Claude Opus reportedly built Jev classifiers that matched or beat fine-tuned RoBERTa on six tasks
Lightfield says the agent wrote readable classification instructions for Jev rather than tuning a model’s weights.
TLDR
Lightfield reports that, given the same training labels, natural-language classifiers built by Claude Opus and run on Jev matched or beat fine-tuned RoBERTa across six tasks. In a separate setup with no labels at the start, the builder used questions and labels from a simulated user; its classifiers beat zero-shot results on four tasks where that input helped. Lightfield says it is releasing its builder recipes, tools and paper.
Combined views
—
2 Sources, first seen ago
