Can You Build a Self-Improving AI? One Writer Put It to the Test

The biggest names in artificial intelligence are locked in a race to build models that can improve themselves. The theory is that once an AI starts tweaking its own code in a feedback loop, it could eventually outthink its creators. But one journalist wondered if the same trick could handle something far more mundane: the busywork of putting out a weekly newsletter.

After a week of tinkering, the answer was a surprising yes. The experiment didn't just save time—it revealed a different vision for AI's future, one where small teams and solo operators can train specialized models without relying on the handful of companies that dominate the industry. To start, the writer handed the heavy lifting to Claude, using a tool called AutoResearch created by AI pioneer Andrej Karpathy. While Claude adjusted parameters and training methods, the writer provided the hardware—an Nvidia DGX desktop supercomputer that ran hot for days—and gave the model unusual freedom to skip permission checks.

Early results were rough. Asked to complete the phrase "In the beginning," the first model produced gibberish. But as Claude autonomously refined the smaller model's training, the output grew coherent. It wasn't GPT-5, but it showed steady improvement. The writer then turned to a startup called Prime Intellect, which recently raised $15 million to make recursive self-improvement accessible to everyone. Using 100 past newsletter entries as training data, Claude helped build a model called Frontier_Paper_Curator, designed to find and summarize research papers. The CEO, Vincent Weisser, argues that democratizing this kind of training unlocks far more value than any single lab can: "We don't want one centralized, almost godlike intelligence—we want a billion intelligences that go into all the niches that create beautiful things."

Other startups are chasing the same idea. Adaption offers AutoScientist, which automates model training, and its CEO notes that many large companies are burning through tokens without in-house experts. The risk of relying on a single frontier model became clear when Anthropic blocked certain requests to its latest model, Fable 5. Palantir's Alex Karp has warned that using frontier labs means handing over data and control. After less than a day with Prime Intellect, the writer had a usable model for curating research—still overeager and a bit generic, but a promising start toward freeing up time for the work that matters.