“Key algorithmic ideas for training frontier AI models are becoming increasingly proprietary, which will lead to greater performance divergence between labs.”