It is almost as if there’s a documented limit in how much you can squeeze out of autoregressive transformers by throwing compute at it