I Built a Self-Improving AI, and So Can You
A journalist experiments with recursive self-improvement using accessible tools and finds that this technology can democratize AI development rather than leaving it to frontier labs.
Experimenting with Self-Improvement
A technology journalist's experiment with self-improving AI models has shown that this technology can be useful not only for leading labs but for everyone. The journalist started with a simple loop using AutoResearch, a tool created by Andrej Karpathy, to train a small model from scratch. He used Claude's assistance and an Nvidia DGX "supercomputer," allowing the model to operate without usual permission checks.
Initial results were poor—the model produced meaningless repetitions—but after several corrections by Claude, it became more coherent. While not comparable to cutting-edge AI, it demonstrated the potential for continuous improvement.
A More Complex Tool: Prime Intellect
Continuing the experiment, the journalist used a tool from startup Prime Intellect to create a model named "Frontier_Paper_Curator" that can find and summarize interesting research papers. He collected about 100 previous newsletter entries, created a training environment, and asked Claude to help build the model. Claude added synthetic data, while another model evaluated the output, and the training environment used reinforcement learning.
Prime Intellect CEO Vincent Weisser emphasizes that the company aims to make recursive self-improvement accessible to everyone. "We don't want one centralized, almost godlike intelligence, we want a billion intelligences that go into all the niches that create beautiful things," he says.
Other Players and Risks
Another startup, Adaption, offers a tool called AutoScientist that automates AI model training. CEO Sara Hooker notes that large companies without in-house AI experts could benefit from such solutions. Meanwhile, the editorial notes that reliance on frontier labs carries risks, as when Anthropic blocked certain requests, revealing dependency problems.
In the end, after less than a day with Prime Intellect, the journalist created a surprisingly good model that, while not perfect, can select and summarize research. It's a step toward freedom from daily busywork.
