Earlier this year, Elon Musk, head of Tesla and xAI, predicted that 2026 would be the “year of the singularity.” The Singularity has been described as the moment when artificial intelligence becomes so advanced that it surpasses human intelligence and begins self-improvement. Only three months into the year, Musk’s predictions already seem to be showing early signs. A new experimental project by Andrej Karpathy, one of OpenAI’s early members and former AI director at Tesla, shows how AI systems can autonomously run experiments to improve other models.
Mr. Karpathy recently introduced a project called ‘autoresearch’. He built a system in which an AI agent repeatedly modifies and tests the training code of a language model in an automated loop. Instead of human researchers manually tweaking code, testing parameters, and evaluating results, this AI agent itself suggests changes to the model, runs experiments, measures performance, and maintains versions with improved performance.
Simply put, the Karpathy experiment shows how AI systems can attempt to improve themselves without continuous human intervention. Andrei Karpathy shared the results of the project online, revealing how the AI agent autonomously ran hundreds of experiments and gradually improved the model.
“Who knew early singularities could be so much fun? 🙂 We’ve confirmed that the improvements found in automated research over the past two days (about 650) experiments on the depth 12 model translate well to depth 24. So nanochat is also trying to get a new leaderboard entry for ‘Time to GPT-2’. Good luck,” he said. I wrote in one of my posts about.
Karpathy’s project quickly became a topic of discussion across the AI community, highlighting how it can provide an early indication of what engineers call a technological singularity. “The singularity has begun. There are so many signs,” Tobi Lutke wrote in response to this development. Musk echoed Lutke’s comments.
What is singularity?
The term singularity refers to the point at which artificial intelligence is expected to surpass human intelligence and begin rapidly self-improving without human intervention. The ability for machines to design better versions of themselves, often coupled with the idea of artificial general intelligence, could rapidly accelerate progress and lead to breakthroughs far beyond current capabilities, researchers suggest.
Currently, Karpathy’s project is still a small-scale research prototype, a far cry from the self-improving intelligence imagined in science fiction. However, we can get a glimpse of how AI can improve itself.
Technology leaders say AI will reach singularity in 2026
As for when machines will truly become “superintelligent,” some leaders in the AI industry think it could happen sooner than expected.
Speaking previously at the World Economic Forum, Musk suggested that AI could surpass the intelligence of a single human by the end of 2026, and the intelligence of all humans combined by around 2030.
He also argued that if AI becomes so powerful, the combination of superintelligent systems and humanoid robots could eventually eliminate the need for traditional jobs and lead to what he called “universal high incomes.”
Musk is not alone in making such predictions. Previously, in his June 2025 essay “The Gentle Singularity,” Sam Altman argued that humanity may already be approaching the “event horizon” of the singularity, and that a technological “takeoff” has begun.
Meanwhile, Demis Hassabis also suggested that the world is nearing a “breaking point” with the rise of autonomous, agent-based AI systems. However, he notes that true AGI is still a long way off, potentially taking five to eight years.
– end
