AI Creating AI: Beyond LLMs to 'Recursive Self-Improvement' [Tech Talk]

Recursive Self-Improvement AI Draws Industry Attention
Automating Performance Gains Beyond LLM Limitations
Massive Investments Flow to AlphaGo Creator's Startup

Editor's NoteThis series unpacks essential yet unfamiliar technologies—ranging from AI and semiconductors to telecommunications and biotech—in a way that's easy to understand and relevant to our everyday lives.

The Einstein Test.

This is the standard for “Artificial General Intelligence (AGI)” proposed by Demis Hassabis, CEO of Google DeepMind, in February 2026. Much like how Einstein established the theory of relativity in 1915 through a thought experiment, this test demands that AI must be able to independently formulate such a theory without any prior knowledge.

Demis Hassabis, CEO of Google DeepMind, attending the 'Artificial Intelligence Summit' in India last February. At this event, he proposed the 'Einstein Test' as the next goal for artificial intelligence. YouTube screenshot.

Today’s large language model (LLM) AIs are certainly smart, but in reality, LLMs are essentially taking an open-book exam. Developers provide them with carefully curated, high-quality data, from which the models extract and combine answers. While they may seem like experts who know everything, they cannot truly engage in creative thinking. So, what does AI need to reach real human-level intelligence? The current AI frontrunners see “recursive self-improvement” as the answer.

AI That Finds Answers Through Trial and Error—Without Training

Recursive Self-Improvement (RSI) refers to AI systems that discover solutions to problems on their own through trial and error, then further enhance themselves.

Let’s imagine a human scientist gives an AI a specific coding challenge. RSI places countless AI agents in a virtual environment to tackle the problem, with each agent trying a different approach. Most will fail, but some will succeed. Then, a “supervising AI” selects the best outcome and uses it to further improve itself. From the outside, it would appear as if AI is building even more advanced AI.

NVIDIA's ARM-based VERA central processing unit (CPU). Unlike the existing large language models (LLM) where data exchange between memory semiconductors and AI accelerators was crucial, recursive self-improvement elevates the CPU as the most important hardware essential for operating AI agents. NVIDIA

Major AI companies such as OpenAI, Google DeepMind, and Anthropic have already made RSI the next competitive frontier. Jack Clark, co-founder of Anthropic, wrote on X (formerly Twitter) in early May that he believes the probability of achieving RSI before 2028 is over 60%, adding, “A world where AI creates itself is coming.” OpenAI has also laid out an internal roadmap to develop a “fully automated AI researcher” by March 2028.

RSI: The Alternative to LLMs Hitting a Dead End

The motivation for leading AI companies to invest in RSI is clear. If AI can build and improve itself, it can dramatically reduce research and development (R&D) costs. Moreover, the process of self-improvement can be massively accelerated using ever-more powerful supercomputers. Coding tasks that used to take humans years can be condensed into mere seconds, leading to explosive gains in performance.

However, RSI’s greatest strengths are its virtually unlimited learning opportunities and versatility. LLMs are limited by data constraints: to work well, they require high-quality human data, which must be processed into textbooks for pre-training. Additionally, whenever AI needs to be tailored for a specific task, further adjustment—known as “fine-tuning”—is required. RSI, on the other hand, can learn endlessly without relying on data, and its performance does not sharply decline even when prior knowledge or data is limited. As CEO Hassabis put it, “Einstein-level intelligence” can be demonstrated across all industries and academic fields.

Astronomical Investments Already Pouring In: “A Fundamental Question About Intelligence”

Nevertheless, RSI is still an area where full-scale research has only just begun. While traditional LLMs required memory semiconductors to store vast datasets, RSI—where processing AI agents is essential—demands investment in previously neglected fields such as central processing units (CPUs) and data input/output (I/O). Even for Big Tech, it is still largely uncharted territory.

Professor David Silver of University College London, who secured the largest seed investment in European history. He is the person who led the development of self-learning artificial intelligence such as AlphaGo and AlphaZero at DeepMind, laying the foundation for RSI. David Silver personal homepage

This is why startups are now emerging that are boldly abandoning the already-commercialized LLMs of Big Tech to go “all-in” on RSI research. One clear example is Ineffable Intelligence, founded by David Silver, a University College London (UCL) professor and former DeepMind researcher. Upon its launch, the company attracted nearly 1.1 billion dollars in investment from some of the world’s largest venture capital firms, reaching a valuation of 5.1 billion dollars. The fact that such massive funding is flowing into a company without a concrete product roadmap yet reveals the level of anticipation surrounding RSI.

Professor Silver, revered as the mastermind behind AlphaGo and AlphaZero and a pioneer of RSI, argues that RSI—not LLMs—is the true candidate for the artificial general intelligence that could transform human civilization. In a statement posted on the day Ineffable Intelligence was founded, Silver said, “AI that generates language, video, and source code is already good enough and will be further improved by talented people,” but emphasized, “We must confront fundamental questions about the very concept of ‘intelligence.’”

He added, “While the probability of RSI failing is extremely high, the benefits, should it succeed, are beyond imagination,” and, “We aim to achieve ultra-fast learning capabilities, enabling the acquisition of knowledge and skills without relying on human data and through continuous self-driven improvement.”

This content was produced with the assistance of AI translation services.