AI
Self-Improving AI Is No Longer Locked Behind Closed Lab Doors
A WIRED editor shows how recent tools are making recursive AI training accessible to anyone—not just the biggest research labs.
WIRED — AI
The race to build self-improving AI has long been dominated by well-funded frontier labs, each trying to achieve models that refine themselves in an autonomous feedback loop. But emerging tooling from startups now suggests that recursive self-improvement needn't remain the exclusive domain of OpenAI, Anthropic, and Google.
To test this theory, a WIRED editor spent a week experimenting with accessible tools to train and continuously refine custom AI models for real-world tasks—from simple language models to a specialized paper-discovery agent. The results were remarkably functional, and the implications ripple beyond novelty: a more distributed ecosystem of specialized AI systems could undermine the concentrated power now held by a handful of frontier companies.
Starting small with open-source infrastructure#
The first experiment used AutoResearch, a tool built by Andrej Karpathy (formerly of OpenAI and Tesla, recently joining Anthropic). The approach was deliberately low-tech: dump a training task into Claude, provide raw compute power via an Nvidia DGX workstation, and let the model autonomously adjust parameters and refine outputs.
Over a few days, Claude iteratively improved a small language model's coherence and reduced its tendency toward repetitive gibberish. The results weren't cutting-edge, but they demonstrated a practical self-reinforcing loop available to anyone with the right hardware and API access.
Building specialized tools for real work#
The second experiment scaled up the ambition. Using Prime Intellect—a platform designed to let users train custom models for specific tasks—the editor built a model called Frontier_Paper_Curator to automatically find and summarize AI research papers.
The process was surprisingly autonomous: Claude collected past newsletter entries as training examples, generated synthetic data, and orchestrated multiple models to evaluate and refine outputs through reinforcement learning. Within a day, the system produced usable summaries of research papers, though still prone to over-eagerness and generic descriptions.
Vincent Weisser, CEO of Prime Intellect (which recently raised $15 million), frames the company's mission as democratization. Rather than concentrating AI power in elite institutions, the goal is to scatter specialized "intelligences" across industries and niches through accessible training infrastructure.
