Claude is building its own smaller models now — here is how
a wired writer used claude and autoresearch to autonomously train and refine a smaller language model, proving that recursive self-improvement is possible outside frontier labs.
a journalist used claude to drive the autoresearch framework, automating the training and iterative refinement of a small language model on an nvidia dgx without human intervention. while early outputs were incoherent, the model improved autonomously over time. the experiment extended to using prime intellect to create a custom model, frontier_paper_curator, for research curation using synthetic data and reinforcement learning. this shift suggests specialized, self-improving models can be built by individuals, potentially reducing the industry's reliance on large, centralized frontier labs for specific applications.