The development of the AI field rests on whether future machines can improve themselves. In 1966, British mathematician I. J. Good wrote:
"An ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind."
For decades, researchers have both anticipated and guarded against this "recursive self-improvement" (RSI). Now, as AI capabilities rapidly advance, a question becomes realistic: Could this process already be underway?
But the term RSI itself is fraught with ambiguity.
Some use it as a buzzword to push for regulation; others treat it as a marketing slogan. To some, RSI means a fully autonomous, closed-loop system; for others, as long as "technology helps create technology," that already counts as some form of self-improvement.
The safest way to understand it is perhaps as a continuum.
Under the strictest definition, RSI refers to a system that can improve not only its outputs but also its own improvement process: it can generate ideas, evaluate results, and modify its methods entirely without human intervention. By this standard, most AI systems today still fall short. They can indeed help build better AI, but they still rely on humans to set goals, define success criteria, and decide which modifications are worth keeping.
The real question isn't "does self-improvement already exist?" but rather to what extent this loop has already been closed.
Stepping Stones Toward Self-Improvement
In fact, researchers have been paving the way for RSI for many years.
Machine learning algorithms have long been able to automatically adjust program parameters; evolutionary algorithms can continuously generate, screen, and iterate design proposals; over the past decade, AutoML has begun automating parts of the neural network architecture design, training, and evaluation process.
Today, large language models—such as OpenAI's GPT, Google DeepMind's Gemini, Anthropic's Claude, and xAI's Grok—are pushing this trend even further. One of the most important uses of these models is writing code. Including the code used to generate the next generation of models.
This past February, OpenAI indicated that GPT-5.3-Codex has played a significant role in its own development process, helping to debug training, manage deployment, and analyze evaluation results. Meanwhile, Anthropic claims that most of its code is now written by Claude Code. Despite this, these systems still require humans to direct and validate the entire process.
In 2025, Google DeepMind also unveiled a system called AlphaEvolve—a "coding agent for scientific and algorithmic discovery." It uses large language models to guide the evolution of solution spaces, for example, optimizing neural network structures, data center scheduling, and chip design. While it still requires humans to define problems and evaluation criteria, each algorithmic breakthrough in turn enhances the capability of AI research and development itself.
Caption: High-level overview of AlphaEvolve.
Computer scientist Matej Balog, who is involved with AlphaEvolve, commented: "It is a highly collaborative process." Often, human researchers gain new inspiration from novel solutions discovered by the AI.
At the same time, the co-lead of DeepMind's earlier chip design system, AlphaChip, has founded a new company called Ricursive Intelligence, aiming to use AI to design AI chips.
Co-founder Azalia Mirhoseini states they hope to compress the traditional chip design cycle of one to two years down to "a few days":
The first stage: AI-assisted human design;
The second stage: AI automatically completes chip development for companies without specialized teams;
The third stage is using AI to design better AI chips, and then using those chips to train more powerful AI.
However, the research team emphasizes that this process will still retain human oversight.
Some other research directly targets the goal of a "system modifying its own behavior." For instance, last year, the University of British Columbia and Sakana AI released Darwin Gödel Machines (DGMs): a system that uses evolutionary algorithms to continuously improve code agents based on LLMs.
Related link: https://spectrum.ieee.org/evolutionary-ai-coding-agents
Caption: DGMs and Super-Agents.
Although they cannot yet modify the underlying language model itself, they are becoming increasingly adept at "improving themselves"; more advanced versions have even begun to modify the very mechanisms of their own improvement.
The same research team later developed AI Scientist, a system designed to attempt to automate the entire scientific research loop. This meant that it is no longer just "coding" being automated, but experimentation, evaluation, and even knowledge production itself.
The 'Intelligence Explosion' Still Faces Significant Friction
That said, not everyone believes the "singularity" is just around the corner.
Many researchers point out that today's AI is still only "reasonably good" at the steps of generating ideas, implementing code, and evaluating results, and is far from being fully autonomous.
Nathan Lambert recently suggested that rather than a future of "recursive self-improvement," we might instead see a form of "lossy self-improvement." As systems grow more complex, friction and coordination costs will gradually slow down the entire flywheel.
Related link: https://www.interconnects.ai/p/lossy-self-improvement
Another practical issue is cost. The development of today's most cutting-edge AI systems has reached costs in the tens of billions of dollars. No company is willing to truly hand over such an expensive system entirely to AI for autonomous operation.
Furthermore, even if an AI can design better software, it doesn't mean it can immediately take over the complex production systems of the real world. Achieving full RSI might require not only the AI to design chips and algorithms, but also for it to build data centers, run power generation systems, mine minerals, and manage robotic production chains.
These capabilities are currently still deeply dependent on human society and industrial infrastructure.
Caption: Self-improving AI versus collaborative AI.
The 'AI Cambrian Explosion'
Some researchers believe the very way people imagine RSI might be wrong.
Many have envisioned an increasingly powerful single super AI, but the reality might be more akin to biological evolution. It could resemble something like a Cambrian explosion of artificial life forms. At that time, a vast number of different types of AI agents would emerge simultaneously, possessing their own ecological, cultural, and economic systems.
But would humans be cut out of the scientific research loop at that point?
Perhaps, but likely more slowly.
Human researchers would first withdraw from low-level work, no longer personally debugging details, acting more like professors or team leaders responsible for choosing research directions. Then, humans might function more like project managers or CEOs, setting broader goals. Further down the line, the human role would gradually morph into that of a supervisor.
Though perhaps, when AI evolves to the point of curing cancer, some scholars would be quite happy to give up their beloved career.
Original article link: https://spectrum.ieee.org/recursive-self-improvement