Frontier researchers have found that dumping more and more data into training is effective at improving LLM capabilities. Nobody has a gears-level understanding of how training on some particular kind of data leads to some particular behaviors, so they generally take the attitude that more is better.