“According to a study by Nathan Lambert, it is no longer economically sensible to pre-train a foundation model from scratch once it reaches 30 billion parameters, as post-training optimization becomes more efficient.”