“OpenAI sees a promising path to overcoming the lack of computer-use training data by using existing models to bootstrap the creation of synthetic data.”