AI systems are already writing code, running experiments and optimizing parts of the technology used to build other AI. The next step, closing that improvement loop, could change how quickly artificial intelligence advances.
THE UNIVERSAL RECORD
Sourced reporting. No opinions.
Brad Socha | September 25, 2026 | 8:09 AM EST
Artificial intelligence is beginning to participate directly in the process that creates better AI. It can write research code, design experiments, optimize algorithms and evaluate potential improvements. Researchers have now demonstrated systems that can even modify parts of their own software and keep changes that make them perform better.
That does not mean an AI has become an autonomous intelligence capable of repeatedly designing more powerful versions of itself. Humans still determine research priorities, provide computing resources, establish evaluation criteria and decide whether changes are accepted or deployed. But the boundary between AI as a product of research and AI as a participant in that research is becoming less distinct.
The idea is known as recursive self-improvement, or RSI. In its strongest theoretical form, an AI improves its ability to conduct AI research, uses that improved capability to make another improvement, and repeats the process. Each generation could potentially become better at producing the next.
Evidence published in 2026 suggests that pieces of this feedback loop already exist. What remains uncertain is whether those pieces can be combined into sustained, open-ended improvement, and whether progress would accelerate dramatically if they were.
AI Is Already Helping Build Better AI
The clearest evidence comes from inside AI laboratories.
Anthropic reported in September that Claude “leads” 26% of its measured AI research and development work, meaning the system can complete most of those tasks from a high-level instruction while a person supervises. More than 90% of the company’s measured AI R&D involves Claude at least performing substantial portions of the work. Anthropic emphasized that none of the measured work is fully autonomous.
OpenAI reported a similar shift on September 6. It said coding agents are allowing its researchers to write code faster, run more experiments and delegate increasingly complex tasks. By mid-August, the company measured approximately 3.1 agent workdays for every human workday across its research organization. OpenAI nevertheless said people still determine research priorities, judge which results deserve further investigation and make decisions about scaling or deploying systems.
Google DeepMind’s AlphaEvolve provides another example. The system uses Gemini models to propose computer programs, automatically evaluates them and builds on successful candidates. DeepMind reported that AlphaEvolve has optimized parts of Google’s computing infrastructure and AI training, including technology involved in training the language models underlying AlphaEvolve itself. One optimization reduced the training time for Gemini by about 1%, according to DeepMind.
Researchers have gone further with experimental systems.
The Darwin Gödel Machine, introduced in 2025 by researchers from Sakana AI and the University of British Columbia, modifies the code of its own coding agent and evaluates whether those modifications improve performance. Its researchers reported an increase on the SWE-bench software-engineering benchmark from 20% to 50%, while experiments were conducted with sandboxing and human oversight.
A new preprint submitted on September 22, 2026 describes AIDE², an AI research agent designed specifically to improve its own research process. During an autonomous multi-day experiment, the system repeatedly proposed modifications to its own code, tested the resulting agents and retained successful changes. The researchers reported that improvements transferred to previously unseen tasks, although the findings remain a preprint and have not undergone peer review.
What Recursive Self-Improvement Would Change
Today’s systems demonstrate bounded improvement: people largely determine the environment, objectives and measurements of success.
Full recursive self-improvement would be different.
Imagine an AI researcher that discovers a more efficient training technique. That technique produces a more capable AI researcher, which discovers a better model architecture. The improved system then develops stronger research tools, allowing another generation to search more effectively.
The important factor would not simply be intelligence. It would be feedback.
Human scientific progress is constrained by researchers’ time, communication, experimentation and training. Software agents can potentially operate simultaneously, copy successful methods, run experiments continuously and perform some tasks much faster than people.
That possibility is why recursive self-improvement has sometimes been associated with an “intelligence explosion”, the hypothesis that improvements could eventually compound fast enough that AI capabilities rise extremely rapidly.
There is no evidence that such an uncontrolled intelligence explosion is occurring today.
A 2026 survey examining 1,250 AI research papers distinguished existing “bounded self-refinement” from open-ended recursive self-improvement. The authors concluded that demonstrated RSI remains constrained by requirements for reliable evaluation, computing resources and problems such as model or diversity collapse. They identified choosing worthwhile research directions as a particularly important area where humans remain involved.
Evaluation may ultimately be one of the largest obstacles. An AI can optimize a computer program when success can be measured precisely, speed, accuracy or a benchmark score. Determining whether a new training strategy creates a genuinely better, safer and more reliable intelligence is much harder.
An AI optimizing the wrong measurement could become extremely effective at improving a score without producing the result humans actually wanted.
The Question Is Whether Humans Remain Inside the Loop
Recursive improvement could have substantial benefits. Automated researchers could test more ideas, reduce the cost of experiments, improve software and hardware efficiency and potentially accelerate work in areas ranging from mathematics to medicine.
The same feedback mechanism creates a fundamental control problem.
If an AI eventually becomes capable of choosing research directions, modifying its own research tools, conducting experiments, evaluating the results and building its successor, human researchers would no longer be performing the most important steps in the development cycle.
Anthropic says that point has not been reached. Its researchers report that Claude performs strongly when executing well-defined experiments but continues to show important weaknesses when deciding which goals should be pursued. OpenAI similarly says its automated research systems still operate under human direction and that it does not know how to safely achieve fully aligned recursive self-improvement.
That distinction separates what exists today from the more dramatic scenarios surrounding self-improving machines.
AI is already helping researchers improve AI. Experimental systems can rewrite parts of themselves and preserve modifications that perform better. Those are measurable developments, not science fiction.
What has not been demonstrated is an AI independently creating progressively more capable successors without humans setting the objectives, supplying the infrastructure or judging the results.
If that final loop eventually closes, the central question may no longer be simply how intelligent an AI system becomes. It will be whether humans still determine what “better” means.
Sources:
Anthropic Institute — Measurements for Understanding the Pace of AI Development Inside Frontier Labs
https://www.anthropic.com/institute/measuring-pace-of-ai-development
Anthropic Institute — When AI Builds Itself
https://www.anthropic.com/institute/recursive-self-improvement
OpenAI — Research Acceleration: The View Inside OpenAI
https://openai.com/index/research-acceleration-view-inside-openai/
Google DeepMind — AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced Algorithms
https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
arXiv — Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
https://arxiv.org/abs/2505.22954
Sakana AI — The Darwin Gödel Machine: AI That Improves Itself by Rewriting Its Own Code
https://sakana.ai/dgm/
arXiv — Recursive Self-Improvement of AI Research Agents
https://arxiv.org/abs/2609.26457
arXiv — Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
https://arxiv.org/abs/2607.07663
About the Author
Brad Socha is the founder of The Universal Record, focused on sourced, factual global reporting. Coverage includes international news, geopolitics, technology, and major developments.