OpenAI Cofounder Builds AI System That Can Improve Itself

Artificial intelligence may be moving closer to a future long discussed by scientists and technologists — the technological singularity. Earlier this year, Elon Musk, CEO of Tesla and founder of xAI, predicted that 2026 could become the “year of the Singularity.”

The concept refers to a moment when artificial intelligence becomes powerful enough to surpass human intelligence and begins improving itself without human intervention.

Now, only a few months into the year, new research experiments are giving the technology community a glimpse into what that future might look like. A new experimental project created by Andrej Karpathy — one of the early members of OpenAI and former Director of AI at Tesla — demonstrates how AI systems could autonomously run experiments to enhance other AI models.

The project suggests that self-improving AI systems, once considered theoretical, may be starting to appear in early experimental forms.

Andrej Karpathy’s ‘Autoresearch’ Project

Karpathy recently introduced a research prototype called “autoresearch.” The experiment focuses on automating the process of improving machine learning models — a task traditionally performed by human researchers.

How the System Works

In the autoresearch framework, an AI agent repeatedly modifies and evaluates the training code of a language model through an automated loop.

Instead of human researchers manually adjusting training parameters, writing new code, testing models, and analysing results, the AI system handles these steps independently.

The agent performs several tasks automatically:

Key Functions of the AI Agent

• Proposes changes to the model architecture or training code

• Runs experiments on the modified model

• Measures performance metrics

• Retains model versions that perform better

This process allows the system to gradually improve the model through continuous experimentation.

AI Running Hundreds of Experiments on Its Own

The results of the project show that AI systems may already be capable of improving their own performance through repeated testing cycles.

Sharing his findings online, Karpathy explained that the AI agent conducted hundreds of experiments autonomously, gradually refining the model over time.

“Who knew early singularity could be this fun? :) I just confirmed that the improvements autoresearch found over the last two days of (~650) experiments on a depth-12 model transfer well to depth-24, so nanochat is about to get a new leaderboard entry for ‘time to GPT-2’ too. Works,” he wrote in one of his posts on X.

The experiment quickly gained attention within the AI research community and sparked discussions about the potential emergence of self-improving AI systems.

Tech Leaders React to Early Signs of Self-Improving AI

The development drew reactions from several influential figures in the technology sector.

Tobi Lütke, CEO of Shopify, described the development as a sign of rapid progress in AI systems.

“The singularity has begun. so many signs,” wrote Tobi Lutke in response to the development.

Elon Musk also agreed with Lutke’s observation, adding further momentum to discussions about the possibility that AI may be entering a new phase of rapid advancement.

What Is the Technological Singularity?

The technological singularity is a theoretical point in the future when artificial intelligence surpasses human intelligence and begins to improve itself without human assistance.

Key Characteristics of the Singularity

Rapid Self-Improvement

Once machines become capable of designing better versions of themselves, the pace of innovation could accelerate dramatically.

Link to Artificial General Intelligence

The singularity is often associated with the development of Artificial General Intelligence (AGI) — AI systems that can perform any intellectual task that humans can.

Potential for Breakthrough Innovations

Experts believe this stage could lead to technological breakthroughs far beyond today’s capabilities.

However, Karpathy’s autoresearch project is still far from the type of superintelligent system imagined in science fiction. Instead, it serves as an early demonstration of how AI systems might eventually learn to improve themselves.

Predictions That AI Could Reach Singularity by 2026

Several technology leaders believe that the timeline for superintelligent AI may be closer than previously expected.

During remarks at the World Economic Forum, Elon Musk predicted that artificial intelligence could exceed the intelligence of any individual human by the end of 2026.

He further suggested that AI might surpass the combined intelligence of all humans by around 2030.

Impact on the Future of Work

According to Musk, highly advanced AI systems combined with humanoid robots could fundamentally reshape the global economy.

He suggested that such technologies could eliminate the need for many traditional jobs and potentially lead to what he described as a “universal high income.”

Other AI Leaders Share Similar Views

Musk is not the only technology leader predicting a rapid acceleration in AI development.

Sam Altman’s “The Gentle Singularity”

Sam Altman, CEO of OpenAI, discussed the concept in his essay “The Gentle Singularity,” published in June 2025.

In the essay, Altman argued that humanity may already be approaching the “event horizon” of the singularity and suggested that the technological “takeoff has started.”

Demis Hassabis on the Rise of Agentic AI

Demis Hassabis, CEO of Google DeepMind, has also pointed out that the world may be approaching a “threshold moment” due to the rapid emergence of autonomous and agent-based AI systems.

However, Hassabis believes that fully developed AGI may still take time and estimates that it could arrive within five to eight years.

Conclusion

Andrej Karpathy’s autoresearch experiment offers an intriguing glimpse into the future of artificial intelligence. While the system remains a research prototype, it demonstrates how AI agents could eventually conduct experiments and optimize machine learning models without constant human involvement.

Although the technology is still far from achieving the true technological singularity, developments like this highlight how quickly AI capabilities are advancing. With major industry leaders predicting that superintelligent systems could emerge within the next decade, experiments such as autoresearch may represent the earliest steps toward a future where AI systems learn, improve, and innovate on their own.