'we can interpret a model’s cross-entropy loss as “how distinguishable” the model is from its training distribution, and use that to upper bound the difficulty of training a model to perform reliable, high-quality reasoning over long sequences.'

https://epochai.org/blog/the-direct-approach