The biggest names in artificial intelligence are chasing a holy grail: models that can refine and improve themselves without human intervention. Some researchers believe this recursive loop could eventually produce systems far smarter than any person. But for Will Knight, a tech journalist with a weekly newsletter to produce, the question was more practical: Could he use this same approach to automate some of his own grunt work?
After a week of tinkering, the answer turned out to be a resounding yes. Knight didn't build a model from scratch—he used existing tools to train a small language model with help from Anthropic's Claude. He installed AutoResearch, a project by AI luminary Andrej Karpathy, which lets a larger AI model guide the training of a smaller one. Claude handled the heavy lifting: adjusting parameters, tweaking training regimens, and evaluating the smaller model's output. Knight provided the hardware—an Nvidia DGX supercomputer that ran hot for days—and gave Claude permission to skip safety checks. Early results were rough: an attempt to complete the phrase "In the beginning…" produced gibberish. But later versions, refined autonomously, became more coherent.
Knight then tried a different approach using a startup called Prime Intellect, which uses AI to train custom models for specific tasks. He fed it around 100 of his past newsletter entries and asked Claude to build a model, dubbed Frontier_Paper_Curator, to find and summarize interesting research papers. Claude generated synthetic training data and tapped another model to evaluate the results. Vincent Weisser, CEO of Prime Intellect, says his company wants to make this kind of recursive improvement accessible to anyone, not just a handful of frontier labs. "Give every company access to frontier training infrastructure," he says, "and the collective creativity of the market unlocks far more than any handful of labs can." After less than a day, Knight had a working model. It was overeager and its summaries were generic, but it showed promise. The experiment hints at a future where AI improvement isn't controlled by a few giants—and where even a journalist can build a tool that saves him from busywork.
Source: Wired