well as in the sections on the Anthropic case and technical limitations.

How modern LLMs help create new models

Here, I want to immediately manage expectations. Modern large language models are already actively involved in the development of next-generation AI – but primarily as a tool, not as an autonomous architect. In practice, it looks like this: models write and test code for training pipelines, generate synthetic data for fine-tuning, help analyze experiment results, and even propose hypotheses for further research.

Both industry leaders confirm this trend officially, albeit in different words. Anthropic reports that the share of code written by Claude and merged into the company's production codebase has grown from single digits in early 2025 to over 80% in May 2026 (Anthropic, "When AI builds itself"). OpenAI, meanwhile, has publicly outlined a more ambitious goal: CEO Sam Altman, Chief Scientist Jakub Pachocki, and co-founder Wojciech Zaremba have stated their intention to create an "AI research intern" by September 2026, and a fully autonomous AI researcher by March 2028 (MIT Technology Review).

However, researchers themselves admit that empirical data on how much AI-assisted automation actually accelerates AI development is still contradictory. In a pilot study for the International AI Safety Report 2026, expert forecasters estimated the probability that the progress of the next few years will compress six years of development (2018–2024) into just two years at only 20% median, and superforecasters even more conservatively, at 8% (International AI Safety Report 2026).

⚙️ How Exactly Does Recursion Happen: Three Specific Mechanisms

In the previous sections, we deliberately described RSI at a high level – "the model writes code," "the model optimizes architecture." But to understand how realistic this is, rather than metaphorical, it's worth examining the specific technical methods behind these phrases. There are three of them, and each already has documented, measurable results – not in laboratory promises, but in the production infrastructure of leading labs.

1. Searching for Architectures and Hyperparameters Through Reinforcement Learning (NAS + RL)

Neural Architecture Search (NAS) is an approach where, instead of an engineer manually selecting the structure of a neural network (number of layers, type of connections, attention window size), this search is entrusted to a separate RL system. This system treats each architectural variant as an "action" and receives a "reward" based on how well a model trained on it performs. This isn't a new idea – Google Brain's first works on NAS via RL appeared as early as 2016-2017 – but the scale of application is qualitatively different now.

The most illustrative modern example is AlphaEvolve, a Google DeepMind system unveiled in May 2025. It combines large language models (based on Gemini) with automated evolutionary evaluation: the LLM proposes code or algorithm variants, an automated evaluator checks the result against a given metric, and the best variants "survive" and mutate further – a classic evolutionary scheme, only instead of random mutation, the language model itself generates the variants. AlphaEvolve has already been used to optimize chip design, data center scheduling, and – notably for the topic of this article – to optimize the Gemini training process itself, reducing training time by approximately 1%. The figure sounds modest, but at the scale of a training cluster costing hundreds of millions of dollars, this is a significant saving (MindStudio).

2. Evolutionary Self-Modification of Agent Code

The second mechanism is not optimizing the model's architecture, but self-modifying the code of the agent that uses that model. The most striking example is the Darwin Gödel Machine (DGM), developed by Sakana AI in collaboration with the University of British Columbia. This is a system where an agent, based on a foundational LLM, proposes changes to its own code, tests them on real tasks, and if a change improves the result, it becomes the new "version" of the agent, which then proposes further changes itself. An important technical clarification: DGM modifies the agent's code and logic (how it plans steps, handles errors, interacts with tools), not the foundational language model itself – meaning it's self-improvement of the "wrapper," not the "brain." On the SWE-bench benchmark (real-world bug fixing tasks in open-source code), such self-modification yielded improvements of 20-50% compared to the agent's initial version (AI Wiki).

This category also includes the generation of hard negative examples – when the model itself creates the most difficult, edge-case training examples for itself (situations where the previous version most often made mistakes), and trains specifically on them, rather than on a random data sample. This accelerates learning on the model's specific weak points, but – and this is a direct bridge to the next section – this technique only works when there is a reliable way to verify that the generated "difficult example" is actually correct, and not just a plausible error the model generated for itself.

3. Automating Its Own Development Cycle

The third, most straightforward mechanism is when the model simply participates in creating the next version of itself through ordinary engineering work, only on an increasingly larger scale. In February 2026, OpenAI reported that the GPT-5.3-Codex model played a significant role in its own creation: it helped debug the training process, manage deployment, and analyze the evaluation results of the next version (IEEE Spectrum). This aligns with Anthropic's data on the proportion of Claude's code in the company's own codebase, which we cited earlier.

A common feature of all three mechanisms – and simultaneously their most important practical limitation – is that each still operates within a clearly defined, measurable evaluation system set by a human engineer (AlphaEvolve's metric, DGM's SWE-bench test set, GPT-5.3-Codex's success criteria). This is precisely the boundary of "who defines the task" that was discussed in the previous section.

🌀 Mathematical Dead End: The Problem of "Model Collapse"

One of the most important counterarguments against optimistic RSI predictions rarely makes it into popular retellings – and it shouldn't, because it's not rhetorical, but mathematical. If an AI system increasingly learns from data generated by previous AI models themselves (rather than from "raw" human data), there's a risk of so-called model collapse – a progressive degradation of the model's quality and diversity with each new cycle.

The phenomenon was formally described by Shumaylov et al. in a paper published in Nature in July 2024. The mechanism is simple and relentless: each generation of a model, trained primarily on the output of the previous generation, gradually loses the "tails" of the original data distribution – rare, atypical, but perfectly valid answers and examples. With each cycle, the model's output distribution narrows, increasingly converging towards a "statistical average," шаблонный response, leading to an almost degenerate, nearly identical output for different queries (Digital Applied). The researchers identify three components of this error: statistical approximation error (limited training examples at each step), model expressivity error (the architecture physically cannot reproduce the entire original distribution), and functional learning error.

This is not purely a theoretical risk. Subsequent research has identified specific threshold values. Ziedek et al. demonstrated that even mixing in 5% of real, human data into each training generation significantly slows down and curbs long-term collapse, whereas training on 100% synthetic data leads to linear, relentless deterioration with each generation. Suresh et al. in 2025 formalized this dependency mathematically: if the proportion of real data in each cycle is greater than zero, the divergence from the true distribution remains bounded; if it tends towards zero, the divergence becomes unbounded (arXiv, Suresh et al.).

Here lies the direct connection to the very idea of RSI: a closed loop where model N trains model N+1 exclusively on its own generated data without an influx of new human data or external verification is a textbook recipe for model collapse, not accelerated progress. It's telling that researchers Feng et al. in 2025 found an important nuance that explains why leading labs still avoid this trap in practice: sufficiently effective verification of synthetic data can significantly mitigate collapse, even if the proportion of "raw" human data is small (arXiv). This is precisely why the mechanisms from the previous section – AlphaEvolve, DGM, hard negative example generation – all rely not on "blind" self-learning, but on an external, objective verification criterion (compiled code that is actually faster; a test that actually passes; a task that is actually solved correctly) – this is the very verification that theoretically protects them from collapse.

However, balance is also important here: not all researchers consider the scale of the collapse threat equally catastrophic. A 2026 position paper with the telling title "Model Collapse Does Not Mean What You Think" warns that some public concern is based on a simplified reading of Shumaylov's initial research, and that in more realistic scenarios – where synthetic data doesn't replace but accumulates alongside real data, and where there is some quality verification – collapse manifests much more slowly and mildly than the first alarmist headlines suggested (arXiv, "Position: Model Collapse Does Not Mean What You Think").

The practical takeaway for this article: model collapse is not an argument that "RSI is impossible," but a specific, mathematically grounded explanatory mechanism for why every documented example of RSI we've examined – both in this article and in the Anthropic case – relies precisely on verified, objectively verifiable success criteria, rather than pure self-learning by the model on its own responses. This is another reason why "who defines and verifies the success criterion" is, in essence, no less important a question than "who defines the task itself," which we discussed earlier.

Anthropic Case Study (April 2026): Agents Closed 97% of an AI Safety Research Task

The most concrete and well-documented example of how this works in practice is Anthropic's April experiment, described in the report "When AI builds itself" dated June 4, 2026. Claude-based agents were given a real open-ended AI safety research task: could a weaker model reliably control (supervise) a stronger one. The task had a clearly measurable "floor" (how well the weaker supervisor performed on its own) and "ceiling" (how well the stronger model performed when trained on correct answers).

Nine parallel agents independently formulated hypotheses, conducted experiments, exchanged results via a shared forum, and iterated on solutions. Over 800 cumulative hours of work and approximately $18,000 in compute costs, the agents recovered 97% of the gap between the "floor" and the "ceiling." For comparison, two human researchers working on the same task for about a week recovered only 23% (Anthropic).

Separately, another notable result from the same report is worth mentioning: the previous Claude Mythos Preview model demonstrated a 52-fold acceleration on the task of optimizing model training code – whereas a skilled human researcher would need four to eight hours for a four-fold improvement on the same code. For comparison, the previous version of the model – Claude Opus 4 – showed only a threefold acceleration on a similar task a year earlier (Anthropic).

Why Anthropic States RSI is "Not Yet Inevitable"

I have deliberately placed this nuance in a separate section because it's where Anthropic's message is most often distorted in retellings. The company itself, despite the impressive figures, explicitly and repeatedly emphasizes caveats. Firstly, the result of the April experiment was not directly transferable to industrial-scale models – the task was intentionally limited and isolated. Secondly, and most importantly: humans in this experiment still chose the research task itself and wrote the rubric for evaluating results – meaning the direction of research, not just its execution, remained a human decision (Anthropic).

It's telling that Anthropic used this same data not for a triumphant announcement, but as an argument for global AI safety coordination, publicly calling on the industry to consider a controlled pause in the development of the most powerful systems if several leading labs simultaneously approach the threshold of self-improvement (Dallas Express). This is an unusual communication strategy for the industry: the company publishes data that could be interpreted as proof of its technological advantage, but presents it primarily as a warning signal, not as marketing.

Where is the Line Between Automation and Self-Improvement?

In my opinion, this is the most important practical question in the entire topic of RSI – and simultaneously the most often ignored in popular retellings. The line is not drawn by how well a model performs a task, but by who decides which task is worth solving at all.

Modern models – including the agents from Anthropic's April experiment – have already confidently surpassed the "template executor" level: they independently design experiments, write and debug complex code, and analyze conflicting results. However, the choice of research direction, prioritization of problems, and definition of success criteria – all still remain human prerogatives in the vast majority of documented cases. The authors of the Anthropic report explicitly separate these two categories: engineering automation, where progress is already very significant, and research decision-making, where the gap between humans and models remains substantial (SmartCR, review of Anthropic report).

There is also intermediate data that shows how quickly this line is moving. In a separate internal Anthropic experiment, they tested whether a model could independently choose the best "next step" at complex junctures of real research sessions – and in November 2025, Claude agreed with a human researcher's decision in 51% of cases (TheNextWeb). This is not an autonomous choice of research direction in its purest form, but a clear signal that the line between "execution" and "decision-making" is no longer absolute.

Technical Limitations of Modern Models

It's worth listing the specific barriers that, in my opinion, are most often underestimated in optimistic forecasts:

• Transferability of results to a larger scale. Success on a limited, specially prepared task (as in Anthropic's April case) does not guarantee the same result on industrial-scale models with a much more complex and less structured problem definition.
• Goal selection and prioritization. The most persistent and documented gap is precisely in the model's ability to independently decide which problem is worth solving first and why, not just how to solve it.
• Cost and computation time. Even the impressive 97% recovered gap cost nearly $18,000 and 800 hours of agent work for one relatively narrow task – this is still far from an instantaneous or cheap process.
• Evaluation of result quality. In the mentioned experiment, the success evaluation rubric was still written by a human. Autonomous, reliable self-assessment of the quality of one's own scientific conclusions by the model is a separate unsolved problem.

Is a Fully Autonomous AI Development Cycle Possible?

The honest answer as of mid-2026: individual elements of such a cycle have been demonstrated, but not the closed loop itself. Models already write a significant portion of the code that trains subsequent models, independently design and conduct experiments within clearly defined boundaries, and even make decisions at the level of individual research junctures in some cases. What has not yet been documented by any of the leading labs is a model that, from start to finish, independently chooses the research direction, formulates success criteria, conducts the work, and in doing so produces a next-generation model significantly better than itself, without a defining human contribution at each of these stages.

Both OpenAI and Anthropic, despite their different public rhetoric, essentially acknowledge the same fact: the key missing link is not the technical complexity of automating individual steps, but the autonomous determination of which steps are worth taking at all. This is precisely why OpenAI is cautiously formulating its goal in stages – first an "intern researcher" by September 2026, and only then, in an optimistic scenario, a fully autonomous researcher by 2028 (Marketing AI Institute), rather than claiming a ready-made closed loop already.

Frequently Asked Questions

What is Recursive Self-Improvement in simple terms?

This is a mechanism by which an AI system participates in creating its own successor – a version smarter than itself – and this cycle can theoretically repeat at an increasing speed.

Is AI recursive self-improvement already happening?

A fully closed, human-independent cycle has not yet been documented by any leading laboratory. At the same time, individual elements of the cycle – code automation, conducting experiments, partial participation in choosing the next research steps – are already actively used.

How much code does Anthropic's Claude write itself?

According to the company, as of May 2026, over 80% of the code that enters Anthropic's production codebase is written by Claude, compared to single-digit figures in early 2025.

Why is Anthropic calling for a pause in AI development if it itself demonstrates such results?

The company presents this data not as marketing, but as an argument for the industry: if several leading laboratories simultaneously approach the threshold of self-improvement, a controlled and coordinated pause, in its opinion, would be safer than an uncontrolled race.