If a model's max
input size is \(k\), we then approximate the likelihood of a token \(x_t\) by conditioning only on the
\(k-1\) tokens that precede it rather than the entire context.