“The performance gap between closed-source and open-source models appears to be narrowing, contrary to expectations that the capital intensity of training would cause it to widen.”