A WIRED writer describes experimenting with "self-improving" AI to automate parts of his newsletter, arguing that recursive self-improvement — often framed as the frontier labs' path to superintelligence — is now within reach of ordinary users. His experiments suggest that this technology need not be monopolised by a handful of large AI companies, pointing to a more decentralised future in which specialised models are trained cheaply for niche tasks.

Using Andrej Karpathy's AutoResearch tool, the author had Anthropic's Claude autonomously train and refine a small language model on his own Nvidia hardware, watching its output improve from gibberish to greater coherence. He then used a tool from the startup Prime Intellect to build "Frontier_Paper_Curator", a custom model that finds and summarises research papers, produced in under a day using synthetic data and reinforcement learning. Prime Intellect's chief executive, Vincent Weisser, and rival firm Adaption champion democratising such training, while the piece notes the risks of over-relying on a single frontier model — citing Anthropic's blocking of certain requests to its Fable 5 model.